Paper deep dive

Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces

Francesca Carlon, Vincent Ginis, Andres Algaba

Year: 2026Venue: arXiv preprintArea: cs.LGType: PreprintEmbeddings: 162

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Large language models often reason at length before answering, increasing cost and latency. Prompts and trained settings can shorten this reasoning, but a shorter trace may only show that the model stopped sooner. Here, we evaluate paired runs of the same question at matched reasoning horizons across 198 GPQA Diamond and 500 MMLU-Pro questions. We test a numeric/concision prompt that announces a token limit for Qwen3-14B and the trained effort settings of gpt-oss-20b and -120b. The Qwen prompt shortens reasoning traces by 12-17%, while accuracy changes at matched token limits are small and mixed. A concise/early-answer instruction raises MMLU-Pro accuracy by 3.8 percentage points at 512 tokens, including +2.7 points when both runs are unfinished. Its gain at 2,048 tokens is uncertain. For gpt-oss, candidate-logit answers from completed low- and medium-effort reasoning are 14.5-26.3 points more accurate than matched-horizon high-effort answers. Most of the 512-token advantage comes from lower effort finishing earlier, while differences among unfinished runs are smaller and mixed. Wrong early answers often concentrate probability on the chosen option, so earlier stopping does not uniformly improve probability quality. In these tests, a tight deadline can favor lower effort or a concise instruction, whereas allowing high effort to finish can recover higher final accuracy. Evaluations should report correct completion before a deadline, the answer obtained when a run is stopped, differences among unfinished runs, and probability assigned to the correct answer separately.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

160,597 characters extracted from source content.

Expand or collapse full text

SHORTER REASONING, EARLIER ANSWERS? AN EVALUATION OF REASONING INTERFACES Francesca Carlon 1,2,* 0009-0004-2152-2745 Vincent Ginis 1,2,3 0000-0003-0063-9608 Andres Algaba 1,2 0000-0002-0532-3066 1 Data Analytics Lab, Vrije Universiteit Brussel, Pleinlaan 5, 1050 Brussels, Belgium 2 imec-SMIT, Vrije Universiteit Brussel, Pleinlaan 9, 1050 Brussels, Belgium 3 School of Engineering and Applied Sciences, Harvard University, Cambridge, Massachusetts 02138, USA August 5, 2026 ABSTRACT Large language models often reason at length before answering, increasing cost and latency. Prompts and trained settings can shorten this reasoning, but a shorter trace may only show that the model stopped sooner. Here, we evaluate paired runs of the same question at matched reasoning horizons across198GPQA Diamond and500MMLU-Pro questions. We test a numeric/concision prompt that announces a token limit for Qwen3-14B and the trained effort settings of gpt-oss-20b and -120b. The Qwen prompt shortens reasoning traces by12–17%, while accuracy changes at matched token limits are small and mixed. A concise/early-answer instruction raises MMLU-Pro accuracy by3.8 percentage points at512tokens, including+2.7points when both runs are unfinished. Its gain at2,048 tokens is uncertain. For gpt-oss, candidate-logit answers from completed low- and medium-effort reasoning are14.5–26.3points more accurate than matched-horizon high-effort answers. Most of the 512-token advantage comes from lower effort finishing earlier, while differences among unfinished runs are smaller and mixed. Wrong early answers often concentrate probability on the chosen option, so earlier stopping does not uniformly improve probability quality. In these tests, a tight deadline can favor lower effort or a concise instruction, whereas allowing high effort to finish can recover higher final accuracy. Evaluations should report correct completion before a deadline, the answer obtained when a run is stopped, differences among unfinished runs, and probability assigned to the correct answer separately. Keywords large language models· reasoning interfaces· reasoning budgets· Qwen3· gpt-oss 1 Introduction Large language models often solve hard questions by writing intermediate reasoning before giving a final answer, a behavior that began as chain-of-thought prompting and is now trained into models whose reasoning length or effort can be adjusted by the user [1,2,3,4]. Accuracy on competition mathematics, graduate science, and code often grows as models spend more reasoning tokens during generation, sampling, or aggregation [5, 6, 7, 8]. This reasoning is costly and variable. A single answer can consume tens of thousands of tokens, and additional tokens often add little once the answer has effectively been chosen [9,10,11,12]. Controlling that cost is therefore a practical constraint when models operate under deadlines [13, 14, 15]. We use the term reasoning control for an intervention that can change how much intermediate reasoning a model produces before its final answer. Examples include training that favors shorter reasoning [16,17,18], settings such as gpt-oss reasoning effort [19], methods that steer generation [20,21,22], and prompts that request concise reasoning [23,24,25]. These controls reach the model through different reasoning interfaces: instructions written into *Corresponding author: francesca.carlon@vub.be 1 arXiv:2608.03401v1 [cs.LG] 4 Aug 2026 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces the prompt or settings trained into the model. Most studies evaluate these interfaces by plotting final accuracy against the length of the completed trace [26]. That comparison measures the final cost–accuracy tradeoff, but it cannot show whether a control improves the answer after a fixed amount of reasoning or simply makes the model stop sooner. More broadly, reported benchmark scores depend on how much test-time compute an evaluation allows, and recent work calls for protocols that make this allocation explicit [27]. A model that stops early has a completed answer at a point where another model may still be working. Comparing those states can answer a useful deployment question, but it does not show that the unfinished reasoning itself is better. Prior work uses stopped trajectories and forced answers to study answer stabilization and early stopping [28,29,30]. We build on that work by comparing paired runs at matched reasoning horizons and by recording separately whether each run has already finished. At each stopping point, we measure whether the model has finished correctly, what answer an unfinished run gives when stopped, how the runs compare while both are still reasoning, and how much probability the model places on the correct answer. For Qwen3-14B, we compare ordinary generation with generation under the numeric/concision prompt at the same externally imposed pointB(Figure 1). We separately test a concise/early-answer instruction that asks the model to reason concisely and prioritize an early usable answer. For gpt-oss, we compare a completed low- or medium-effort run with high effort at the point where the lower-effort run chose to stop. We validate the forced-answer measurement against completed answers, controls using the same questions, a scripted positive control, and generated-answer continuations. The main paired experiments cover Qwen3-14B and gpt-oss-20b/-120b on GPQA Diamond [31] and a500-item MMLU-Pro subset sampled across subject areas [32]. Qwen3-4B/-8B calibrate the measurement, and Omni-MATH- 2 [33,34] extends the evaluation to open-ended mathematics. In the tested Qwen3 case, the numeric/concision prompt shortens traces without a consistent same-horizon gain. A nine-arm MMLU-Pro experiment finds an early gain from the concise/early-answer instruction but no clear benefit from matching the stated number to the stopping point. In the tested gpt-oss case, lower effort gives a correct answer earlier mainly because it stops and commits sooner. These within-family case studies separate completion timing from the answer obtained when a run is stopped. They do not provide a causal comparison between prompts and trained effort settings. 2 Matched-horizon probes separate evaluation points from generation policies Letiindex items,rindex independent generation replicates,zindex the generation condition, andL z ir denote the condition’s terminal reasoning length. We writeY z,ρ ir (B) ∈ 0, 1 for correctness when conditionzis evaluated at horizonBwith readoutρ. The standardized prefix-probe readout (ρ = P) closes the stopped reasoning block and selects among the candidate answers. The natural-terminal readout (ρ = N) uses the answer given when the model finished without interruption. The generated-answer continuation (ρ = C) closes an externally stopped reasoning block and greedily generates an unconstrained final-channel answer. These readouts measure which answer is available, not when the model decided. Figure 1 and Table A1 specify where each readout enters the two case studies. Token-exact replay and candidate-logit scoring describe different stages of the measurement. Token-exact replay constructs the stopped input directly from stored prompt, reasoning, and closure token IDs, without decoding and retokenizing the reasoning prefix. The candidate-logit probe then scores every valid one-token spelling of each answer letter from full-vocabulary logits and aggregates spellings by letter. We report the total probability mass assigned to valid options before normalization, conditional log loss, multiclass Brier score, and argmax accuracy. These proper scores and valid-option mass diagnose the standardized probe; they do not make it a natural continuation. We distinguish four readouts. The policy-context readout preserves each condition’s generation prompt at answer time and measures the full interface-level contrast. The common-context readout uses the ordinary base prompt for every Qwen arm or the medium-effort answer context for every gpt-oss arm, replays the stored reasoning IDs, and applies a common closure. Differences at answer time are then carried by the generated prefixes. The both-active readout applies the common context only to pairs in which both traces remain unfinished atB; it is diagnostic because conditioning on survival changes the compared population. The behavioral continuation greedily generates a short final answer from a stopped prefix instead of selecting an option from first-position logits. Completed reasoning blocks are replayed under the common context when estimating prefix-carried differences, whereas the terminal-aware policy readout retains their natural answers. Full closure texts and prompt variants appear in Tables A1 and A2 and Section A. For Qwen3-14B, lethdenote ordinary generation with the horizon hidden (the surprise stop) andvthe textual numeric/concision treatment with the horizon visible (the bundled prompt). The terminal-aware policy-context readout uses Y z,T pol ir (B) = Y z,N ir (L z ir ) when the trace finishes before B, and Y z,P pol ir (B) otherwise. The primary curve is τ Q,pol (B) =E i,r h Y v,T pol ir (B)− Y h,T pol ir (B) i . 2 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces source context generated prefix answer readoutpaired score a Qwen3-14B: numeric/concision prompting Only one source trace sees the numeric/concision prompt before reasoning; both are cut at the same B. surprise stop budget hidden while reasoning question only C 1 : B surprise budget hidden limited-time closure S surprise (B) bundled-prompt stop numeric/concision prompt question + prompt B C 1 : B prompt prompt visible limited-time closure S prompt (B) same imposed horizon B discarded continuation Δ prompt (B) = S prompt (B) −S surprise (B) source context generated prefix answer readoutpaired score b gpt-oss-20b/120b: effort setting No numeric target is imposed; high effort is truncated to the lower-effort realized length L c . low/medium effort effort setting c ∈low, med question + effort c C 1 : L c c terminal prefix common readout S c (L c ) high effort matched same item, same length question + high effort C 1 : L c high truncated prefix common readout S high (L c ) same length L c discarded high-effort tail Δ c −high = S c (L c ) −S high (L c ) Figure 1: We compare paired runs at matched reasoning horizons. a, For Qwen3-14B, the terminal-aware policy contrast evaluates two traces from the same item and generation replicate at the same imposed token horizonB: an ordinary trace with the horizon hidden (“surprise stop”) and a trace generated under the textual numeric/concision treatment (“bundled-prompt stop”). Truncated prefixes are read in their policy context, while a trace that finishes before Bkeeps its natural terminal answer. Common-context and both-active diagnostics hold the answer-time prompt fixed (Table A1). b, For gpt-oss-20b and gpt-oss-120b, the effort contrast compares a completed low- or medium-effort reasoning prefix with high-effort reasoning at that same itemwise lower-effort lengthL c (c∈low, medium), under a common medium-effort answer context. If high effort finishes beforeL c , its completed reasoning block is replayed; the high-active subset is diagnostic. Replayed reasoning prefixes exclude the original final-answer tokens. Both arms therefore face the same imposed horizon, although a naturally completed trace can contain fewer thanB tokens. For any tested instruction a, we also report correct natural completion by the deadline, K a ir (B) = 1L a ir ≤ BY a,N ir , and the both-active common-context contrast τ a Q,BA (B) =E i,r h Y a,P com ir (B)− Y h,P com ir (B)| L a ir > B, L h ir > B i . The four-stratum decomposition partitions every pair into both active, instruction complete/ordinary active, ordinary complete/instruction active, and both complete. It reports each stratum’s frequency, within-stratum contrast, and contribution to the unconditional terminal-aware difference. The literal cap-then-answer experiment and the continuation check of the instruction arms replace candidate-logit scoring with the behavioral readout C at B ∈512, 2048. For gpt-oss, letc∈low, mediumdenote the lower effort. The matched horizon is the lower-effort policy’s item- and replicate-specific terminal lengthL c ir . LetR c ir = 1when both token-exact reasoning prefixes can be reconstructed: high effort either reachesL c ir or has a recoverable terminal boundary before it. For replayable pairs, letℓ high,c ir = L c ir when high effort remains active at the matched horizon andℓ high,c ir = L high ir when high effort finishes earlier. Under a common medium-effort answer context, define D c ir = Y c,P com ir (L c ir )− Y high,P com ir (ℓ high,c ir ). This definition selects the firstL c ir high-effort reasoning IDs when high effort remains active and the complete high-effort reasoning block when it finishes earlier. The primary all-replayable estimand is τ c G,R =E i,r [D c ir | R c ir = 1]. 3 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces 10 3 10 4 imposed horizon B (reasoning tokens) 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 terminal-aware policy accuracy GPQA Diamond a Qwen3-4B Qwen3-8B Qwen3-14B gpt-oss-20b gpt-oss-120b 10 3 10 4 imposed horizon B (reasoning tokens) 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 MMLU-Pro b 10 3 10 4 imposed horizon B (reasoning tokens) 0.0 0.2 0.4 0.6 0.8 1.0 survival P ( L > B ) c GPQA Diamond MMLU-Pro original swapped trace shu ffl ed tokens random tokens 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 accuracy (pooled over budgets) d Qwen3-14B gpt-oss-20b Figure 2: Longer stopped traces reveal more answer information, and the effect depends on the question-specific reasoning. Surprise-stop probes use the full-vocabulary candidate-logit forced-answer readout on GPQA Diamond (n=198) and the stratified MMLU-Pro subset (n=500); colors and markers denote models. a,b, Terminal-aware policy accuracy increases with the imposed horizonBand approaches the cohort-matched natural full-trace reference shown at the right of each panel, with95%item-bootstrap bands. Token positions are paired within model; cross-family positions are descriptive because each model uses its own tokenizer. c, SurvivalP (L > B)falls sharply withB, so late-budget survivor accuracy describes a selected set of long-running items. d, Length-matched prefix controls under the same candidate-logit readout, available for Qwen3-14B and gpt-oss-20b, show that recovered accuracy comes from the item-specific reasoning prefix: original prefixes beat swapped-trace, shuffled-token, and random-token controls on the common four-arm item–budget cohort pooled over budgets. Figure A1 reports the returned-top-20readout comparison. The high-active subset, in which high effort has not reached a terminal boundary byL c ir , is a diagnostic rather than the primary analysis. Fixed-checkpoint both-active strata address the narrower comparison between two unfinished policies. These conditional estimands are not commensurable with τ Q,pol (B). We retain every matched item–replicate pair with an observed lower-effort terminal outcome. When a high-effort trace cannot be replayed, worst-case bounds set its outcome first to correct and then to incorrect (Section A). Greedy final- channel continuations repeat the all-replayable matched-horizon comparison withρ = Cand provide the behavioral check. We apply these probes to Qwen3-4B, -8B, -14B [35] and gpt-oss-20b, -120b [19] on GPQA Diamond (198graduate science questions, four choices) [31], MMLU-Pro (12,032questions, ten choices) [32], and Omni-MATH-2 (4,181open- ended problems) [33,34]. The main paired Qwen and gpt-oss comparisons use GPQA Diamond and a category-stratified 500-item MMLU-Pro subset. Omni-MATH-2 supplies natural full-trace accuracy and imposed- or matched-horizon open-ended comparisons. Full prompts, generation settings, parsing rules, and statistical procedures are in Section A. 3 Stopped reasoning retains question-specific answer information Stopped prefixes reveal increasingly accurate, question-specific answers as the imposed horizon grows (Figure 2a,b,d). For Qwen3-14B on MMLU-Pro, terminal-aware policy accuracy rises from50.4%without reasoning to60.0%,70.8%, and76.6%atB = 512,2,048, and8,192, approaching the cohort-matched full-trace accuracy of76.8%(384/500; 4 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A3). Across all five models, saturation follows natural reasoning length: the shorter gpt-oss traces plateau by B = 2,048, while Qwen remains on the rising part of the curve and smaller models show a few late-horizon reversals or near-ties (Figure A2). The calibration curves and natural references share one natural-trace cohort. The Qwen prompt comparison below uses separately generated, replicate-matched cohorts (Section A). Survival falls as terminal-aware accuracy rises, making survivor-only accuracy a selected comparison (Figure 2c and Table A3). The model’s decision to continue reasoning changes which items remain at a large horizon, so survivor-only accuracy is flat or non-monotone even as the terminal-aware curve improves. AtB = 8,192, only18.7%of Qwen3-14B and9.1%of gpt-oss-20b GPQA trajectories are still reasoning. We therefore report survival with every horizon curve rather than conflating the evaluation point with the selected set of items that reaches it. The candidate-logit probe also reproduces completed answers without replaying final-channel text. A separate validation cohort yields98.3–100.0%agreement between replayed-prefix and parsed natural answers across its model–benchmark– policy combinations (Table A4). A single-run check also finds no final-channel delimiter in2,788terminal gpt-oss replay prefixes (Table A30). These cohorts validate terminal replay but do not exhaustively check the three-replicate token-exact analysis. The fixed-stop and prefix-control probes have no errors or missing valid answer-token IDs. Answers asserted inside the reasoning remain part of the measured behavior. Length-matched controls confirm that the recovered answer depends on the question-specific prefix (Figure 2d). Swapping in reasoning from another question reduces accuracy by18.1points for gpt-oss-20b and17.8points for Qwen3-14B, while shuffled-token and random-token controls fall between the swap and original prefixes (Figure A3). These checks support comparisons between generated prefixes at matched evaluation points. 4 The numeric/concision prompt shortens Qwen reasoning without a consistent early accuracy gain The numeric/concision prompt shortens typical Qwen3-14B reasoning by 12–17% without a consistent accuracy gain at the same imposed horizon (Figure 3a,b,f). Across the32-fold range fromB = 512to16,384, the treated median changes only from3,740.5to3,925tokens on GPQA Diamond and remains between881and929on MMLU-Pro. The within-item slopes oflogLonlogBare0.012and0.018, with rank correlations of0.058and0.121(Table A10). GPQA adherence rises from1.9%at the tightest budget to99.2%at the largest, showing that the prompt changes typical length but weakly targets the announced count. The12–17%contrast concerns geometric-mean reasoning length, not arithmetic-mean token cost. The candidate-logit comparison is a controlled analogue of Qwen’s documented external-stop-then-answer procedure [36,35]. It evaluates the externally stopped prefix with deterministic candidate- logit scoring, while the literal generated-answer validation implements the two-stage answer continuation. The prompts and closures are specified in Section A. The same-horizon accuracy differences remain small and inconsistent across the main curve (Figure 3a,b and Table A6). GPQA Diamond estimates range from−0.5to+1.3points, while the four MMLU-Pro estimates are+2.0,+0.7, −0.9, and+0.1points. Only the MMLU-ProB = 512estimate has both a pointwise interval ([+0.4, +3.7]) and simultaneous band ([+0.1, +3.9]) above zero. Termination-stratum contributions are also small and benchmark- dependent (Table A18). Replicate-sensitive, answer-context, proper-score, and literal cap-then-answer analyses show no consistent gain across benchmarks and horizons (Tables A11 to A13 and A19). Dense trajectory probes show positive log-odds AUC shifts but no clear stable-correct gain (Table A17). The proper-score table stops atB = 2,048because the jointly active sets atB = 8,192/16,384contain only25/1GPQA Diamond items and19/2MMLU-Pro items. The companion data release contains those sparse estimates. A concise/early-answer instruction produces a clear early gain, whereas matching the announced number to the stop does not (Figure 3c,d). The nine-arm experiment uses500MMLU-Pro items with eight replicates per item. At B = 512, the concise/early-answer instruction raises terminal-aware accuracy from59.7%to63.5%(+3.8points, 95%CI[+2.1, +5.6]) and reduces median reasoning length from1,083to679tokens (Table A7). Its simultaneous interval over all16instruction–horizon contrasts is[+1.7, +6.0]points, and direct comparisons place it+2.3to+3.4 points above the numeric and exact-stop prompts at this deadline (Table A9). AtB = 2,048, its+1.0-point accuracy difference is unclear ([−0.4, +2.4]). Matching the announced and imposed horizons gives−0.2points ([−1.0, +0.6]) for the numeric/concision prompt and+0.1points ([−0.6, +0.8]) for advance exact-stop notice. Both intervals lie within about one percentage point of zero, although this is a descriptive comparison rather than an equivalence test. The512-token concise/early-answer gain combines earlier completion with a change in unfinished prefixes (Table A8). Correct natural completion rises from 18.2% to 31.9% (+13.8 points), while replaying all prefixes under the ordinary prompt retains a+3.1-point accuracy difference ([+1.5, +4.8]). Among the2,476pairs in which both traces remain active, the common-context difference is+2.7points ([+0.6, +5.0]). AtB = 2,048, correct completion remains+7.6 5 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces 10 3 10 4 imposed horizon B (reasoning tokens) 0.5 0.6 0.7 terminal-aware policy accuracy a GPQA Diamond, prompt GPQA Diamond, surprise MMLU-Pro, prompt MMLU-Pro, surprise 10 3 10 4 imposed horizon B (reasoning tokens) −0.10 −0.05 0.00 0.05 0.10 Δ prompt ( B ) b GPQA Diamond MMLU-Pro −20246 instruction minus ordinary accuracy (p) Irrelevant number; states 512 Concise/early-answer instruction Number only; states 512 Limited budget; no number Advance stop notice; states 2,048 Advance stop notice; states 512 Numeric/concision prompt; states 2,048 Numeric/concision prompt; states 512 MMLU-Pro, stopped at 512 tokens c −20246 instruction minus ordinary accuracy (p) Irrelevant number; states 512 Concise/early-answer instruction Number only; states 512 Limited budget; no number Advance stop notice; states 2,048 Advance stop notice; states 512 Numeric/concision prompt; states 2,048 Numeric/concision prompt; states 512 MMLU-Pro, stopped at 2,048 tokens d −6−4−20246 numeric/concision prompt minus hidden horizon (p) 95% item bootstrap CI (observed replicates averaged within item) GPQA Diamond, B = 512 GPQA Diamond, B = 2,048 MMLU-Pro, B = 512 MMLU-Pro, B = 2,048 Literal cap-then-answer validation e 10 3 10 4 instructed budget B 10 3 10 4 reasoning tokens L L = B within-item estimates (GPQA/MMLU) log–log slope: 0.012/0.018 Spearman ρ: 0.058/0.121 f GPQA Diamond MMLU-Pro Figure 3: The numeric/concision prompt shortens Qwen reasoning without a consistent early accuracy gain. a, Terminal-aware policy accuracy for the broader numeric/concision prompt (solid) and surprise stop (dashed), pairing three explicit generation replicates on GPQA Diamond (n = 198) and MMLU-Pro (n = 500) with token-exact replay and candidate-logit scoring. b, The corresponding contrast∆ prompt (B). Primary95%intervals average the three paired replicate outcomes within item and bootstrap items; two-way item×replicate and crossed-intercept intervals are sensitivity analyses. c,d, MMLU-Pro mechanism controls at imposed stops of512and2,048tokens. Points are instruction minus ordinary accuracy, with pointwise95%intervals from a nested bootstrap over500items and eight paired replicates per item. Generation and probing use stored token IDs. Table A7 reports the instruction grid and matching-number test, Table A8 reports completion and both-active results, and Table A9 reports multiplicity-adjusted, direct-contrast, and generated-answer checks. e, Literal cap-then-answer validation at external capsB = 512and2,048, following Qwen’s documented two-stage procedure. Points are accuracy differences between the numeric/concision- prompt and hidden-horizon arms; primary bars condition on the three observed trajectory sets and bootstrap items (Table A6). f, Median realized reasoning length with interquartile bands versus the instructed budget. Realized length changes little with the announced budget: within-item slopes oflogLonlogBare0.012on GPQA Diamond and0.018 on MMLU-Pro, with Spearman correlations of 0.058 and 0.121 (Table A10). 6 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces points higher, but the both-active difference is+0.5points ([−4.8, +5.5];644pairs) and the pooled candidate-logit gain is unclear. Greedy answer continuations retain positive concise/early-answer point estimates at both stops, including differences of +4.1 points in the policy context and +3.5 under the common context at B = 512 (Table A9). The literal cap-then-answer validation of the numeric/concision prompt reproduces the benchmark split but no broad gain (Figure 3e and Table A6). A text-reconstructed four-budget crossed analysis shows no diagonal matching-number pattern, and a separate four-horizon concise/early-answer comparison likewise shows no broad advantage (Figure A4 and Table A23). In a pooled MMLU-ProB = 512decomposition, category differences range from−7.1to+7.8 points (Table A23). A single-rollout Qwen3-8B comparison also shows no broad same-horizon gain (Table A22). A 200-item open-ended Omni-MATH-2 comparison finds differences of+3.5,+2.0, and−4.1points across the three tested horizons, with every interval spanning zero (Table A5). A scripted control moves from25.0%to100.0%at B = 64and reaches100.0%in both arms byB = 512, confirming that the readout detects a known early-answer policy (Table A29). A text-reconstructed MMLU-Pro subset replication broadens the item sample, but its saved Qwen records lack generation seeds and completion token IDs and therefore do not enter the token-exact primary analysis. On a second500-item subset, the numeric/concision contrast is+2.5points atB = 512([+0.4, +4.8]), and pooling the original and independent subsets gives+2.1points ([+0.6, +3.6]). On that independent subset atB = 2,048, the numeric/concision estimate is−0.2points ([−2.6, +2.2]), whereas the concise/early-answer estimate is+2.7points ([+0.5, +5.0]; Table A24). A separate text-reconstructed, eight-instruction prompt-control comparison atB = 512likewise shows small, mixed GPQA estimates and larger early MMLU-Pro point estimates, with the concise/early-answer instruction again largest. Its clearest pattern is instruction-dependent reasoning length rather than a consistent accuracy gain (Table A25). Two diagnostic analyses ask what remains after an early answer has formed. Continuing an incorrect Qwen prefix raises rescue from20.0–30.1%under immediate answering to26.5–38.2%with a long hidden continuation, but35.8–55.8% of continuations remain anchored to the original wrong answer and stating the remaining budget gives no consistent benefit (Table A26). Hidden-state readouts recover the exact announced budget at the prompt boundary with balanced accuracy0.938on GPQA Diamond and0.905on MMLU-Pro, then mostly fall to label-shuffle levels after reasoning begins. Binary bundled-prompt membership remains modestly decodable at most post-start token positions, with balanced accuracy from0.588to0.700(Table A27). These analyses characterize answer rescue, anchoring, and persistence of prompt information; they are not estimates of the matched-horizon effects. 5 Lower gpt-oss effort improves early accuracy mainly by stopping sooner At the low- or medium-effort stopping point, candidate-logit accuracy from the completed lower-effort reasoning block exceeds that from the matched-horizon high-effort prefix in all eight model–benchmark–effort comparisons (Figure 4a,b). The primary comparison replays each completed lower-effort block and its corresponding reconstructable high-effort prefix under a common medium-effort answer context; if high effort finishes before the matched horizon, its completed reasoning block is replayed. Because lower effort chooses the horizon, the result includes its decision to stop and does not establish a uniform advantage among unfinished prefixes. The high-active subset and the fixedB = 512 termination strata isolate those narrower comparisons. Full estimands and replay rules are in Table A1 and Section A. Medium is the default effort setting, and none of the settings specifies a target length [19]. At the lower-effort terminal horizon, candidate-logit accuracy differences range from+14.5to+26.3points, and every primary interval excludes zero (Figure 4a,b and Table A14). The high-active diagnostic remains positive in all eight comparisons, and all eligible outcome pairs were replayable (nonreplayable= 0; arm-level accuracies and coverage are in Table A20). Greedy answer continuations preserve the ordering without option normalization, with differences of +15.5 to +25.9 points and intervals excluding zero throughout (Table A14). Earlier completion explains most of the gpt-oss advantage at the fixedB = 512checkpoint (Figure 4c–e). The lower-complete/high-active stratum contributes+3.4to+15.8points and is the largest positive component in all eight comparisons. Contributions from pairs that are both active are smaller and mixed (−1.0to+3.2points), and none of their within-stratum accuracy intervals excludes zero. Both-complete contributions range from−1.5to+0.1points (Table A16). Although the lower-complete/high-active share varies from13.5%to73.7%, its contribution remains positive. Thus, at B = 512, lower effort’s earlier completion accounts for most of the fixed-checkpoint gap. Proper scores qualify the accuracy gain (Figure 4f). Brier estimates favor the terminal lower-effort arm in seven of eight comparisons, with intervals excluding zero in five, whereas conditional log loss favors the high-effort prefix in all eight, with all intervals excluding zero (Table A15). This divergence is concentrated in the lower tail of correct-option probability. Across comparisons,20.9%–36.0%of probed terminal answers assign it less than10 −4 , compared with 0.2%–11.7%of matched high-effort prefixes (Table A15). An incorrect completed trace can concentrate mass on its 7 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces −505101520253035 accuracy difference (p) at the same realized length 95% item bootstrap CI (observed seeds averaged within item) GPQA Diamond, low minus high GPQA Diamond, medium minus high MMLU-Pro, low minus high MMLU-Pro, medium minus high gpt-oss-20b a −505101520253035 accuracy difference (p) at the same realized length 95% item bootstrap CI (observed seeds averaged within item) GPQA Diamond, low minus high GPQA Diamond, medium minus high MMLU-Pro, low minus high MMLU-Pro, medium minus high gpt-oss-120b b low minus high medium minus high −100102030 termination-stratum accuracy contribution (p) at B= 512 GPQA Diamond, low minus high GPQA Diamond, medium minus high MMLU-Pro, low minus high MMLU-Pro, medium minus high gpt-oss-20b c −100102030 termination-stratum accuracy contribution (p) at B= 512 GPQA Diamond, low minus high GPQA Diamond, medium minus high MMLU-Pro, low minus high MMLU-Pro, medium minus high gpt-oss-120b d both active lower complete, high active both complete high complete, lower active −1001020 both-active accuracy difference (p) at B= 512 95% item bootstrap CI gpt-oss-20b, GPQA Diamond gpt-oss-20b, MMLU-Pro gpt-oss-120b, GPQA Diamond gpt-oss-120b, MMLU-Pro Fixed-checkpoint jointly active prefixes e 05 terminal minus high negative favors terminal Log loss f −0.50.0 terminal minus high negative favors terminal Brier Figure 4: Lower gpt-oss effort improves early accuracy mainly by stopping sooner, while proper scores qualify the accuracy gain. a,b, Matched-horizon accuracy gaps for gpt-oss-20b and gpt-oss-120b on GPQA Diamond and MMLU-Pro, pairing each item within three explicit generation replicates. Each completed low- or medium-effort prefix is read under a common medium-effort final-answer context and compared with every reconstructable high-effort prefix at that itemwise horizon. The primary analysis includes early high-effort completions; the high-active subset is reported separately. Primary95%intervals average the three paired replicate outcomes within item and bootstrap items. c,d, Fixed-checkpointB = 512decomposition by termination stratum. Stacked bars show each stratum’s signed contribution to the lower-effort-minus-high accuracy gap; the stratum legend in d applies to both panels. e, Within-stratum accuracy differences for the pairs in which both traces remain active atB = 512. f, Conditional log-loss and multiclass Brier-score contrasts at the policy-chosen matched horizon. Values are the completed lower-effort readout minus the matched-horizon high-effort readout, so negative proper-score differences favor lower effort. Matched- horizon accuracy, the high-active diagnostic, and generated-answer validation are in Table A14; fixed-checkpoint strata are in Table A16; and proper scores with their correct-option probability tails are in Table A15. The descriptive accuracy–median-length plot is in Figure A6. 8 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces chosen letter, producing near-zero correct-option probabilities that dominate unbounded log loss; the corresponding incorrect-answer score tails are in Table A21. Brier score caps these tails, whereas argmax accuracy ignores confidence. Replicate-specific, two-way, and crossed-intercept calculations agree with the primary finite-run result (Table A11). Answer-context, source-context, and closure checks at the reported checkpoints do not reverse the family-level pattern (Tables A12 and A19). In model-specific paired cohorts of 475 items for gpt-oss-20b and 490 for gpt-oss-120b from a difficulty-stratified 500-item Omni-MATH-2 draw, completed lower-effort answer accuracy exceeds matched-horizon high-effort answer accuracy by+33.5to+42.0points across four comparisons; fixed-checkpoint comparisons are reported in Table A22, but the single rollout does not establish that the estimate generalizes across sampled generations (Table A5). Dense probes place the lower-effort answer earlier in the trajectory: terminal-aware AUC differences are positive in all eight gpt-oss comparisons, whereas jointly active early-prefix differences are much smaller (Table A17 and Figure A5). Surface diagnostics show more digits and operators, fewer explicit alternative-strategy cues, and effort classification accuracy of0.70from the first64tokens (Figures A8 and A9). These item-conditional diagnostics are not used to compare effect magnitudes across the two case-study estimands. Adding numeric-budget wording to the gpt-oss effort instruction produces no broadB = 512advantage over effort alone (Table A28). Ten of the twelve intervals span zero. The gpt-oss-20b medium-effort difference on GPQA Diamond is−5.6points ([−10.6,−0.5]); the gpt-oss-20b high-effort difference on MMLU-Pro is+2.7points, with a rounded lower endpoint of zero ([+0.0, +5.5]). The effort-only baseline ends atB = 512, leaving longer announced-budget contrasts unmatched. High effort can nevertheless be more accurate when allowed to finish (Figure A6). The completed-effort view describes accuracy against median trace length rather than arithmetic-mean token use or expected API cost; it provides completed-trace context but is not the matched-horizon estimand. At the same fixedB = 512checkpoint, low- or medium-minus-high differences range from+2.8to+19.0points (Figure A7 and Table A16). Effort changes both when the model finishes and what it answers at a given checkpoint. 6 Discussion Matched-horizon comparisons reveal two distinct ways to answer earlier: a control can make the model finish sooner, change the answer available in an unfinished prefix, or do both. The Qwen3-14B numeric/concision prompt mainly reduces typical reasoning length, whereas the concise/early-answer instruction improves the early answer through both earlier completion and a change in unfinished prefixes. Lower gpt-oss effort produces a substantially better answer at its chosen stopping point, but most of its512-token advantage comes from having finished while high effort is still reasoning. Differences between jointly active gpt-oss prefixes are smaller and depend on the model and task, and high effort can be more accurate once allowed to finish. Reasoning-interface evaluations should therefore separate correct completion by a deadline, the answer obtained by stopping an unfinished run, differences between jointly active runs, and probability assigned to the correct answer. The gpt-oss results show how earlier completion can improve the first two outcomes, the Qwen concise/early-answer result shows that an instruction can also change unfinished reasoning, and the gpt-oss log-loss result shows that a better selected answer need not imply uniformly better probability quality. Matched-horizon evaluation connects language-model reasoning to work on anytime algorithms, deadline-aware planning, and rational metareasoning [13, 14, 15]. Those traditions separate the value of an intermediate answer from the policy that allocates computation before a deadline. A shorter final trace does not establish that the model planned around a deadline, and a better answer at lower terminal cost does not imply continuous remaining-budget tracking. The same distinction may matter when reasoning traces are used to study behavior under resource pressure [37,38], although the present tasks contain no competing objective or safety constraint and do not test behavioral misalignment. Limitations. Three limitations bound our conclusions. First, primary experiments cover two model families and two multiple-choice benchmarks with noncomparable prompt and effort interventions. We therefore restrict claims within families; single-rollout Qwen3-8B and Omni-MATH-2 estimates describe only the sampled trajectories. Second, three generation replicates limit stochastic inference; text-reconstructed Qwen subset and prompt-control cohorts lack token IDs and verified seeds. We report replicate-specific, two-way, and crossed-intercept sensitivity analyses; these cohorts inform subset replication and prompt patterns, not primary token-exact estimates. Third, forced-answer readout affects unfinished-prefix accuracy, and near-zero correct-option probabilities affect gpt-oss log loss. We validate against completed answers and continuations and report valid-option mass, Brier scores, medians, tails, and a scripted positive control. Conclusions remain specific to the tested models, tasks, horizons, and answer procedures. 9 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Acknowledgements This research was supported by funding from the Flemish Government under the “Onderzoeksprogramma Artificiële Intelligentie (AI) Vlaanderen” program. Andres Algaba acknowledges support from the Francqui Foundation (Belgium) through a Francqui Start-Up Grant and a fellowship from the Research Foundation Flanders (FWO) under Grant No.1286924N. Vincent Ginis acknowledges support from Research Foundation Flanders under Grant No.G032822N and G0K9322N. The computational resources and services used in this work were provided by the VSC (Flemish Supercomputer Center), funded by the Research Foundation Flanders (FWO) and the Flemish Government - department WEWIS. Author contributions A.A. and F.C. conceived the study and designed the experiments. A.A. implemented the probing and generation code and performed the analysis. F.C., V.G., and A.A. interpreted the results and wrote the manuscript. Data and code availability The code for this publication is publicly available at https://github.com/AndresAlgaba/shorter_reasoning. Data associated with this study are available in a public repository athttps://doi.org/10.5281/zenodo. 21773770. The GPQA Diamond dataset [31] is available at https://huggingface.co/datasets/Idavidrein/gpqa. The MMLU-Pro dataset [32] is available at https://huggingface.co/datasets/TIGER-Lab/MMLU-Pro. The Omni-MATH-2 dataset [34] is available athttps://huggingface.co/datasets/martheballon/ Omni-MATH-2. The Qwen3 model family [35] is available at https://huggingface.co/collections/Qwen/qwen3. The gpt-oss model family [19] is available at https://huggingface.co/collections/openai/gpt-oss. We used Python 3.13.1 across separate environments. Data processing, analysis, and visualization used pandas 3.0.0, pyarrow 23.0.0, scikit-learn 1.7.2, SciPy 1.17.0, statsmodels 0.14.6, and matplotlib 3.10.8. Model generation used vLLM 0.12.0 with flashinfer-python 0.5.3; local candidate-logit readouts used PyTorch 2.11.0 and Transformers 5.8.0. 10 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces References [1]Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022. doi: 10.52202/068431-1800. URLhttps://doi.org/10.52202/ 068431-1800. [2]Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199–22213, 2022. doi: 10.52202/068431-1613. URL https://doi.org/10.52202/068431-1613. [3]Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. OpenAI o1 system card. arXiv preprint arXiv:2412.16720, 2024. URL https://arxiv.org/abs/2412.16720. [4]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature, 645(8081):633–638, 2025. ISSN 1476-4687. doi: 10.1038/s41586-025-09422-z. URLhttps://doi.org/10. 1038/s41586-025-09422-z. [5]Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. URL https://arxiv.org/abs/2203.11171. [6]Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling LLM test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314, 2024. URLhttps://arxiv.org/ abs/2408.03314. [7]Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori B Hashimoto. s1: Simple test-time scaling. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 20286–20332. Association for Computational Linguistics, 2025. doi: 10.18653/v1/2025.emnlp-main.1025. URLhttps://aclanthology. org/2025.emnlp-main.1025/. [8]Pranjal Aggarwal, Marjan Ghazvininejad, Seungone Kim, Ilia Kulikov, Jack Lanchantin, Xian Li, Tianjian Li, Bo Liu, Graham Neubig, Anaelia Ovalle, et al. Reasoning over mathematical objects: on-policy reward modeling and test time aggregation. arXiv preprint arXiv:2603.18886, 2026. URLhttps://arxiv.org/abs/2603. 18886. [9] Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, et al. Do not think that much for 2+3=? On the overthinking of o1-like LLMs. arXiv preprint arXiv:2412.21187, 2024. URL https://arxiv.org/abs/2412.21187. [10] Marthe Ballon, Andres Algaba, and Vincent Ginis. The relationship between reasoning and performance in large language models—o3 (mini) Thinks Harder, Not Longer. arXiv preprint arXiv:2502.15631, 2025. URL https://arxiv.org/abs/2502.15631. [11]Jinyan Su, Jennifer Healey, Preslav Nakov, and Claire Cardie. Between underthinking and overthinking: An empirical study of reasoning length and correctness in LLMs. arXiv preprint arXiv:2505.00127, 2025. URL https://arxiv.org/abs/2505.00127. [12] Renliang Sun, Wei Cheng, Dawei Li, Haifeng Chen, and Wei Wang. Stop when enough: Adaptive early-stopping for chain-of-thought reasoning. arXiv preprint arXiv:2510.10103, 2025. URLhttps://arxiv.org/abs/2510. 10103. [13] Thomas Dean and Mark Boddy. An analysis of time-dependent planning. In Proceedings of the Seventh National Conference on Artificial Intelligence, pages 49–54, 1988. [14] Shlomo Zilberstein. Using anytime algorithms in intelligent systems. AI Magazine, 17(3):73–83, 1996. [15]Stuart J. Russell and Eric H. Wefald. Do the Right Thing: Studies in Limited Rationality. MIT Press, Cambridge, MA, 1991. doi: 10.7551/mitpress/2474.001.0001. URLhttps://doi.org/10.7551/mitpress/2474.001. 0001. [16]Pranjal Aggarwal and Sean Welleck. L1: Controlling how long a reasoning model thinks with reinforcement learning. arXiv preprint arXiv:2503.04697, 2025. URL https://arxiv.org/abs/2503.04697. [17]Tingxu Han, Chunrong Fang, Shiyu Zhao, Shiqing Ma, Zhenyu Chen, and Zhenting Wang. Token-budget-aware LLM reasoning. In Findings of the Association for Computational Linguistics: ACL 2025, pages 24842– 24855. Association for Computational Linguistics, 2025. doi: 10.18653/v1/2025.findings-acl.1274. URL https://aclanthology.org/2025.findings-acl.1274/. 11 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces [18]Daman Arora and Andrea Zanette.Training language models to reason efficiently.arXiv preprint arXiv:2502.04463, 2025. URL https://arxiv.org/abs/2502.04463. [19]Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K Arora, Yu Bai, Bowen Baker, Haiming Bao, et al. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925, 2025. URL https://arxiv.org/abs/2508.10925. [20]Junyan Li, Wenshuo Zhao, Yang Zhang, and Chuang Gan. Steering LLM thinking with budget guidance. arXiv preprint arXiv:2506.13752, 2025. URL https://arxiv.org/abs/2506.13752. [21]Yi Sun, Han Wang, Jiaqiang Li, Jiacheng Liu, Xiangyu Li, Hao Wen, Yizhen Yuan, Huiwen Zheng, Yan Liang, Yuanchun Li, and Yunxin Liu. An empirical study of LLM reasoning ability under strict output length constraint. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 7652–7671. Association for Computational Linguistics, 2025. doi: 10.18653/v1/2025.emnlp-main.389. URL https://aclanthology.org/2025.emnlp-main.389/. [22] Hao Wen, Xinrui Wu, Yi Sun, Feifei Zhang, Liye Chen, Jie Wang, Yunxin Liu, Yunhao Liu, Ya-Qin Zhang, and Yuanchun Li. BudgetThinker: Empowering budget-aware LLM reasoning with control tokens. arXiv preprint arXiv:2508.17196, 2025. URL https://arxiv.org/abs/2508.17196. [23] Matthew Renze and Erhan Guven. The benefits of a concise chain of thought on problem-solving in large language models. arXiv preprint arXiv:2401.05618, 2024. URL https://arxiv.org/abs/2401.05618. [24] Sania Nayab, Giulio Rossolini, Marco Simoni, Andrea Saracino, Giorgio Buttazzo, Nicolamaria Manes, and Fabrizio Giacomelli. Concise thoughts: Impact of output length on LLM reasoning and cost. arXiv preprint arXiv:2407.19825, 2024. URL https://arxiv.org/abs/2407.19825. [25]Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. Chain of draft: Thinking faster by writing less. arXiv preprint arXiv:2502.18600, 2025. URL https://arxiv.org/abs/2502.18600. [26]Yang Sui, Yu-Neng Chuang, Guanchu Wang, et al. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419, 2025. URL https://arxiv.org/abs/2503.16419. [27]Jessica McFadyen, Ole Jorgensen, Harry Coppock, Kevin Wei, and Cozmin Ududec. How inference compute shapes frontier LLM evaluation. arXiv preprint arXiv:2606.17930, 2026. URLhttps://arxiv.org/abs/2606. 17930. [28]Marthe Ballon, Brecht Verbeken, Vincent Ginis, and Andres Algaba. Probing the trajectories of reasoning traces in large language models. arXiv preprint arXiv:2601.23163, 2026. URLhttps://arxiv.org/abs/2601.23163. [29]Ayan Datta, Zhixue Zhao, Bhuvanesh Verma, Radhika Mamidi, Mounika Marreddy, and Alexander Mehler. Large language models decide early and explain later. arXiv preprint arXiv:2604.22266, 2026. URLhttps: //arxiv.org/abs/2604.22266. [30]Tsung-Han Wu, Mihran Miroyan, David M. Chan, Trevor Darrell, Narges Norouzi, and Joseph E. Gonzalez. Are large reasoning models interruptible? arXiv preprint arXiv:2510.11713, 2025. URLhttps://arxiv.org/abs/ 2510.11713. [31]David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. GPQA: A graduate-level Google-proof Q&A benchmark. In First Conference on Language Modeling, 2024. URL https://openreview.net/forum?id=Ti67584b98. [32]Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. MMLU-Pro: A more robust and challenging multi-task language un- derstanding benchmark. Advances in Neural Information Processing Systems, 37:95266–95290, 2024. doi: 10.52202/079017-3018. URL https://doi.org/10.52202/079017-3018. [33]Bofei Gao, Feifan Song, Zhe Yang, Zefan Cai, Yibo Miao, Qingxiu Dong, Lei Li, Chenghao Ma, Liang Chen, Runxin Xu, et al. Omni-MATH: A universal olympiad level mathematic benchmark for large language models. arXiv preprint arXiv:2410.07985, 2024. URL https://arxiv.org/abs/2410.07985. [34]Marthe Ballon, Andres Algaba, Brecht Verbeken, and Vincent Ginis. Benchmarks saturate when the model gets smarter than the judge. arXiv preprint arXiv:2601.19532, 2026. URLhttps://arxiv.org/abs/2601.19532. [35]An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. URLhttps: //arxiv.org/abs/2505.09388. [36] Qwen Team. Qwen3 thinking budget.https://github.com/QwenLM/Qwen3/blob/main/docs/source/ getting_started/thinking_budget.md, 2025. Accessed 20 June 2026. 12 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces [37]Bowen Baker, Joost Huizinga, Leo Gao, Zehao Dou, Melody Y Guan, Aleksander Madry, Wojciech Zaremba, Jakub Pachocki, and David Farhi. Monitoring reasoning models for misbehavior and the risks of promoting obfuscation. arXiv preprint arXiv:2503.11926, 2025. URL https://arxiv.org/abs/2503.11926. [38]Tomek Korbak, Mikita Balesni, Elizabeth Barnes, Yoshua Bengio, Joe Benton, Joseph Bloom, Mark Chen, Alan Cooney, Allan Dafoe, Anca Dragan, et al. Chain of thought monitorability: A new and fragile opportunity for AI safety. arXiv preprint arXiv:2507.11473, 2025. URL https://arxiv.org/abs/2507.11473. [39] Quinn McNemar. Note on the sampling error of the difference between correlated proportions or percentages. Psy- chometrika, 12(2):153–157, 1947. doi: 10.1007/BF02295996. URLhttps://doi.org/10.1007/BF02295996. 13 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces A Methods A.1 Datasets and sampled items We use GPQA Diamond [31],198expert-validated graduate-level multiple-choice questions in biology, chemistry, and physics with four choices (A–D); MMLU-Pro [32],12,032questions across14categories with ten choices (A–J); and Omni-MATH-2 [33],4,181open-ended problems with free-form answers, introduced by Ballon et al.[34]. The Qwen numeric/concision experiments use the full GPQA Diamond set and a500-item MMLU-Pro subset drawn by proportional stratified sampling over the14categories with seed0, so that every budget condition is evaluated on the identical items. The nine-arm mechanism experiment reuses this MMLU-Pro subset. A.2 Models, prompts, and generation We evaluate Qwen3-4B, -8B, -14B [35] with a32,768-token context, and gpt-oss-20b, -120b [19] with a131,072- token context, all served locally from pinned Hugging Face snapshots. Qwen models run in bfloat16 without weight quantization; gpt-oss models use their released MXFP4 quantization. Full reasoning traces are generated in thinking mode with model-recommended sampling: Qwen3 uses temperature 0.6, top-p 0.95, and top-k 20with up to28,000reasoning tokens; gpt-oss uses temperature1.0, top-p 1.0, and top-k 50with up to126,000tokens. The main Qwen budget curve and gpt-oss experiments use three independently seeded global generation replicates. We store full completion token IDs and seeds, require every planned generation to be present, and reject any pair of global replicates whose complete per-item token-sequence collections are identical within a model–benchmark–condition combination. All23,034planned generations are present, no two global replicate collections are identical, and the largest incidental item-level exact-match rate within a combination is 0.6%. The primary candidate-logit readouts and controls use local Transformers/PyTorch forward passes rather than vLLM’s returned log probabilities. Effort-conditioned gpt-oss traces are generated through the chat template with thinking enabled and the reasoning-effort setting low, medium, or high. In the released template this rendersReasoning: low|medium|highbefore the user message and assistant generation prompt. Prompted Qwen3 traces append the tested textual instruction to the system prompt and otherwise leave generation unconstrained, so the external horizonBis applied identically to prompted and unprompted trajectories at probe time. The returned-top-20readout is retained only for the fixed-stop calibration (Figure A1). The numeric/concision prompt is present during treated generation and absent during ordinary generation. Prompt- control generations replace it with the alternatives in Table A2, using the same imposed horizon and the same ordinary-generation baseline within that control cohort. The policy-context readout reconstructs each arm from its stored source-prompt IDs, so the generation-time instruction remains visible when an active prefix is probed. The common-context readout instead uses the ordinary base-prompt IDs for every Qwen arm, replays the arm’s exact reasoning IDs, and appends the same neutral closure. Completed reasoning blocks are also replayed under this common context; their original final answers are retained only in the terminal-aware policy-context estimand. Budget-present controls add the numeric/concision prompt only at answer time while holding the source prefix fixed. The token horizon is defined with the source model’s tokenizer, so absolute horizons are paired within model and remain descriptive across model families. A.3 Trace parsing and forced-answer probes Qwen3 reasoning is the token sequence between the thinking tags. gpt-oss reasoning is the analysis-channel token sequence, including the channel marker needed to form a valid replay. For the primary token-exact paths, we locate the terminal boundary in the stored completion IDs, exclude all following final-answer IDs, and concatenate the chosen prompt IDs, the firstBreasoning IDs or the complete reasoning block, and the closure IDs. No generated reasoning prefix is decoded and retokenized. We verify the replayed prompt tokens against the prompts used at generation, check boundary and special-token placement, and record hashes of the source completion and the complete replay input. For gpt-oss at the lower-effort matched horizonL c , a high-effort trace is replayable when it contains at leastL c stored reasoning IDs, whether or not it later reaches a terminal boundary, or when a terminal boundary identifies a completed reasoning block beforeL c . A stored sequence that ends beforeL c without such a boundary is retained among the matched pairs but has an unavailable high-effort outcome. This rule includes capped high-effort traces whenever their stored reasoning IDs reach the matched horizon. The primary analysis uses all replayable pairs; the high-active diagnostic retains only the first case. The nine-arm Qwen mechanism experiment follows the same token-exact construction atB ∈512, 2,048. All36,000 generations and72,000policy-context replays are present, with no probe errors or context truncations. Cohorts without 14 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces stored token IDs use text reconstruction: the reasoning text is tokenized with the source model, cut to the requested horizon, decoded, and embedded in the probe prompt. They do not supply primary token-exact evidence. The candidate-logit probe contains the single-letter system instruction, question, replayed reasoning prefix, and a family-specific suffix that closes reasoning. For gpt-oss, the suffix closes the analysis channel and opens the final channel. One deterministic forward pass extracts full-vocabulary logits at the first answer position. The valid set contains every supported one-token bare-letter and leading-space spelling; spellings for the same answer are aggregated by log-sum-exp. Ifp y is the resulting unnormalized probability for optiony, we report valid-option massM = P y∈Y p y and the conditional distributionq y = p y /M. The predicted answer isarg max y q y , conditional log loss is− logq y ⋆ , and multiclass Brier score is P y (q y − 1y = y ⋆ ) 2 . Lower log loss and Brier score are better. Qwen policy-context proper-score contrasts are restricted to paired active prefixes because a retained natural terminal answer has no associated probe distribution. Common-context terminal replay supplies probe distributions for completed reasoning blocks as well. Each candidate-logit probe records the valid answer-token IDs, missing IDs, coverage, and any probe error. Terminal- prefix agreement is reported in Table A4. In a separate single-run cohort, none of the2,788terminal gpt-oss replay prefixes contains a final-channel delimiter (Table A30); this check does not exhaustively cover the three-replicate primary cohort. The returned-top-20calibration is incomplete by construction and supports no primary estimate (Figure A1). For the nine Qwen instruction arms and the central gpt-oss comparisons, we also append the appropriate final-channel closure to the exact stored prefix and greedily generate up to16unconstrained answer tokens. Missing or malformed answer letters count as incorrect and as parse failures. These continuations decode answers deterministically from independently sampled reasoning prefixes, so the reasoning replicate is the stochastic unit. Terminal-prefix readouts reproduce the naturally sampled final answer for98.3–100.0%of completed prefixes in the terminal-agreement check. The Omni-MATH-2 analyses use free-form answers judged withgpt-5-mini-2025-08-07, following Ballon et al. [34]. The Qwen comparison uses a200-item subset and includes item–horizon comparisons for which both prompt and surprise answers were judged. The gpt-oss comparison draws500items using Omni-MATH’s native difficulty metadata [33]; the analyzed model-specific paired cohorts contain 475 items for gpt-oss-20b and 490 for gpt-oss-120b. All gpt-oss prefixes are replayed under a common medium-effort final-answer context, and the analysis includes all 15,440judged answers across the matched-horizon and fixed-checkpoint comparisons. These comparisons use one rollout per condition and therefore do not characterize variation across sampled generations. An independent full-trace judging pass agrees on94–99%of items across the five models. Structural-marker annotations use the same snapshot with the released JSON-schema prompt. A.4 Readout contexts and answer availability We run policy- and common-context readouts to separate the full interface contrast from differences carried by the generated reasoning prefix. The Qwen policy-context readout preserves the stored condition-specific prompt. The Qwen common-context readout uses the ordinary base prompt and neutral closure for every arm and replays completed as well as active reasoning blocks. A paired answer-context control then reintroduces the numeric/concision prompt only at answer time while holding the bundled-prompt prefix fixed. For gpt-oss, the primary matched-horizon readout uses a common medium-effort answer context for every low-, medium-, and high-source prefix. Additional controls cross the source effort with the effort setting used only at readout. We treat candidate-logit scoring as an option-normalized, deterministic paired instrument for answer availability at a standardized prefix. Greedy final-answer continuations provide the corresponding behavioral readout. A source-context branch readout reconstructs the source prompt and reasoning prefix at the answer point instead of placing the prefix in a neutral context. Qwen surprise-stop traces keep the budget absent, Qwen bundled-prompt traces retain the numeric-budget instruction, and gpt-oss traces retain the source effort setting. This text-based recomputation approximates the trajectory’s original prompt context rather than recovering the generation-time KV cache. To test when answers become available in the trajectory, we run a dense readout over the generated prefixes at checkpoints0, 64,..., 512. For gpt-oss, the source prefixes come from low, medium, and high effort traces, and all are read with a common medium-effort readout. For Qwen3-14B, the source prefixes compare bundled-prompt and surprise-stop traces at theB = 512condition, and the readout uses a neutral budget-free closure. At each checkpoint we extract candidate-letter probabilities from full-vocabulary next-token logits and compute the gold-answer log-odds, logp(y ⋆ )− log P y̸=y ⋆ p(y) . Natural termination is treated as an absorbing state. For a trace that terminates at length L ic < 512, we carry its terminal forced-answer distribution forward, so thatq ic (t) = q ic (L ic )for everyt≥ L ic . The resulting terminal-aware answer-availability AUC is a token-normalized trapezoidal average that includes both evidence accumulated before termination and the practical consequence of having completed the answer early. We also report 15 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces retrospective stable correctness, which counts a checkpoint as correct only if the predicted answer remains correct at all later checkpoints through 512. As a sensitivity analysis, we recompute the terminal-aware answer-availability AUC over0–128tokens after retaining only items for which both compared traces are still reasoning after128tokens (L > 128). We call this an active-pair sensitivity analysis rather than a separately identified causal estimate, because conditioning on survival changes the target comparison. A.5 Literal Qwen cap-then-answer validation The literal Qwen validation follows the documented two-stage thinking-budget shape while retaining the paired item– replicate design of the candidate-logit analysis. The design crosses GPQA Diamond (198items) and the primary MMLU-Pro subset (500items) withB ∈512, 2,048, hidden/surprise and numeric/concision-visible arms, and the same three generation replicates. In the first call, vLLM generates the reasoning block with the Qwen sampling settings and an external maximum of B tokens, stopping early if </think> appears. If generation hits the cap, we append the Qwen limited-time closure and close the thinking block; otherwise we retain the naturally closed state. In the second call, the model greedily generates up to16answer tokens. Malformed or missing choice labels are parse failures and count as incorrect. We store the full reasoning- and answer-stage token IDs, hash the reasoning and answer stages together, and reject a benchmark–arm–budget combination if two complete replicate collections are record-identical across its items. Accuracy contrasts pair arms within item, replicate, and budget. Primary intervals average the three observed replicate outcomes within item and bootstrap items; replicate-specific, two-way item×replicate, and crossed-intercept calculations are sensitivities. Candidate-logit deltas are included only as a readout comparison. A.6 Scripted positive control To check that the same-imposed-horizon readout detects a known early-answer trajectory, we construct a scripted synthetic positive control with96balanced four-choice items. The question contains no answer information. All options are generic labels A–D, and the gold label is assigned by a deterministic balanced key. The long policy uses a neutral reasoning prefix in which the candidate-answer line appears after the first64source tokens but before512source tokens. The short policy uses a prefix beginning withcandidate answer: X, whereXis the gold label. A shortB = 512 policy reuses the long trace, so the positive control should be large at64tokens and vanish once both traces reveal the candidate answer. We generate probes directly from these scripted reasoning prefixes, truncate them with the Qwen tokenizer, and apply the same candidate-logit answer-token readout used for the paired multiple-choice contrasts and controls. Accuracy moves from25.0%for the long policy to100.0%for the short policy atB = 64; both reach100.0%atB = 512 (Table A29). This positive control validates the truncation and readout on a known text-level policy difference. It is not evidence that Qwen3-14B or gpt-oss spontaneously condition on exact numeric budgets. A.7 Trajectory diagnostics and controls To test whether a condition is legible in the early reasoning, we train a logistic-regression classifier to predict the condition from the firstkreasoning tokens of a trace, fork ∈64, 128, 256, 512, 1024, 2048. The input is the surface text of the firstkgenerated reasoning tokens and nothing else: no question, no answer, no trace length, and no model identity. Traces shorter thankare excluded at thatkrather than padded, so termination or residual length is not provided as an implicit feature. We use two feature sets, TF-IDF over wordn-grams (1–3) and charactern-grams (3–5), and a set of length-normalized style counts (backtracking and verification cue words, punctuation, operator and digit density, answer-letter mentions, mean word length). The label is the effort setting (low, medium, high) for gpt-oss and the bundled-prompt condition (bundled prompt versus surprise-stop) for Qwen3. We evaluate with5-fold cross-validation grouped by item, so the same question never appears in both train and test, and compare against a label-shuffle baseline of 20 permutations. We analyze the generated reasoning text rather than treating prompt-boundary hidden-state decodability as evidence that the reasoning policy changed. The effort or budget instruction is present in the prompt, whereas the surface text is generated by the policy itself. We therefore use text classifiers for the early-policy diagnostic. To characterize how the early reasoning differs across conditions, we compute length-normalized structural markers on the firstmin(256,L)reasoning tokens of each trace: the position of the first computational cue, the position of the first answer assertion, the rates of backtracking, verification, and alternative-strategy cue phrases per thousand words, characters per word, and digit and operator density. Raw trace length is not among the reported markers. Markers are compared between paired conditions with bootstrap intervals over items on the same GPQA Diamond and stratified 16 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces MMLU-Pro analysis subsets used for the main probe contrasts. These are deterministic lexical patterns and capture surface form rather than meaning. To validate them, we annotate a stratified sample of250trace excerpts (50per condition) with GPT-5 mini for the same constructs and correlate the cue-based markers with the annotations (n=204 matched). The backtracking and alternative-strategy markers agree with the annotations (Spearmanρ=0.32and0.19, respectively; bothp < 0.01), whereas the verification marker is weaker and not significant (point-biserialr=0.11, p = 0.109), and the two position-based markers, restatement length and first-answer commitment, do not validate. We base our structural interpretation on the digit and operator density and the alternative-strategy marker, which are either objective character statistics or annotation-supported. To test whether the recovered accuracy is instance-specific, we re-probe each surprise-stop prefix against three length- matched controls at the same budgetB: a random prefix ofBtokens sampled uniformly from the vocabulary excluding special tokens; a swap prefix taken as the firstBreasoning tokens of a different question’s trace; and a shuffle prefix that permutes the firstBoriginal reasoning tokens. Swap donors must be distinct and contain at leastBsource tokens. We exclude31swap prefixes whose donor traces are shorter thanBand restrict all pooled contrasts to the remaining common four-arm item–budget cohort. The controls isolate, respectively, context length and position, coherent but irrelevant reasoning, and token identity without order. A.8 Statistical inference The primary Qwen curve and gpt-oss contrasts use three independently sampled global generation replicates. Conditions are paired within item and replicate. Primary intervals condition on the three trajectory sets observed in this experiment: we first average the three paired replicate outcomes within item and then draw10,000bootstrap resamples of items. Bootstrap resampling uses NumPy seed0by default, and every reported interval uses a fixed, recorded resampling seed. We report each replicate separately with an item bootstrap. A paired two-way item×replicate bootstrap and a crossed item–replicate variance decomposition are sensitivity analyses. With only three global replicate levels, those sensitivities do not provide precise population-over-seed inference. For the model-based sensitivity analysis, we fit the paired differences with crossed item and replicate random intercepts. Variance components are estimated by method of moments on complete item–replicate matrices; uncertainty for the grand mean uses the item, replicate, and residual components with a Satterthwaitetapproximation. With three global replicates, this calculation is a diagnostic of sensitivity to trajectory-set variation rather than a precise variance- component analysis. The nine-arm Qwen mechanism experiment uses eight item-specific replicates for each of500MMLU-Pro items. A deterministic hash of the item ID, replicate number, and base seed defines one generation seed for each item–replicate pair; the same seed is reused across all nine prompt arms, and all4,000item–replicate seeds are distinct. The arms are hidden horizon; numeric/concision prompts announcing512or2,048tokens; advance exact-stop notices announcing the same two horizons; generic limited-budget wording; number-only wording; a concise/early-answer instruction; and an irrelevant-number placebo. Primary intervals use a nested paired bootstrap that first resamples items and then resamples the eight paired replicates within each selected item. The arm–horizon family is covered by simultaneous intervals obtained from the maximum absolute deviation across the treatment-versus-ordinary contrasts in each bootstrap draw. Direct prompt-to-prompt contrasts use the same paired draws. The direct diagonal statistic averages the advantage of announcing512rather than2,048tokens when the imposed stop is512with the reverse contrast when the imposed stop is 2,048. A positive value would indicate that matching the announced and imposed horizons helps. For the three-replicate Qwen accuracy curve, each item’s paired outcomes are first averaged over the observed replicates. The primary bootstrap resamples items and preserves each selected item’s complete budget curve. The simultaneous 95%band uses the95th percentile of the maximum absolute deviation from the estimated curve. A scalar summary integrates the accuracy difference overlog 2 Bwith trapezoidal weighting and divides by the observed log-budget span; its percentile interval is computed from the same item-bootstrap draws. A two-way item×replicate version of both summaries is retained as a sensitivity analysis. The±3percentage-point margin is a descriptive scale, not a preregistered utility threshold or an equivalence test. For the gpt-oss matched-horizon comparison, we retain every matched item–replicate pair with an observed lower-effort terminal outcome and bound the effect of unreplayable high-effort traces. LetO c andM c denote the replayable and missing-high sets, letN c =|O c | +|M c |, and write the observed completed lower-effort outcome asY c,N ir . Worst-case binary bounds set every missing high-effort outcome first to correct and then to incorrect: τ c G = P (i,r)∈O c D c ir + P (i,r)∈M c (Y c,N ir − 1) N c , τ c G = P (i,r)∈O c D c ir + P (i,r)∈M c Y c,N ir N c . 17 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Several gpt-oss and Qwen run pairs have duplicate decoded response strings. Because their token IDs and seeds are unavailable, we cannot determine whether the underlying token sequences are identical or reconstruct the cause; these runs are used only for run-level diagnostics. Prompt-control, readout-context, crossed-horizon, and subset- replication cohorts without stored token IDs or verified seeds are excluded from the token-exact primary estimates. The Omni-MATH-2 comparisons use one rollout per condition. Single-run binary matched-pair checks use item-bootstrap intervals. Where McNemar’s test is reported, it is exact when the discordant count is below25and otherwise uses the continuity-correctedχ 2 approximation [39]. Repeated-budget prefix-control summaries first average paired outcomes over budgets within item and then use an item-cluster bootstrap. Every budget curve reports the total item count, the number surviving toB, the survival rate, and both survivor-only and terminal-aware policy accuracy. A.9 Replication, interface comparisons, and diagnostic analyses A.9.1 Text-reconstructed Qwen subset replication and prompt comparisons The text-reconstructed MMLU-Pro subset replication compares the500-item subset used in the main design with a second category-stratified500-item subset that excludes those items. Both subsets compare the numeric/concision and surprise-stop conditions atB = 512over three nominal runs, and the independent subset also compares the numeric/concision and concise/early-answer conditions with the same surprise-stop cohort atB = 2,048. Accuracy and survival use the candidate-logit readout. Intervals resample items and nominal runs in a paired two-way bootstrap. Generation seeds and completion token IDs are unavailable for these cohorts, so the estimates describe the observed nominal runs and do not support generalization across generation seeds. Only aggregate estimates and intervals are available for the twoB = 2,048rows, so they cannot be regenerated independently and are treated as descriptive evidence (Table A24). The text-reconstructed prompt-control comparison evaluates eight Qwen3-14B instructions against the same surprise- stop cohort atB = 512on GPQA Diamond and MMLU-Pro. Each estimate averages three nominal runs within item before an item-cluster bootstrap. Table A25 reports terminal-aware accuracy and median full-trace length; paired log-length contrasts are included in the companion data release. Completion token IDs and verified generation seeds are unavailable for these cohorts. A.9.2 Numeric wording with gpt-oss effort The within-model gpt-oss comparison crosses the numeric-budget instruction with low, medium, and high effort; B ∈512, 2,048, 8,192, 16,384; both gpt-oss model sizes; and GPQA Diamond or MMLU-Pro. The directly matched comparison adds the numeric wording to each effort setting and evaluates both conditions atB = 512(Table A28). Longer announced-budget comparisons lack a directly matched effort-only baseline because the effort-only comparison covers only B = 512, and low-effort traces often terminate before later checkpoints. A.9.3 Wrong-prefix continuation The wrong-prefix continuation analysis selects natural Qwen3-14B traces with an incorrect parsed final answer, retains the firstBsource-tokenized reasoning tokens, and presents that prefix with the same question. The answer-now condition closes reasoning immediately with the forced-answer suffix. Continuation conditions instead allowR additional reasoning tokens before the forced answer, with the remaining budget either hidden or stated in the prompt. Both conditions use the returned-top-20readout, whose incomplete option coverage applies symmetrically. We evaluate B ∈ 512, 2,048, 8,192andR ∈ 512, 2,048, 8,192, 32,768on GPQA Diamond and the stratified MMLU-Pro subset. The displayed R = 32,768 conditions are limited by the remaining context window (Table A26). A.9.4 Activation readout of the announced budget The activation analysis tests whether Qwen3-14B hidden states retain the announced numeric budget. For each item with one natural trace and all four numeric-budget traces, we feed the prompt and the firstk ∈0, 64, 256, 512, 1024, 2048 generated reasoning tokens. We random-project hidden states to256dimensions at layers0,10,20,30, and39, then fit logistic-regression readouts with item-grouped cross-validation. One readout predicts whether the trace uses the numeric/concision prompt; the other predicts which of the four announced budgets generated a treated trace. Each is compared with20label-shuffle baselines. A focused check uses100shuffles andC ∈0.1, 1, 10for the post-start exact-budget comparisons (Table A27). 18 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces B Figures The appendix figures are organized by the question each analysis addresses. Figures A1 to A3 characterize readout sensitivity and natural trace lengths and test whether recovered information depends on the question-specific prefix. Figures A4 to A5 test horizon matching and show when answer evidence becomes available. Figures A6 to A7 compare completed-trace and fixed-checkpoint gpt-oss accuracy, while Figures A8 to A9 characterize how effort and prompting alter early traces. The crossed-horizon Qwen cohorts in Figure A4 lack stored completion token IDs and verified generation seeds. They provide a horizon-pattern check and do not enter the token-exact effect estimates. 10 3 10 4 imposed horizon B (reasoning tokens) 0.2 0.4 0.6 0.8 terminal-aware policy accuracy GPQA Diamond a Qwen3-4B Qwen3-8B Qwen3-14B gpt-oss-20b gpt-oss-120b 10 3 10 4 imposed horizon B (reasoning tokens) 0.2 0.4 0.6 0.8 MMLU-Pro b 10 3 10 4 imposed horizon B (reasoning tokens) 0.0 0.2 0.4 0.6 0.8 1.0 survival P ( L > B ) c GPQA Diamond MMLU-Pro Figure A1: A returned-top-20readout reproduces the qualitative fixed-stop pattern. Terminal-aware policy accuracy (a, GPQA Diamond; b, MMLU-Pro) and survival (c) for surprise-stop probes under a returned-top-20readout, which takes the argmax over the option letters present in the returned top-20log-probabilities and has incomplete option-letter coverage (Section A). The full-vocabulary candidate-logit readout (Figure 2) is used for the main fixed-stop calibration and all primary multiple-choice estimates. 10 0 10 1 10 2 10 3 10 4 reasoning tokens 0.00 0.25 0.50 0.75 1.00 ECDF GPQA Diamond a 10 0 10 1 10 2 10 3 10 4 reasoning tokens MMLU-Pro b 10 0 10 1 10 2 10 3 10 4 10 5 reasoning tokens Omni-MATH-2 c Qwen3-4BQwen3-8BQwen3-14Bgpt-oss-20bgpt-oss-120b Figure A2: Natural reasoning-length distributions by model and benchmark. Empirical cumulative distribution of reasoning-trace token length for the five models on GPQA Diamond, MMLU-Pro, and Omni-MATH-2. Natural length varies severalfold across models and benchmarks, which sets where the imposed-horizon curves of Figure 2 saturate. Panel c includes980/4,011Qwen3-8B and223/4,039Qwen3-14B responses that immediately close an empty<think></think>block. We count each empty block as one source-model token and exclude the following final-answer text from reasoning length. 19 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces original swapped trace shu ffl ed tokens random tokens 0.0 0.1 0.2 0.3 0.4 accuracy (pooled over budgets) 0.39 0.21 0.33 0.31 gpt-oss-20b original swapped trace shu ffl ed tokens random tokens 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.52 0.34 0.43 0.39 Qwen3-14B Figure A3: Truncation-recovered accuracy is instance-specific: the original prefix beats length-matched controls. Accuracy of the surprise-stop probe under the original budget-Bprefix and three length-matched controls, pooled over budgets on the common four-arm item–budget cohort, for gpt-oss-20b and Qwen3-14B on GPQA Diamond and MMLU- Pro. The swapped-trace control (coherent reasoning from a different question) is the weakest, below the shuffled-token and random-token controls, and all sit below the original. Inference averages paired original-minus-control outcomes over budgets within item and uses an item-cluster bootstrap plus item-level sign-flip randomization test. 512 2k 8k 16k announced horizon H +1.0-1.7-0.8-0.2 -0.3-1.9-0.8-1.0 -1.0-1.7-2.9-3.0 -1.0-0.2+0.8+1.9 Numeric/concision, GPQA Diamond a +2.2+1.8-0.3-0.7 +2.9+1.1-0.8-1.3 +2.8+1.8-0.2-0.7 +2.0+2.2-0.5-0.6 Numeric/concision, MMLU-Pro b 512 2k8k16k imposed horizon B 512 2k 8k 16k announced horizon H +0.7+0.5-2.8-1.3 -1.3-0.8-3.5-1.7 -1.9-2.2-1.5-1.7 -0.8-0.8-2.5-1.9 Advance exact-stop notice, GPQA Diamond c 512 2k8k16k imposed horizon B +1.3+1.2-1.2-2.1 +2.4+1.6-0.2-0.5 +2.1+2.0-0.4-0.5 +2.7+1.0-0.6-1.0 Advance exact-stop notice, MMLU-Pro d −4 −3 −2 −1 0 1 2 3 4 prompt-control minus surprise (p) Figure A4: Crossing announced and imposed horizons shows no diagonal matching-number pattern. Heatmaps show prompt-control minus surprise-stop accuracy deltas in percentage points for Qwen3-14B under the candidate-logit readout. Rows vary the announced horizonHin the prompt, columns vary the imposed horizonB, and dotted diagonals markH = B. a,b, Numeric/concision prompt for GPQA Diamond and MMLU-Pro. c,d, Advance exact-stop notice for the same benchmarks. MMLU-Pro positives concentrate at early imposed horizons, while GPQA remains small or negative across the grid. Completion token IDs and verified generation seeds are unavailable for these four-budget and GPQA cohorts; the paired MMLU-Pro experiment likewise finds no clear matching-number benefit at two imposed stops (Table A7). 20 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces 0 1 2 3 4 5 6 7 8 Δ gold log-odds gpt-oss-20b a 0 1 2 3 4 5 6 7 8 gpt-oss-120b b 0 1 2 3 4 5 6 7 8 Qwen3-14B c low minus high medium minus high bundled prompt minus surprise 0128256384512 checkpoint tokens −0.05 0.00 0.05 0.10 0.15 0.20 Δ stable correctness d 0128256384512 checkpoint tokens −0.05 0.00 0.05 0.10 0.15 0.20 e 0128256384512 checkpoint tokens −0.05 0.00 0.05 0.10 0.15 0.20 f GPQA Diamond MMLU-Pro Figure A5: Shorter gpt-oss effort settings make answers available earlier under the candidate-logit readout. Curves show paired condition differences at dense checkpoints throughB = 512; solid lines are GPQA Diamond, dotted lines are MMLU-Pro, and shaded bands are pointwise95%paired-bootstrap intervals over items. a–c, Difference in option-normalized gold-answer log-odds for gpt-oss-20b, gpt-oss-120b, and Qwen3-14B, computed from full- vocabulary next-token logits over every valid answer-token ID. d–f, Difference in retrospective stable correctness, where a checkpoint is counted correct only if all later checkpoints through512remain correct. The gpt-oss contrasts compare low or medium source prefixes against high source prefixes under a medium-effort readout; Qwen compares bundled-prompt and surprise prefixes under a neutral budget-free closure. 21 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces 10 3 10 4 0.4 0.5 0.6 0.7 0.8 accuracy gpt-oss-20b, GPQA Diamond a low (run to completion) medium (run to completion) high (run to completion) high prefix cut to low/medium length 10 2 10 3 0.4 0.5 0.6 0.7 0.8 gpt-oss-20b, MMLU-Pro b 10 3 median reasoning tokens 0.4 0.5 0.6 0.7 0.8 accuracy gpt-oss-120b, GPQA Diamond c 10 2 10 3 median reasoning tokens 0.4 0.5 0.6 0.7 0.8 gpt-oss-120b, MMLU-Pro d Figure A6: Completed gpt-oss effort settings trade accuracy against median trace length. Results for gpt-oss-20b (a,b) and gpt-oss-120b (c,d) on GPQA Diamond and the stratified MMLU-Pro subset pool three generation replicates. When run to completion, the three effort settings define the observed relationship, while high-effort readouts at the itemwise low- or medium-effort terminal horizon sit below it; a completed high-effort reasoning block is replayed when it finishes before that horizon. Because the horizontal axis uses medians, this figure describes typical trace length rather than arithmetic-mean token use or expected API cost. 22 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces low minus high medium minus high gpt-oss-20b, GPQA Diamond a low minus high medium minus high gpt-oss-20b, MMLU-Pro b −50510152025 accuracy difference at B= 512 (p) low minus high medium minus high gpt-oss-120b, GPQA Diamond c −50510152025 accuracy difference at B= 512 (p) low minus high medium minus high gpt-oss-120b, MMLU-Pro d Figure A7: Lower gpt-oss effort remains more accurate at the same 512-token checkpoint. Accuracy differences compare low or medium effort with high effort for gpt-oss-20b (a,b) and gpt-oss-120b (c,d) on GPQA Diamond and MMLU-Pro. Naturally stopped traces retain their final answer. Primary95%intervals average the three paired replicate outcomes within item and bootstrap items. Positive values favor the lower-effort setting. 23 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces lowmediumhigh 0.00 0.01 0.02 0.03 restatement length (frac. before first work cue) lowmediumhigh 0.0 0.2 0.4 0.6 0.8 time to first answer (frac. of trace) lowmediumhigh 0.0 0.5 1.0 1.5 2.0 2.5 3.0 backtracking (per 1k words) lowmediumhigh 0.00 0.25 0.50 0.75 1.00 1.25 1.50 verification (per 1k words) lowmediumhigh 0.00 0.05 0.10 0.15 0.20 0.25 0.30 alternatives (per 1k words) lowmediumhigh 0 2 4 6 8 chars per word lowmediumhigh 0.00 0.02 0.04 0.06 digit density lowmediumhigh 0.00 0.01 0.02 0.03 0.04 0.05 operator density gpt-oss-20b structural markers by effort (GPQA + MMLU-Pro subset, first 256 reasoning tokens) Figure A8: Structural markers of the first256reasoning tokens by gpt-oss-20b effort setting. Each panel is one length-normalized marker, averaged over GPQA Diamond and the stratified MMLU-Pro subset with95%bootstrap intervals over items, for low (blue), medium (green), and high (orange) effort. Digit density, operator density, and characters per word decrease monotonically from low to high effort, while the explicit alternative-strategy marker rate increases. Together, these surface measures show that shorter-effort traces allocate more of their opening tokens to calculation-like text and contain fewer explicit alternatives. The time-to-first-answer marker is near-saturated, and neither position-based marker validates against GPT-5 mini annotations; the backtracking cue rate agrees with the annotations but is not monotone across effort, and the verification marker does not validate in the agreement analysis on this subset. We interpret the density and alternative-strategy markers, which are objective or annotation-supported (Section A). 6412825651210242048 prefix tokens 0.3 0.4 0.5 0.6 0.7 balanced accuracy gpt-oss effort Qwen bundled prompt Figure A9: Condition classifiers distinguish gpt-oss effort more strongly than Qwen bundled prompting. Balanced accuracy of a classifier predicting the condition from the firstkreasoning tokens, for gpt-oss effort (three-class) and Qwen3-14B bundled-prompt versus surprise-stop (two-class), with dotted label-shuffle baselines. Traces shorter thank are excluded at that prefix length, so the classifier is a diagnostic of condition legibility among surviving traces rather than a causal estimate. The effort setting is strongly legible from the opening tokens; the bundled-prompt signal is weaker and decays to chance by 512 tokens. 24 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces C Tables The appendix tables report the fine-grained estimates needed to check the primary matched-horizon results. Tables A1 to A4 define the estimands and interventions and validate the fixed-stop readouts, while Tables A6 to A16 report the Qwen and gpt-oss estimates and sensitivity analyses. Tables A17 to A21 give dense trajectory, termination-stratum, answer-context, arm-level, and score-tail detail; Tables A5 and A22 report single-rollout comparisons for Qwen3-8B and Omni-MATH-2 and characterize sampled trajectories. Qwen subset-replication and prompt-control results from cohorts without verified token IDs or seeds are in Tables A23 to A25; readout validations, wrong-prefix continuation, activation readouts, and the prompt–effort interaction are in Tables A26 to A30. These tables broaden item or prompt coverage and address measurement or diagnostic questions; the primary matched-horizon claims rest on the token-exact experiments above. Table A1: We evaluate each condition at a common stopping point and separate completion from prefix quality. The terminal-aware readout retains a natural answer after completion and probes an unfinished prefix in its source context. Common-context replay applies the same answer context to both conditions; both-active comparisons further restrict the analysis to jointly unfinished pairs. Behavioral continuations generate an unconstrained short answer from the stopped prefix. ComparisonWhere the comparison stopsHow the answer is measuredMain result Qwen textual treatment: visible numeric/concision prompt (v) vs ordinary generation (h) Common imposed horizon B; retain either arm’s natural answer if it completed before B Terminal-aware policy-context readout; τ Q,pol (B) = E[Y v,T pol (B)− Y h,T pol (B)] Shorter traces; no broad accuracy gain at the same imposed horizon Qwen prefix diagnostics: tested instruction (a) vs ordinary generation (h) Common imposed horizon B; replay both prefixes under the ordinary base prompt Correct completion by B, common-context terminal replay, both-active contrast, and four-stratum decomposition Separates stopping and commitment from differences between unfinished prefixes Qwen literal cap: visible numeric/concision prompt vs hidden horizon Common external cap B ∈512, 2048; both arms then receive an answer continuation Greedy generated-answer continuation; E[Y v,C (B)− Y h,C (B)] Benchmark split checked without option normalization gpt-oss effort: completed c∈low, medium vs high effort Itemwise lower-effort terminal length L c ir ; replay every reconstructable high prefix, including early completions All-replayable common-context probe; τ c G,R =E[D c | R c = 1], with worst-case bounds Lower effort reaches correct terminal answers earlier gpt-oss active diagnosticThe same matched pairs restricted to high effort still active at L c ir Common-context candidate-logit probe Separates the policy-cost result from early high-effort completion gpt-oss continuation check: the same matched prefixes The same all-replayable L c ir pairs Greedy unconstrained final-channel answer;E[Y c,C −Y high,C | R c = 1] Tests whether the ordering survives a generated-answer continuation 25 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A2: Verbatim generation and readout interventions. Braces indicate the interpolated token budget. ConditionMessage roleExact text or template Qwen numeric/concision prompt system, appended to base instruction You have at most B tokens inside the <think>...</think> reasoning block. Use a strategy appropriate for this budget. When you have enough evidence, close the reasoning block and answer with only the final answer. Qwen limited-budget control system, appended to base instruction You have a limited token budget inside the <think>...</think> reasoning block. Use a strategy appropriate for this limit. When you have enough evidence, close the reasoning block and answer with only the final answer. Qwen number-only control system, appended to base instruction You have at most B tokens inside the <think>...</think> reasoning block. Qwen advance exact-stop control system, appended to base instruction Your <think>...</think> reasoning block will be stopped after exactly B tokens. Qwen approximate-budget control system, appended to base instruction Aim to use approximately B tokens inside the <think>...</think> reasoning block. Qwen irrelevant-number control system, appended to base instruction The control number for this problem is B. Do not treat this number as a token budget. Qwen concise/early-answer control system, appended to base instruction Use a very concise reasoning strategy. Prioritize reaching a usable answer quickly. Qwen thorough semantic control system, appended to base instruction Use a thorough reasoning strategy. Explore alternatives and check your answer carefully before finalizing. Qwen limited-time closure continuation Considering the limited time by the user, I have to give the solution based on the thinking directly now. </think> Qwen paraphrased closure continuation Based on my reasoning so far, my answer is: </think> Qwen neutral closurecontinuation </think> Qwen budget-present readout system, answer-time only You have at most B tokens inside the <think>...</think> reasoning block. Use a strategy appropriate for this budget. When you have enough evidence, close the reasoning block and answer with only the final answer. Qwen answer instruction system Solve the following problem. Please make sure that your response only consists of a single letter corresponding to the correct answer choice. Do not include anything else in your final response. gpt-oss effortchat-template kwarg/system scaffold reasoning_effort = low | medium | high, rendered by the snapshot chat template as Reasoning: low|medium|high in the system message before the developer/user messages 26 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A3: Fixed-stop calibration separates completed answers from the selected set of surviving traces. The full-vocabulary candidate-logit readout is evaluated atB ∈ 0, 512, 2,048, 8,192, 16,384. Survival isP (L > B), survivor-only accuracy conditions onL > B, and terminal-aware accuracy retains the natural answer after completion. The natural full-trace reference uses the same cohort. Rate columns report percentages with95%item-bootstrap intervals. Accuracy omits missing parsed outcomes; n retains the full cohort. ModelBenchmarkB nSurvival (%) Survivor-only accuracy (%) Terminal-aware policy accuracy (%) Natural full-trace accuracy (%) Qwen3-4BGPQA Diamond0 198 100.0 [100.0, 100.0]31.8 [25.3, 38.4]31.8 [25.3, 38.4]55.6 [48.5, 62.2] Qwen3-4BGPQA Diamond512 198 100.0 [100.0, 100.0]34.8 [28.3, 41.4]34.8 [28.3, 41.4]55.6 [48.5, 62.2] Qwen3-4BGPQA Diamond2,048 19885.9 [80.8, 90.4]37.6 [30.6, 44.7]43.4 [36.9, 50.5]55.6 [48.5, 62.2] Qwen3-4BGPQA Diamond8,192 19823.2 [17.7, 29.3]50.0 [34.8, 65.2]56.1 [49.5, 62.6]55.6 [48.5, 62.2] Qwen3-4BGPQA Diamond 16,384 1982.5 [0.5, 5.1]40.0 [0.0, 80.0]54.8 [47.7, 61.9]55.6 [48.5, 62.2] Qwen3-4BMMLU-Pro0 500 100.0 [100.0, 100.0]39.0 [34.8, 43.2]39.0 [34.8, 43.2]69.5 [65.5, 73.5] Qwen3-4BMMLU-Pro512 50085.4 [82.2, 88.4]43.3 [38.9, 48.0]49.4 [45.2, 53.8]69.5 [65.5, 73.5] Qwen3-4BMMLU-Pro2,048 50035.2 [31.0, 39.4]44.3 [36.9, 51.7]64.3 [60.1, 68.5]69.5 [65.5, 73.5] Qwen3-4BMMLU-Pro8,192 5008.0 [5.6, 10.4]27.5 [15.0, 42.5]68.3 [64.3, 72.3]69.5 [65.5, 73.5] Qwen3-4BMMLU-Pro16,384 5001.0 [0.2, 2.0]0.0 [0.0, 0.0]69.3 [65.3, 73.3]69.5 [65.5, 73.5] Qwen3-8BGPQA Diamond0 198 100.0 [100.0, 100.0]30.8 [24.2, 37.4]30.8 [24.2, 37.4]57.9 [50.8, 64.5] Qwen3-8BGPQA Diamond512 19898.0 [96.0, 99.5]35.6 [28.9, 42.3]36.9 [30.3, 43.4]57.9 [50.8, 64.5] Qwen3-8BGPQA Diamond2,048 19886.4 [81.3, 90.9]35.7 [28.7, 42.7]42.9 [35.9, 49.5]57.9 [50.8, 64.5] Qwen3-8BGPQA Diamond8,192 19827.8 [21.7, 34.3]30.9 [18.2, 43.6]56.9 [49.7, 63.5]57.9 [50.8, 64.5] Qwen3-8BGPQA Diamond 16,384 1984.0 [1.5, 7.1]37.5 [0.0, 75.0]57.9 [50.8, 64.5]57.9 [50.8, 64.5] Qwen3-8BMMLU-Pro0 500 100.0 [100.0, 100.0]43.0 [38.8, 47.4]43.0 [38.8, 47.4]72.2 [68.1, 76.0] Qwen3-8BMMLU-Pro512 50086.8 [83.8, 89.8]46.5 [41.7, 51.2]52.6 [48.2, 57.0]72.2 [68.1, 76.0] Qwen3-8BMMLU-Pro2,048 50043.0 [38.8, 47.4]43.7 [37.2, 50.2]65.1 [60.9, 69.3]72.2 [68.1, 76.0] Qwen3-8BMMLU-Pro8,192 50010.0 [7.4, 12.8]34.0 [22.0, 48.0]70.2 [66.2, 74.2]72.2 [68.1, 76.0] Qwen3-8BMMLU-Pro16,384 5001.8 [0.8, 3.0]33.3 [0.0, 66.7]71.8 [67.7, 75.8]72.2 [68.1, 76.0] Qwen3-14BGPQA Diamond0 198 100.0 [100.0, 100.0]37.4 [30.8, 43.9]37.4 [30.8, 43.9]62.6 [55.6, 69.2] Qwen3-14BGPQA Diamond512 19899.0 [97.5, 100.0]41.8 [35.2, 49.0]42.4 [35.8, 49.5]62.6 [55.6, 69.2] Qwen3-14BGPQA Diamond2,048 19876.8 [70.7, 82.3]44.1 [36.2, 52.0]53.5 [46.5, 60.6]62.6 [55.6, 69.2] Qwen3-14BGPQA Diamond8,192 19818.7 [13.6, 24.2]29.7 [16.2, 45.9]59.1 [52.5, 66.2]62.6 [55.6, 69.2] Qwen3-14BGPQA Diamond 16,384 1980.0 [0.0, 0.0]—62.6 [56.1, 69.2]62.6 [55.6, 69.2] Qwen3-14BMMLU-Pro0 500 100.0 [100.0, 100.0]50.4 [46.0, 54.8]50.4 [46.0, 54.8]76.8 [73.0, 80.4] Qwen3-14BMMLU-Pro512 50080.6 [77.0, 84.0]52.4 [47.4, 57.1]60.0 [55.8, 64.4]76.8 [73.0, 80.4] Qwen3-14BMMLU-Pro2,048 50031.0 [27.0, 35.0]50.3 [42.6, 58.1]70.8 [66.8, 74.8]76.8 [73.0, 80.4] Qwen3-14BMMLU-Pro8,192 5005.8 [3.8, 8.0]51.7 [34.5, 69.0]76.6 [72.8, 80.2]76.8 [73.0, 80.4] Qwen3-14BMMLU-Pro16,384 5001.0 [0.2, 2.0]0.0 [0.0, 0.0]76.4 [72.8, 80.2]76.8 [73.0, 80.4] gpt-oss-20bGPQA Diamond0 198 100.0 [100.0, 100.0]40.9 [33.8, 48.0]40.9 [33.8, 48.0]62.6 [55.6, 69.2] gpt-oss-20bGPQA Diamond512 19883.8 [78.3, 88.9]42.2 [34.3, 49.4]50.0 [42.9, 57.1]62.6 [55.6, 69.2] gpt-oss-20bGPQA Diamond2,048 19843.9 [37.4, 51.0]36.8 [26.4, 47.1]60.6 [53.5, 67.2]62.6 [55.6, 69.2] gpt-oss-20bGPQA Diamond8,192 1989.1 [5.6, 13.1]38.9 [16.7, 61.1]63.6 [56.6, 70.2]62.6 [55.6, 69.2] gpt-oss-20bGPQA Diamond 16,384 1982.5 [0.5, 5.1]60.0 [20.0, 100.0]63.6 [57.1, 70.2]62.6 [55.6, 69.2] gpt-oss-20bMMLU-Pro0 500 100.0 [100.0, 100.0]43.4 [39.2, 47.8]43.4 [39.2, 47.8]72.9 [68.9, 76.8] gpt-oss-20bMMLU-Pro512 50040.0 [35.8, 44.2]35.5 [29.0, 42.0]64.1 [59.9, 68.3]72.9 [68.9, 76.8] gpt-oss-20bMMLU-Pro2,048 50010.0 [7.4, 12.6]24.0 [12.0, 36.0]71.1 [67.1, 75.2]72.9 [68.9, 76.8] gpt-oss-20bMMLU-Pro8,192 5000.4 [0.0, 1.0]0.0 [0.0, 0.0]72.9 [69.1, 76.8]72.9 [68.9, 76.8] gpt-oss-20bMMLU-Pro16,384 5000.0 [0.0, 0.0]—72.9 [69.1, 76.8]72.9 [68.9, 76.8] gpt-oss-120bGPQA Diamond0 198 100.0 [100.0, 100.0]38.9 [32.3, 46.0]38.9 [32.3, 46.0]69.7 [63.1, 75.8] gpt-oss-120bGPQA Diamond512 19870.2 [63.6, 76.8]36.0 [28.1, 43.9]51.5 [44.4, 58.6]69.7 [63.1, 75.8] gpt-oss-120bGPQA Diamond2,048 19823.7 [18.2, 29.8]40.4 [25.5, 53.2]67.7 [61.1, 74.2]69.7 [63.1, 75.8] gpt-oss-120bGPQA Diamond8,192 1980.5 [0.0, 1.5]0.0 [0.0, 0.0]69.7 [63.1, 75.8]69.7 [63.1, 75.8] gpt-oss-120bGPQA Diamond 16,384 1980.0 [0.0, 0.0]—69.7 [63.1, 75.8]69.7 [63.1, 75.8] gpt-oss-120bMMLU-Pro0 500 100.0 [100.0, 100.0]48.4 [44.0, 52.8]48.4 [44.0, 52.8]78.4 [74.8, 81.8] gpt-oss-120bMMLU-Pro512 50027.2 [23.4, 31.2]40.4 [32.4, 48.5]71.2 [67.2, 75.2]78.4 [74.8, 81.8] gpt-oss-120bMMLU-Pro2,048 5004.2 [2.6, 6.0]42.9 [23.8, 66.7]78.0 [74.4, 81.6]78.4 [74.8, 81.8] gpt-oss-120bMMLU-Pro8,192 5000.0 [0.0, 0.0]—78.4 [74.8, 82.0]78.4 [74.8, 81.8] gpt-oss-120bMMLU-Pro16,384 5000.0 [0.0, 0.0]—78.4 [74.8, 82.0]78.4 [74.8, 81.8] 27 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A4: Completed-prefix readouts agree with natural answers in98.3–100.0%of cases. Each row compares the candidate-logit answer from a replayed completed reasoning prefix with the parsed natural answer from the same trace. This separate cohort validates terminal replay; it does not assess coverage of the three-replicate token-exact analysis. ModelBenchmarkSource policyComparedAgreeAgreement gpt-oss-120bGPQA Diamondgpt-oss low effort198198100.0% gpt-oss-120bGPQA Diamondgpt-oss medium effort198198100.0% gpt-oss-120bMMLU-Progpt-oss low effort499499100.0% gpt-oss-120bMMLU-Progpt-oss medium effort500500100.0% gpt-oss-20bGPQA Diamondgpt-oss low effort196196100.0% gpt-oss-20b GPQA Diamondgpt-oss medium effort196196100.0% gpt-oss-20bMMLU-Progpt-oss low effort49649499.6% gpt-oss-20bMMLU-Progpt-oss medium effort499499100.0% Qwen3-14BGPQA DiamondQwen numeric/concision prompt 1,3171,31599.8% Qwen3-14B GPQA DiamondQwen surprise58858699.7% Qwen3-14BMMLU-ProQwen numeric/concision prompt 4,4514,44799.9% Qwen3-14BMMLU-ProQwen surprise1,4891,46398.3% Table A5: Open-ended continuations show no consistent Qwen prompt gain and a gpt-oss matched-horizon advantage. GPT-5 mini judges free-form Omni-MATH-2 answers. Qwen rows compare numeric/concision-prompt and surprise-stop prefixes at the same imposed horizon on a200-item sample; each row includes only item–horizon comparisons for which both answers were judged. gpt-oss rows compare a completed lower-effort reasoning block with high effort at the itemwise lower-effort terminal horizon, replaying a completed high-effort reasoning block when it finishes before that horizon. Every reasoning block is read under a common medium-effort final-answer context. The one-rollout, difficulty-stratified500-item draw yields model-specific paired cohorts of 475 items for gpt-oss-20b and 490 for gpt-oss-120b. The analysis includes all15,440judged gpt-oss answers across the matched-horizon and fixed-checkpoint comparisons. Intervals bootstrap items. ComparisonModelContrastStop nTreated/lower acc.Comparator acc.∆ [95% CI] Qwen promptQwen3-14Bprompt minus surprise51217022.9%19.4%+3.5 [−1.2, +8.2] Qwen promptQwen3-14Bprompt minus surprise2,04815033.3%31.3%+2.0 [−3.3, +7.3] Qwen promptQwen3-14Bprompt minus surprise8,1927326.0%30.1% −4.1 [−12.3, +4.1] gpt-oss effortgpt-oss-20blow minus high prefixmatched47547.8%14.3%+33.5 [+28.8, +37.9] gpt-oss effortgpt-oss-20b medium minus high prefixmatched47561.5%27.6%+33.9 [+28.8, +38.9] gpt-oss effortgpt-oss-120blow minus high prefixmatched49061.2%19.2%+42.0 [+37.3, +46.7] gpt-oss effortgpt-oss-120bmedium minus high prefixmatched49070.6%33.1%+37.6 [+32.4, +42.7] 28 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A6: The numeric/concision prompt does not consistently improve Qwen accuracy at the same horizon. Panel A compares the numeric/concision (visible-horizon) and ordinary (hidden-horizon) conditions at the same imposed horizon using token-exact replay and the terminal-aware policy readout for all198GPQA Diamond and 500MMLU-Pro items across three generation replicates;ngives pointwise/complete-curve item counts.∆is visible minus hidden accuracy; intervals and simultaneous bands average replicates within item before bootstrapping items. AUC integrates∆overlog 2 B. Panel B applies the literal cap-then-answer continuation and reports parse coverage as visible/hidden. Panel A: terminal-aware policy readout ModelBenchmarkBnReplicatesVisible acc.Hidden acc.∆ [95% CI]Simultaneous 95% band AUC [95% CI] Qwen3-14BGPQA Diamond 512198/198341.8%42.3% −0.5 [−2.5, +1.3][−4.7, +3.7]+0.5 [−1.6, +2.7] Qwen3-14BGPQA Diamond 2,048198/198353.2%52.5%+0.7 [−3.0, +4.4][−3.5, +4.9]— Qwen3-14BGPQA Diamond 8,192198/198360.3%59.6%+0.7 [−2.4, +3.9][−3.5, +4.9]— Qwen3-14BGPQA Diamond 16,384198/198362.0%60.6%+1.3 [−1.9, +4.5][−2.9, +5.6]— Qwen3-14BMMLU-Pro512500/500360.5%58.5%+2.0 [+0.4, +3.7][+0.1, +3.9]+0.4 [−0.7, +1.5] Qwen3-14BMMLU-Pro2,048500/500371.7%71.0%+0.7 [−0.9, +2.1][−1.3, +2.6]— Qwen3-14BMMLU-Pro8,192500/500375.1%75.9% −0.9 [−2.5, +0.7][−2.8, +1.1]— Qwen3-14BMMLU-Pro16,384500/500375.7%75.6%+0.1 [−1.2, +1.3][−1.9, +2.0]— Panel B: literal cap-then-answer continuation ModelBenchmarkBnReplicatesVisible acc.Hidden acc.∆ [95% CI]Parse coverage V/H Policy-readout ∆ Qwen3-14BGPQA Diamond 512198339.9%41.8% −1.9 [−4.9, +1.0]100.0%/100.0% −0.5 Qwen3-14BGPQA Diamond 2,048198353.9%53.4%+0.5 [−2.7, +3.7]99.8%/99.8%+0.7 Qwen3-14BMMLU-Pro512500360.7%58.7%+2.0 [+0.2, +3.9]97.1%/96.7%+2.0 Qwen3-14BMMLU-Pro2,048500371.4%71.3%+0.1 [−1.3, +1.5]99.5%/97.5%+0.7 Table A7: At the 512-token stop, the concise/early-answer instruction gives the largest accuracy gain; matching the stated number to the stop shows no clear benefit. a, Qwen3-14B results on the500-item MMLU-Pro subset with eight paired replicates per item. “Finished” is the share of runs that produce a natural answer byB; otherwise, accuracy uses the exact stored prefix and standardized forced-answer readout. Differences are relative to no advance notice, with nested item–replicate bootstrap intervals. b, Direct tests of whether matching the stated number to the imposed stop helps. a Nine prompt instructions Imposed stop B = 512Imposed stop B = 2,048 Prompt instructionNumber stated HMedian lengthFinishedAccuracyDifference [95% CI] FinishedAccuracyDifference [95% CI] No advance notice—1,08320.3%59.7%—72.1%72.1%— Numeric/concision prompt 51290025.0%60.1%+0.4 [−1.2, +2.0]77.6%72.0% −0.1 [−1.4, +1.2] Numeric/concision prompt 2,04892123.5%60.5%+0.8 [−0.8, +2.4]77.1%72.1% −0.1 [−1.4, +1.3] Advance notice of exact stop 51285725.4%61.3%+1.6 [+0.1, +3.1]78.4%72.6%+0.5 [−0.7, +1.7] Advance notice of exact stop 2,04890023.0%60.7%+1.0 [−0.4, +2.4]77.3%72.3%+0.2 [−1.1, +1.4] Limited budget (no number) —87425.0%60.1%+0.4 [−1.2, +2.0]77.1%71.8% −0.3 [−1.6, +1.0] Number only51292424.1%60.9%+1.2 [−0.3, +2.7]77.1%72.6%+0.5 [−0.8, +1.7] Concise/early-answer instruction —67936.4%63.5%+3.8 [+2.1, +5.6]82.3%73.1%+1.0 [−0.4, +2.4] Irrelevant number5121,02021.2%60.0%+0.3 [−1.2, +1.8]73.0%72.0% −0.1 [−1.2, +1.0] b Direct test of matching the stated number to the stop Prompt instructionMatch advantage (p)95% CI Numeric/concision prompt−0.2[−1.0, +0.6] Advance notice of exact stop+0.1[−0.6, +0.8] 29 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A8: AtB = 512, the concise/early-answer contrast combines earlier completion with better unfinished prefixes. a, Correct completion, all-pair accuracy, and both-active accuracy for concise/early-answer versus ordinary generation. Both-active rows condition on pairs in which neither trace has finished. b, The same 4,000 item–replicate pairs partitioned into four termination strata under terminal-aware policy scoring and common replay. Contributions sum to the corresponding all-pair difference. Intervals use a nested item–replicate bootstrap. a Arm-level estimands EstimandReadout contextBItemsPairsOrdinaryConcise/early-answer∆ [95% CI] Correct natural completionReadout-context invariant5125004,00018.2%31.9%+13.8 [+11.6, +16.0] Correct natural completionReadout-context invariant2,0485004,00058.0%65.6%+7.6 [+5.8, +9.5] All-pair accuracyTerminal-aware policy5125004,00059.7%63.5%+3.8 [+2.1, +5.6] All-pair accuracyTerminal-aware policy2,0485004,00072.1%73.1%+1.0 [−0.4, +2.4] All-pair accuracyCommon replay context5125004,00059.8%62.9%+3.1 [+1.5, +4.8] All-pair accuracyCommon replay context2,0485004,00072.0%73.0%+1.0 [−0.4, +2.4] Both-active accuracyTerminal-aware policy5123892,47645.2%49.2%+4.0 [+1.6, +6.5] Both-active accuracyTerminal-aware policy2,04813264440.2%41.0%+0.8 [−4.7, +6.0] Both-active accuracyCommon replay context5123892,47645.4%48.2%+2.7 [+0.6, +5.0] Both-active accuracyCommon replay context2,04813264440.1%40.5%+0.5 [−4.8, +5.5] b Termination-stratum decomposition Readout contextBTermination stratumShareWithin-stratum ∆ [95% CI]Contribution [95% CI] Terminal-aware policy512Neither trace finished61.9%+4.0 [+1.6, +6.5]+2.5 [+1.0, +4.0] Terminal-aware policy512Concise/early-answer finished; ordinary active 17.8%+8.6 [+4.9, +12.6]+1.5 [+0.9, +2.3] Terminal-aware policy512Ordinary finished; concise/early-answer active 1.7% −7.5 [−20.3, +3.8]−0.1 [−0.4, +0.1] Terminal-aware policy512Both traces finished18.7% −0.3 [−1.4, +0.8]−0.1 [−0.3, +0.2] Terminal-aware policy2,048Neither trace finished16.1%+0.8 [−4.7, +6.0]+0.1 [−0.8, +1.0] Terminal-aware policy2,048Concise/early-answer finished; ordinary active 11.8%+6.2 [+0.2, +12.1]+0.7 [+0.0, +1.4] Terminal-aware policy2,048Ordinary finished; concise/early-answer active 1.7% −16.7 [−34.3, +1.3]−0.3 [−0.6, +0.0] Terminal-aware policy2,048Both traces finished70.5%+0.6 [−0.4, +1.6]+0.4 [−0.3, +1.2] Common replay context 512Neither trace finished61.9%+2.7 [+0.6, +5.0]+1.7 [+0.4, +3.1] Common replay context 512Concise/early-answer finished; ordinary active 17.8%+8.6 [+4.9, +12.6]+1.5 [+0.9, +2.3] Common replay context 512Ordinary finished; concise/early-answer active 1.7% −6.0 [−19.4, +6.5]−0.1 [−0.3, +0.1] Common replay context 512Both traces finished18.7% −0.3 [−1.4, +0.8]−0.1 [−0.3, +0.2] Common replay context 2,048Neither trace finished16.1%+0.5 [−4.8, +5.5]+0.1 [−0.8, +0.9] Common replay context 2,048Concise/early-answer finished; ordinary active 11.8%+6.2 [+0.2, +12.1]+0.7 [+0.0, +1.4] Common replay context 2,048Ordinary finished; concise/early-answer active 1.7% −16.7 [−34.5, +1.4]−0.3 [−0.6, +0.0] Common replay context 2,048Both traces finished70.5%+0.6 [−0.4, +1.7]+0.5 [−0.3, +1.2] 30 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A9: The 512-token concise/early-answer advantage persists under multiplicity control, direct prompt com- parisons, and generated-answer continuation. a, Concise/early-answer minus ordinary generation with simultaneous 95%intervals over the16instruction–deadline contrasts in each readout context. b, Direct comparisons atB = 512. c, Greedy generated-answer continuations from the stored prefixes. Panels b and c use pointwise nested item–replicate bootstrap intervals; each cell contains 500 items and eight paired replicates per item. a Familywise simultaneous inference Readout contextB∆ (p)Simultaneous 95% CI Terminal-aware policy512+3.8[+1.7, +6.0] Terminal-aware policy2,048+1.0[−1.2, +3.1] Common replay context512+3.1[+1.1, +5.1] Common replay context2,048+1.0[−1.0, +3.0] b Direct prompt contrasts at B = 512 Readout contextComparatorConcise/early-answer acc.Comparator acc.∆ [95% CI] Terminal-aware policyNumeric/concision prompt, H = 51263.5%60.1%+3.4 [+1.7, +5.2] Terminal-aware policyNumeric/concision prompt, H = 2,04863.5%60.5%+3.0 [+1.3, +4.8] Terminal-aware policyAdvance notice of exact stop, H = 51263.5%61.3%+2.3 [+0.5, +4.0] Common replay contextNumeric/concision prompt, H = 51262.9%60.7%+2.2 [+0.7, +3.8] Common replay contextNumeric/concision prompt, H = 2,04862.9%60.6%+2.4 [+0.9, +3.9] Common replay contextAdvance notice of exact stop, H = 51262.9%60.6%+2.3 [+0.8, +3.9] c Greedy generated-answer continuation Readout contextBConcise/early-answer acc.Ordinary acc.∆ [95% CI] Terminal-aware policy51263.2%59.1%+4.1 [+2.4, +5.9] Terminal-aware policy2,04873.1%70.9%+2.2 [+0.7, +3.7] Common replay context51262.6%59.1%+3.5 [+1.9, +5.2] Common replay context2,04872.8%70.9%+1.9 [+0.4, +3.4] Table A10: The numeric/concision prompt shortens Qwen reasoning but weakly tracks the announced token count. a, Budget-specific reasoning length and adherence. Visible denotes the numeric/concision condition and hidden denotes matched ordinary generation.L/BandP (L ≤ B)use visible-condition lengths; the log-length contrast is logL visible − logL hidden . b, Responsiveness across the four announced budgets. The slope and Spearman correlation are computed within each item–replicate trajectory. Intervals first average the three observed generation replicates within item and then bootstrap items. a Budget-specific length and adherence ModelBenchmarkBItemsItem–replicate pairsVisible median LHidden median LMedian L/B P (L≤ B)Mean ∆ logL [95% CI] Qwen3-14BGPQA Diamond 5121985943,740.54,486.57.311.9% −0.179 [−0.215, −0.144] Qwen3-14BGPQA Diamond 2,0481985943,647.04,486.51.7829.6% −0.176 [−0.211, −0.142] Qwen3-14BGPQA Diamond 8,1921985943,698.54,486.50.4590.6% −0.159 [−0.196, −0.121] Qwen3-14BGPQA Diamond 16,3841985943,925.04,486.50.2499.2% −0.133 [−0.168, −0.099] Qwen3-14BMMLU-Pro5125001,500884.51,094.01.7323.6% −0.189 [−0.214, −0.165] Qwen3-14BMMLU-Pro2,0485001,500880.51,094.00.4377.5% −0.170 [−0.195, −0.145] Qwen3-14BMMLU-Pro8,1925001,500929.01,094.00.1196.4% −0.134 [−0.159, −0.109] Qwen3-14BMMLU-Pro16,3845001,500905.01,094.00.0699.6% −0.130 [−0.154, −0.106] b Across-budget within-item responsiveness ModelBenchmarkMetricItemsEstimate [95% CI] Qwen3-14BGPQA Diamond Within-item slope of logL on logB198+0.012 [+0.004, +0.021] Qwen3-14BGPQA Diamond Within-item Spearman ρ(B,L)198+0.058 [+0.013, +0.103] Qwen3-14BMMLU-ProWithin-item slope of logL on logB500+0.018 [+0.013, +0.024] Qwen3-14BMMLU-ProWithin-item Spearman ρ(B,L)500+0.121 [+0.091, +0.150] 31 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A11: Replicate-specific estimates and crossed-intercept analyses preserve the absence of a broad Qwen gain and the positive gpt-oss effort differences. Pooled rows first average the three observed generation replicates within item and then bootstrap items. The three named replicate rows report separate paired item-bootstrap estimates and 95% intervals. Two-way intervals resample items and replicates. Crossed intervals use method-of-moments item and replicate random intercepts with a Satterthwaite approximation on complete item–replicate matrices, so crossed point estimates can differ from the primary∆when pairs are ineligible. Qwen effects are visible minus hidden; gpt-oss effects are the completed lower-effort readout minus the matched-horizon high-effort readout. All effects are percentage points. StudyModelBenchmarkComparisonReplicateItems Item-bootstrap ∆ [95% CI] Two-way item×replicate ∆ [95% CI] Crossed ∆ [95% CI] Qwen textualQwen3-14BGPQA Diamond Exact, B = 512 Pooled (3)198−0.5 [−2.5, +1.3] −0.5 [−3.4, +2.2] −0.5 [−2.7, +1.7] Run 41; seed 74,017198−1.0 [−5.1, +3.0] — Run 42; seed 74,029198 +0.5 [−3.0, +4.0] — Run 43; seed 74,047198−1.0 [−4.5, +2.5] — Qwen textualQwen3-14BGPQA Diamond Exact, B = 2,048 Pooled (3)198 +0.7 [−3.0, +4.4] +0.7 [−4.0, +5.7] +0.7 [−3.3, +4.7] Run 41; seed 74,017198−1.5 [−6.6, +3.5] — Run 42; seed 74,029198 +0.0 [−5.1, +5.1] — Run 43; seed 74,047198 +3.5 [−2.0, +9.6] — Qwen textualQwen3-14BGPQA Diamond Exact, B = 8,192 Pooled (3)198 +0.7 [−2.4, +3.9] +0.7 [−3.4, +5.1] +0.7 [−2.5, +3.8] Run 41; seed 74,017198 +1.0 [−4.5, +6.6] — Run 42; seed 74,029198−0.5 [−5.1, +4.0] — Run 43; seed 74,047198 +1.5 [−4.0, +7.1] — Qwen textualQwen3-14BGPQA Diamond Exact, B = 16,384 Pooled (3)198 +1.3 [−1.9, +4.5] +1.3 [−2.9, +5.7] +1.3 [−1.8, +4.5] Run 41; seed 74,017198 +3.0 [−2.0, +8.1] — Run 42; seed 74,029198 +0.0 [−5.1, +5.1] — Run 43; seed 74,047198 +1.0 [−4.5, +6.6] — Qwen textualQwen3-14BMMLU-ProExact, B = 512 Pooled (3)500 +2.0 [+0.4, +3.7] +2.0 [−0.1, +4.1] +2.0 [+0.3, +3.7] Run 41; seed 74,017500 +2.6 [+0.2, +5.2] — Run 42; seed 74,029500 +1.6 [−1.0, +4.2] — Run 43; seed 74,047500 +1.8 [−1.0, +4.6] — Qwen textualQwen3-14BMMLU-ProExact, B = 2,048 Pooled (3)500 +0.7 [−0.9, +2.1] +0.7 [−1.6, +2.9] +0.7 [−1.1, +2.5] Run 41; seed 74,017500 +0.8 [−1.8, +3.4] — Run 42; seed 74,029500 +2.0 [−0.2, +4.4] — Run 43; seed 74,047500−0.8 [−3.0, +1.6] — Qwen textualQwen3-14BMMLU-ProExact, B = 8,192 Pooled (3)500−0.9 [−2.5, +0.7] −0.9 [−3.1, +1.5] −0.9 [−2.5, +0.7] Run 41; seed 74,017500−1.0 [−3.6, +1.6] — Run 42; seed 74,029500 +0.4 [−2.2, +3.0] — Run 43; seed 74,047500−2.0 [−4.6, +0.4] — Qwen textualQwen3-14BMMLU-ProExact, B = 16,384 Pooled (3)500 +0.1 [−1.2, +1.3] +0.1 [−2.3, +2.7] +0.1 [−2.5, +2.6] Run 41; seed 74,017500−1.0 [−3.6, +1.6] — Run 42; seed 74,029500 +2.2 [−0.4, +4.8] — Run 43; seed 74,047500−1.0 [−3.0, +1.0] — Qwen literalQwen3-14BGPQA Diamond Continuation, B = 512 Pooled (3)198−1.9 [−4.9, +1.0] −1.9 [−5.4, +1.7] −1.9 [−4.8, +1.1] Run 41; seed 74,017198−3.0 [−6.6, +0.5] — Run 42; seed 74,029198−1.5 [−6.1, +3.0] — Run 43; seed 74,047198−1.0 [−5.1, +3.0] — Qwen literalQwen3-14BGPQA Diamond Continuation, B = 2,048 Pooled (3)198 +0.5 [−2.7, +3.7] +0.5 [−3.9, +5.4] +0.5 [−3.4, +4.5] Run 41; seed 74,017198−1.5 [−6.1, +3.0] — Run 42; seed 74,029198−0.5 [−5.1, +4.0] — Run 43; seed 74,047198 +3.5 [−1.5, +8.6] — Qwen literalQwen3-14BMMLU-ProContinuation, B = 512 Pooled (3)500 +2.0 [+0.2, +3.9] +2.0 [−0.4, +4.5] +2.0 [+0.2, +3.8] Run 41; seed 74,017500 +2.0 [−0.6, +4.6] — Run 42; seed 74,029500 +0.8 [−2.0, +3.6] — Run 43; seed 74,047500 +3.2 [+0.2, +6.2] — Qwen literalQwen3-14BMMLU-ProContinuation, B = 2,048 Pooled (3)500 +0.1 [−1.3, +1.5] +0.1 [−2.3, +2.4] +0.1 [−2.1, +2.3] Run 41; seed 74,017500−1.0 [−3.4, +1.6] — Run 42; seed 74,029500−0.8 [−3.2, +1.6] — Continued on next page 32 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A11: Replicate-specific estimates and crossed-intercept analyses. Continued. StudyModelBenchmarkComparisonReplicateItems Item-bootstrap ∆ [95% CI] Two-way item×replicate ∆ [95% CI] Crossed ∆ [95% CI] Run 43; seed 74,047500 +2.0 [−0.2, +4.4] — gpt-oss exactgpt-oss-20bGPQA Diamond lowPooled (3)198 +14.5 [+7.7, +21.2] +14.5 [+6.1, +22.7] +14.5 [+6.9, +22.0] Run 41; seed 74,017198 +13.6 [+5.1, +22.2] — Run 42; seed 74,029198 +10.6 [+1.5, +19.7] — Run 43; seed 74,047198 +19.2 [+10.6, +27.8] — gpt-oss exactgpt-oss-20bGPQA Diamond mediumPooled (3)198 +17.4 [+11.3, +23.7] +17.4 [+9.6, +25.8] +17.8 [+11.1, +24.4] Run 41; seed 74,017197 +13.7 [+5.6, +22.3] — Run 42; seed 74,029198 +17.7 [+9.6, +25.8] — Run 43; seed 74,047198 +21.2 [+12.6, +29.8] — gpt-oss exactgpt-oss-20bMMLU-ProlowPooled (3)500 +20.9 [+16.7, +25.0] +20.9 [+16.2, +25.5] +20.9 [+16.8, +25.1] Run 41; seed 74,017500 +19.4 [+14.2, +24.4] — Run 42; seed 74,029500 +21.2 [+16.2, +26.0] — Run 43; seed 74,047500 +22.2 [+17.2, +27.0] — gpt-oss exactgpt-oss-20bMMLU-PromediumPooled (3)500 +18.0 [+14.5, +21.5] +18.0 [+13.7, +22.3] +18.0 [+14.5, +21.5] Run 41; seed 74,017500 +17.2 [+12.4, +21.8] — Run 42; seed 74,029500 +17.0 [+12.2, +21.8] — Run 43; seed 74,047500 +19.8 [+14.8, +24.6] — gpt-oss exactgpt-oss-120bGPQA Diamond lowPooled (3)198 +23.6 [+16.3, +30.6] +23.6 [+15.7, +31.5] +23.6 [+16.3, +30.9] Run 41; seed 74,017198 +20.7 [+12.1, +29.3] — Run 42; seed 74,029198 +26.3 [+18.2, +34.3] — Run 43; seed 74,047198 +23.7 [+15.7, +31.8] — gpt-oss exactgpt-oss-120bGPQA Diamond mediumPooled (3)198 +23.6 [+16.7, +30.5] +23.6 [+16.2, +31.0] +23.6 [+16.6, +30.5] Run 41; seed 74,017198 +23.7 [+15.6, +31.8] — Run 42; seed 74,029198 +24.2 [+16.2, +32.3] — Run 43; seed 74,047198 +22.7 [+14.6, +30.8] — gpt-oss exactgpt-oss-120bMMLU-ProlowPooled (3)500 +26.3 [+21.9, +30.5] +26.3 [+21.7, +30.7] +26.3 [+22.0, +30.6] Run 41; seed 74,017500 +27.4 [+22.8, +32.0] — Run 42; seed 74,029500 +25.4 [+20.6, +30.0] — Run 43; seed 74,047500 +26.0 [+21.0, +30.8] — gpt-oss exactgpt-oss-120bMMLU-PromediumPooled (3)500 +24.2 [+20.3, +28.2] +24.2 [+19.6, +28.9] +24.2 [+19.9, +28.5] Run 41; seed 74,017500 +21.8 [+17.4, +26.2] — Run 42; seed 74,029500 +25.8 [+21.2, +30.6] — Run 43; seed 74,047500 +25.0 [+20.2, +29.8] — Continued on next page 33 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A11: Replicate-specific estimates and crossed-intercept analyses. Continued. StudyModelBenchmarkComparisonReplicateItems Item-bootstrap ∆ [95% CI] Two-way item×replicate ∆ [95% CI] Crossed ∆ [95% CI] gpt-oss continuation gpt-oss-20bGPQA Diamond lowPooled (3)198 +15.5 [+8.8, +22.1] +15.5 [+7.6, +23.2] +15.5 [+8.8, +22.2] Run 41; seed 74,017198 +16.7 [+8.6, +24.7] — Run 42; seed 74,029198 +12.1 [+3.5, +20.7] — Run 43; seed 74,047198 +17.7 [+9.1, +26.3] — gpt-oss continuation gpt-oss-20bGPQA Diamond mediumPooled (3)198 +17.4 [+11.1, +23.8] +17.4 [+9.8, +25.3] +17.8 [+11.4, +24.1] Run 41; seed 74,017197 +14.7 [+6.1, +23.4] — Run 42; seed 74,029198 +17.2 [+9.1, +25.3] — Run 43; seed 74,047198 +20.7 [+12.1, +29.3] — gpt-oss continuation gpt-oss-20bMMLU-ProlowPooled (3)500 +20.5 [+16.2, +24.5] +20.5 [+15.9, +24.9] +20.5 [+16.3, +24.7] Run 41; seed 74,017500 +19.4 [+14.2, +24.4] — Run 42; seed 74,029500 +20.6 [+15.6, +25.6] — Run 43; seed 74,047500 +21.4 [+16.4, +26.0] — gpt-oss continuation gpt-oss-20bMMLU-PromediumPooled (3)500 +18.2 [+14.7, +21.7] +18.2 [+13.9, +22.5] +18.2 [+14.8, +21.6] Run 41; seed 74,017500 +17.2 [+12.4, +21.8] — Run 42; seed 74,029500 +17.4 [+12.6, +22.0] — Run 43; seed 74,047500 +20.0 [+15.0, +24.8] — gpt-oss continuation gpt-oss-120bGPQA Diamond lowPooled (3)198 +23.1 [+15.8, +30.3] +23.1 [+15.2, +30.8] +23.1 [+15.8, +30.3] Run 41; seed 74,017198 +20.2 [+11.6, +28.8] — Run 42; seed 74,029198 +24.2 [+16.2, +32.3] — Run 43; seed 74,047198 +24.7 [+16.7, +32.8] — gpt-oss continuation gpt-oss-120bGPQA Diamond mediumPooled (3)198 +23.1 [+16.2, +30.0] +23.1 [+15.8, +30.3] +23.1 [+16.2, +29.9] Run 41; seed 74,017198 +23.2 [+15.2, +31.3] — Run 42; seed 74,029198 +23.2 [+15.2, +31.3] — Run 43; seed 74,047198 +22.7 [+14.6, +30.8] — gpt-oss continuation gpt-oss-120bMMLU-ProlowPooled (3)500 +25.9 [+21.5, +30.2] +25.9 [+21.3, +30.4] +25.9 [+21.6, +30.3] Run 41; seed 74,017500 +26.2 [+21.4, +31.0] — Run 42; seed 74,029500 +26.4 [+21.4, +31.2] — Run 43; seed 74,047500 +25.2 [+20.4, +30.0] — gpt-oss continuation gpt-oss-120bMMLU-PromediumPooled (3)500 +24.1 [+20.1, +28.0] +24.1 [+19.5, +28.7] +24.1 [+19.9, +28.3] Run 41; seed 74,017500 +22.0 [+17.6, +26.4] — Run 42; seed 74,029500 +26.0 [+21.4, +30.6] — Run 43; seed 74,047500 +24.2 [+19.4, +29.0] — 34 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A12: Selected answer-context checks do not reverse the main interpretation. Qwen neutral replay applies the same budget-free closure to every prefix, including completed traces. The second block holds the source prefix fixed and changes only the answer-time context. All rows use full-vocabulary candidate logits and paired item-bootstrap intervals. Limited-time, paraphrased, and neutral Qwen closure wording gives the same qualitative pattern (Table A19). FamilyContrastBenchmarkB n∆ (p)95% CI Neutral replay contrasts Qwen3-14Bprompt vs. surprise, neutral replay GPQA Diamond 512198 −1.3[−5.4, +2.5] Qwen3-14Bprompt vs. surprise, neutral replay GPQA Diamond 2,048198 −0.3[−4.4, +3.7] Qwen3-14Bprompt vs. surprise, neutral replay GPQA Diamond 8,192198 −3.0[−6.1, +0.0] Qwen3-14Bprompt vs. surprise, neutral replay GPQA Diamond 16,384198+1.9[−1.6, +5.6] Qwen3-14Bprompt vs. surprise, neutral replay MMLU-Pro512500+1.9[+0.1, +3.7] Qwen3-14Bprompt vs. surprise, neutral replay MMLU-Pro2,048500+1.1[−0.5, +2.7] Qwen3-14Bprompt vs. surprise, neutral replay MMLU-Pro8,192500+1.0[−0.6, +2.7] Qwen3-14Bprompt vs. surprise, neutral replay MMLU-Pro16,384500+0.8[−0.9, +2.5] Direct effects of the readout context Qwen3-14Bbudget present vs. absent at readout GPQA Diamond 512198 −2.9[−5.9, +0.0] Qwen3-14Bbudget present vs. absent at readout GPQA Diamond 2,048198 −0.5[−2.9, +1.9] Qwen3-14Bbudget present vs. absent at readout MMLU-Pro512500 −0.4[−1.5, +0.7] Qwen3-14Bbudget present vs. absent at readout MMLU-Pro2,048500+0.1[−0.2, +0.3] gpt-oss-20blow vs. medium readout, high source GPQA Diamond 512191+2.1[−1.6, +6.3] gpt-oss-20blow vs. medium readout, high source MMLU-Pro512361+1.7[+0.0, +3.6] gpt-oss-120blow vs. medium readout, high source GPQA Diamond 512188+0.0[−2.7, +2.7] gpt-oss-120blow vs. medium readout, high source MMLU-Pro512335+0.3[−2.1, +2.7] 35 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A13: Qwen proper scores show no consistent advantage for the numeric/concision prompt at 512 or 2,048 tokens. Visible denotes the numeric/concision condition and hidden the matched hidden-horizon condition.∆is visible minus hidden. Higher valid-option mass is preferred; lower conditional log loss and Brier score are preferred. Rows include token-exact candidate-logit pairs from three generation replicates. Arm means and differences first average replicates within item, after which intervals bootstrap items. ModelBenchmarkBMetricItemsItem–replicate pairsReplicatesVisibleHidden∆ [95% CI] Qwen3-14BGPQA Diamond 512Valid-option mass 197580399.6%99.5%+0.1 [−0.2, +0.3] Qwen3-14BGPQA Diamond 512Conditional log loss 19758033.3103.021+0.289 [+0.142, +0.459] Qwen3-14BGPQA Diamond 512Conditional Brier score 19758030.9420.918+0.024 [+0.001, +0.048] Qwen3-14BGPQA Diamond 2,048Valid-option mass 146394399.1%99.1%+0.1 [−0.3, +0.4] Qwen3-14BGPQA Diamond 2,048Conditional log loss 14639432.8692.738+0.131 [−0.176, +0.470] Qwen3-14BGPQA Diamond 2,048 Conditional Brier score 14639430.9180.923 −0.005 [−0.076, +0.070] Qwen3-14BMMLU-Pro512Valid-option mass 4061,085396.9%96.5%+0.4 [−0.1, +0.9] Qwen3-14BMMLU-Pro512Conditional log loss 4061,08534.2344.351 −0.117 [−0.324, +0.075] Qwen3-14BMMLU-Pro512 Conditional Brier score 4061,08530.8220.861 −0.039 [−0.078, −0.003] Qwen3-14BMMLU-Pro2,048Valid-option mass 132299393.9%93.8%+0.1 [−1.5, +1.7] Qwen3-14BMMLU-Pro2,048Conditional log loss 13229933.6453.811 −0.167 [−0.666, +0.341] Qwen3-14BMMLU-Pro2,048 Conditional Brier score 13229930.8770.898 −0.021 [−0.112, +0.070] Table A14: Lower gpt-oss effort is more accurate at its stopping point under candidate-logit and generated- answer readouts. Each row compares a completed low- or medium-effort reasoning block with high effort at the item- and replicate-specific lower-effort terminal horizon:198GPQA Diamond or500MMLU-Pro items across three generation replicates. If high effort finishes before that horizon, its completed reasoning block is replayed. Candidate-logit and high-active columns use a common medium-effort answer context; the latter restricts to pairs in which high effort remains unfinished. Generated-answer columns continue all replayable prefixes without option normalization. Active/all gives the high-active and replayable pair counts, and parse coverage is the share with two parsed continuations. Differences are the completed lower-effort readout minus the matched-horizon high-effort readout. Intervals first average replicates within item and then bootstrap items (10,000 resamples). ModelBenchmarkEffortCandidate-logit ∆ [95% CI] High-active ∆ [95% CI] Active / all pairsGenerated-answer ∆ [95% CI] Parse coverage gpt-oss-20bGPQA Diamond low+14.5 [+7.7, +21.2] +14.5 [+7.7, +21.2] 594/594+15.5 [+8.8, +22.1] 99.7% gpt-oss-20bGPQA Diamond medium+17.4 [+11.3, +23.7] +17.9 [+11.5, +24.3] 567/593+17.4 [+11.1, +23.8] 98.8% gpt-oss-20bMMLU-Prolow+20.9 [+16.7, +25.0] +21.1 [+16.9, +25.2] 1,496/1,500+20.5 [+16.2, +24.5] 99.3% gpt-oss-20bMMLU-Promedium+18.0 [+14.5, +21.5] +19.2 [+15.6, +22.9] 1,407/1,500+18.2 [+14.7, +21.7] 99.1% gpt-oss-120bGPQA Diamond low+23.6 [+16.3, +30.6] +23.6 [+16.3, +30.6] 594/594+23.1 [+15.8, +30.3] 99.8% gpt-oss-120bGPQA Diamond medium+23.6 [+16.7, +30.5] +23.9 [+17.0, +30.9] 591/594+23.1 [+16.2, +30.0] 100.0% gpt-oss-120bMMLU-Prolow+26.3 [+21.9, +30.5] +26.3 [+21.9, +30.5] 1,499/1,500+25.9 [+21.5, +30.2] 99.7% gpt-oss-120bMMLU-Promedium+24.2 [+20.3, +28.2] +24.9 [+20.9, +28.9] 1,473/1,500+24.1 [+20.1, +28.0] 99.5% 36 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A15: High valid-option mass coexists with proper-score disagreement driven by near-zero correct-option probabilities in terminal readouts. The table uses the all-replayable token-exact prefixes in Table A14, comprising 198GPQA Diamond or500MMLU-Pro items across three generation replicates. a, Valid-option mass. b, Absolute arm means and paired differences for conditional log loss and multiclass Brier score. c, Median conditional log loss and the share of all probes with option-normalizedq gold < 10 −4 . Differences are the completed lower-effort readout minus the matched-horizon high-effort readout. Higher valid-option mass and lower proper scores are preferred. Means and differences first average replicates within item; intervals then bootstrap items. a Valid-option mass ModelBenchmarkEffortItemsReplicatesTerminalHigh prefix∆ [95% CI] gpt-oss-20bGPQA Diamondlow198399.9%98.5%+1.5 [+1.2, +1.8] gpt-oss-20bGPQA Diamondmedium198399.8%98.0%+1.8 [+1.3, +2.5] gpt-oss-20bMMLU-Prolow500399.6%99.2%+0.4 [−0.0, +0.8] gpt-oss-20bMMLU-Promedium500399.9%99.3%+0.6 [+0.4, +0.8] gpt-oss-120bGPQA Diamondlow1983100.0%100.0%+0.0 [+0.0, +0.0] gpt-oss-120bGPQA Diamondmedium1983100.0%100.0%+0.0 [+0.0, +0.0] gpt-oss-120bMMLU-Prolow5003100.0%99.9%+0.1 [+0.0, +0.2] gpt-oss-120bMMLU-Promedium5003100.0%99.9%+0.1 [+0.0, +0.2] b Conditional proper-score means ModelBenchmarkEffortMetricTerminalHigh prefix∆ [95% CI] gpt-oss-20bGPQA DiamondlowConditional log loss5.7321.396+4.336 [+3.599, +5.068] Conditional Brier score0.8780.723+0.155 [+0.051, +0.259] gpt-oss-20bGPQA DiamondmediumConditional log loss5.8031.748+4.056 [+3.224, +4.890] Conditional Brier score0.7020.725 −0.023 [−0.124, +0.079] gpt-oss-20bMMLU-ProlowConditional log loss5.3141.855+3.459 [+2.953, +3.987] Conditional Brier score0.7120.721 −0.009 [−0.072, +0.056] gpt-oss-20bMMLU-PromediumConditional log loss5.0752.328+2.747 [+2.241, +3.264] Conditional Brier score0.5520.651 −0.099 [−0.154, −0.045] gpt-oss-120bGPQA DiamondlowConditional log loss8.8703.351+5.520 [+4.126, +6.923] Conditional Brier score0.7001.014 −0.314 [−0.449, −0.179] gpt-oss-120bGPQA DiamondmediumConditional log loss8.4573.286+5.171 [+3.687, +6.648] Conditional Brier score0.6300.957 −0.327 [−0.460, −0.196] gpt-oss-120bMMLU-ProlowConditional log loss6.2192.974+3.244 [+2.429, +4.068] Conditional Brier score0.4910.812 −0.322 [−0.398, −0.245] gpt-oss-120bMMLU-PromediumConditional log loss5.6233.196+2.427 [+1.655, +3.226] Conditional Brier score0.4190.759 −0.341 [−0.413, −0.270] c Conditional log-loss medians and lower tails ModelBenchmarkEffortPairsTerminal medianHigh-prefix median q gold < 10 −4 , terminal / high prefix gpt-oss-20bGPQA Diamondlow5940.0001.22436.0% / 0.2% gpt-oss-20bGPQA Diamondmedium5930.0001.15534.6% / 2.4% gpt-oss-20bMMLU-Prolow1,5000.0001.63432.3% / 0.4% gpt-oss-20bMMLU-Promedium1,5000.0000.83727.4% / 5.9% gpt-oss-120bGPQA Diamondlow5940.0002.14434.8% / 8.2% gpt-oss-120bGPQA Diamondmedium5940.0001.54231.5% / 8.6% gpt-oss-120bMMLU-Prolow1,5000.0000.94624.5% / 8.5% gpt-oss-120bMMLU-Promedium1,5000.0000.25920.9% / 11.7% 37 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A16: Earlier lower-effort completion accounts for most of the fixed-horizon gpt-oss gap. Active means that reasoning had not terminated atB = 512. Shares and contributions use item–replicate pairs from three generated replicates, and contributions sum to the overall lower-minus-high accuracy difference. Within-stratum intervals average replicate outcomes within item before bootstrapping items. Because contributions use raw pair means, they need not equal the share times the displayed item-averaged difference. ModelBenchmarkContrastTermination stratumItemsPairsShare Within-stratum ∆ [95% CI] Contribution gpt-oss-20bGPQA DiamondLow–highBoth active95178 30.0% −4.2 [−13.3, +4.9]−1.0 gpt-oss-20bGPQA DiamondLow–highLower complete; high active 171403 67.8% +16.2 [+8.1, +24.2]+11.8 gpt-oss-20bGPQA DiamondLow–highBoth complete7132.2% −14.3 [−42.9, +0.0] −0.2 gpt-oss-20bGPQA DiamondMedium–highBoth active182501 84.3% +2.8 [−1.5, +7.1]+1.9 gpt-oss-20bGPQA DiamondMedium–highLower complete; high active 4980 13.5% +23.1 [+12.2, +34.7]+3.4 gpt-oss-20bGPQA DiamondMedium–highBoth complete7132.2% +0.0 [+0.0, +0.0]+0.0 gpt-oss-20bMMLU-ProLow–highBoth active56926.2% +1.2 [−11.0, +13.4] −0.1 gpt-oss-20bMMLU-ProLow–highLower complete; high active 416 1,045 69.9% +7.3 [+3.3, +11.3]+6.2 gpt-oss-20bMMLU-ProLow–highBoth complete169357 23.9% −6.8 [−10.5, −3.5]−1.5 gpt-oss-20bMMLU-ProMedium–highBoth active303710 47.4% −1.6 [−5.7, +2.6]−0.3 gpt-oss-20bMMLU-ProMedium–highLower complete; high active 255432 28.8% +11.6 [+6.9, +16.3]+3.4 gpt-oss-20bMMLU-ProMedium–highBoth complete163344 22.9% −0.4 [−2.5, +1.6]−0.1 gpt-oss-20bMMLU-ProMedium–highHigh complete; lower active 12130.9% −25.0 [−50.0, +0.0] −0.2 gpt-oss-120bGPQA DiamondLow–highBoth active63132 22.2% +10.6 [−1.6, +22.8]+3.2 gpt-oss-120bGPQA DiamondLow–highLower complete; high active 168438 73.7% +22.3 [+15.2, +29.5]+15.8 gpt-oss-120bGPQA DiamondLow–highBoth complete13244.0% +0.0 [+0.0, +0.0]+0.0 gpt-oss-120bGPQA DiamondMedium–highBoth active154421 70.9% +4.5 [−0.1, +9.4]+1.9 gpt-oss-120bGPQA DiamondMedium–highLower complete; high active 68149 25.1% +20.1 [+9.8, +30.4]+5.6 gpt-oss-120bGPQA DiamondMedium–highBoth complete13244.0% +0.0 [+0.0, +0.0]+0.0 gpt-oss-120bMMLU-ProLow–highBoth active46996.6% +6.5 [−6.5, +19.6]+0.4 gpt-oss-120bMMLU-ProLow–highLower complete; high active 353900 60.0% +12.0 [+7.7, +16.3]+7.4 gpt-oss-120bMMLU-ProLow–highBoth complete214500 33.4% +0.2 [−2.1, +2.6]−0.1 gpt-oss-120bMMLU-ProMedium–highBoth active164379 25.3% +2.0 [−3.5, +7.5]+0.0 gpt-oss-120bMMLU-ProMedium–highLower complete; high active 284621 41.4% +11.4 [+7.0, +15.9]+5.3 gpt-oss-120bMMLU-ProMedium–highBoth complete213498 33.2% +0.9 [−0.9, +2.8]+0.1 gpt-oss-120bMMLU-ProMedium–highHigh complete; lower active 220.1% −50.0 [−100.0, +0.0] −0.1 38 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A17: Lower gpt-oss effort makes answer evidence available earlier, while active-prefix differences are smaller. The terminal-aware area under the curve (AUC) averages paired correct-option log-odds differences from0to 512reasoning tokens and carries terminal states forward. Stable-correct∆ 256 compares whether answers at256tokens remain correct through512. The active-pair AUC retains item–replicate pairs for which both traces haveL > 128 and recomputes the AUC over 0–128; conditioning on survival changes the compared population. Qwen contrasts are numeric/concision prompt minus surprise stop. gpt-oss contrasts are lower effort minus high effort. Intervals bootstrap items. a Terminal-aware and stable-correct contrasts ModelBenchmarkContrastnTerminal-aware AUC∆ log-odds Stable-correct ∆ 256 (p) gpt-oss-20bGPQA Diamond low minus high196+1.85 [+0.74, +2.90]+15.8 [+8.2, +23.5] gpt-oss-20bGPQA Diamond medium minus high196+0.83 [+0.49, +1.18]+4.3 [+0.2, +8.3] gpt-oss-20bMMLU-Prolow minus high500+1.44 [+0.76, +2.11]+17.7 [+14.1, +21.3] gpt-oss-20bMMLU-Promedium minus high500+1.36 [+1.04, +1.70]+8.3 [+5.8, +10.7] gpt-oss-120bGPQA Diamond low minus high198+4.22 [+2.38, +6.00]+14.1 [+7.6, +20.7] gpt-oss-120bGPQA Diamond medium minus high198+2.00 [+1.29, +2.70]+3.9 [−0.5, +8.4] gpt-oss-120bMMLU-Prolow minus high500+3.76 [+2.59, +4.91]+15.7 [+11.7, +19.6] gpt-oss-120bMMLU-Promedium minus high500+3.27 [+2.65, +3.87]+9.2 [+6.5, +11.9] Qwen3-14BGPQA Diamond prompt minus surprise 198+0.17 [+0.03, +0.32] −1.2 [−4.2, +1.9] Qwen3-14BMMLU-Proprompt minus surprise 500+0.42 [+0.18, +0.65]+0.8 [−0.7, +2.3] b Jointly active early-prefix contrasts ModelBenchmarkContrastActive item–replicate pairsActive-pair AUC 0:128 gpt-oss-20bGPQA Diamond low minus high414+0.05 [−0.10, +0.20] gpt-oss-20bGPQA Diamond medium minus high580+0.06 [−0.02, +0.16] gpt-oss-20bMMLU-Prolow minus high498+0.12 [−0.06, +0.28] gpt-oss-20bMMLU-Promedium minus high1,314+0.19 [+0.04, +0.36] gpt-oss-120bGPQA Diamond low minus high441 −0.10 [−0.46, +0.25] gpt-oss-120bGPQA Diamond medium minus high580 −0.05 [−0.24, +0.15] gpt-oss-120bMMLU-Prolow minus high534+0.12 [−0.33, +0.55] gpt-oss-120bMMLU-Promedium minus high1,178+0.53 [+0.29, +0.77] Qwen3-14BGPQA Diamond prompt minus surprise 593+0.08 [−0.01, +0.18] Qwen3-14BMMLU-Proprompt minus surprise 1,498+0.06 [−0.03, +0.16] 39 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A18: Qwen’s small prompt effects are distributed across unfinished-prefix and completion strata. Qwen3- 14B numeric/concision-prompt and hidden-horizon traces are partitioned by whether each arm has completed at the imposed horizonB. Shares use all token-exact matched item–replicate pairs and sum to one before rounding within each benchmark–horizon cell. Within-stratum differences are visible minus hidden accuracy. Contributions use the item-normalized pair decomposition and sum before rounding to the primary pooled difference. Intervals average the three observed replicates within item before bootstrapping items. All four horizons are reported. Several late-horizon strata, especially the both-active strata, are sparse and should be read descriptively. BenchmarkB Termination stratumItems Item–replicate pairsShare Within-stratum ∆ [95% CI] Contribution GPQA Diamond512 Both active197580 97.6%−0.8 [−2.8, +1.2]−0.7 Visible complete; hidden active9111.9% +11.1 [+0.0, +33.3]+0.2 Hidden complete; visible active330.5% +0.0 [+0.0, +0.0]+0.0 GPQA Diamond2,048 Both active146394 66.3%−0.5 [−5.8, +4.9]−0.3 Visible complete; hidden active42569.4% +8.3 [+2.4, +15.5]+0.8 Hidden complete; visible active22244.0% +2.3 [−13.6, +18.2]+0.0 Both complete55120 20.2% +1.2 [−1.8, +5.5]+0.2 GPQA Diamond8,192 Both active25416.9% +7.3 [−9.3, +24.0]+0.7 Visible complete; hidden active42569.4% +0.4 [−15.1, +15.1]−0.2 Hidden complete; visible active12152.5%−33.3 [−58.3,−8.3]−0.8 Both complete180482 81.1% +1.0 [−2.2, +4.3]+1.0 GPQA Diamond16,384 Both active120.3% +0.0 [+0.0, +0.0]+0.0 Visible complete; hidden active230.5% +50.0 [+0.0, +100.0]+0.3 Hidden complete; visible active330.5% +0.0 [−100.0, +100.0] +0.0 Both complete197586 98.7% +0.7 [−2.5, +3.9]+1.0 MMLU-Pro512 Both active4061,085 72.3% +2.0 [−0.1, +4.2]+1.3 Visible complete; hidden active991187.9% +6.6 [+0.5, +13.1]+0.5 Hidden complete; visible active54614.1% +0.9 [−6.5, +8.3]+0.1 Both complete118236 15.7% +0.6 [−2.0, +3.1]+0.1 MMLU-Pro2,048 Both active132299 19.9% +2.0 [−3.5, +7.6]+0.2 Visible complete; hidden active1021469.7% +2.6 [−4.7, +9.8]+0.3 Hidden complete; visible active35392.6%−5.7 [−18.6, +7.1]−0.2 Both complete3761,016 67.7% +1.3 [−0.1, +2.9]+0.4 MMLU-Pro8,192 Both active19402.7%−11.4 [−28.1, +2.6]−0.3 Visible complete; hidden active29342.3% +1.7 [−17.2, +20.7]+0.1 Hidden complete; visible active11140.9%−22.7 [−54.5, +4.5]−0.2 Both complete4811,412 94.1%−0.8 [−2.5, +0.8]−0.5 MMLU-Pro16,384 Both active230.2%−50.0 [−100.0, +0.0]−0.1 Visible complete; hidden active570.5% +20.0 [+0.0, +40.0]+0.1 Hidden complete; visible active230.2%−50.0 [−100.0, +0.0]−0.1 Both complete4981,487 99.1% +0.1 [−1.3, +1.5]+0.1 40 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A19: Qwen closure and source-context effects remain small and mixed, whereas gpt-oss source-context estimates are positive atB = 256and512. Panel a compares three forced-answer closures for the same Qwen prompt-minus-surprise contrast. Panel a reports item-bootstrap intervals, but the number of contributing items is unavailable. Panel b reconstructs each source prompt at the branch point. Qwen rows are numeric/concision prompt minus surprise stop; gpt-oss rows are lower source effort minus high source effort. Deltas are percentage points with paired item-bootstrap intervals. The Qwen results cover all four horizons, whereas the gpt-oss results coverB = 256 and 512. a Forced-answer closure wording for Qwen3-14B BenchmarkBLimited-time ∆ [95% CI]Paraphrase ∆ [95% CI]Neutral ∆ [95% CI]Neutral minus limited-time GPQA Diamond512−1.3 [−4.5, +1.7]−1.7 [−4.7, +1.3]−0.8 [−4.9, +3.0]+0.5 [−4.0, +5.1] GPQA Diamond2,048−1.3 [−4.5, +1.7]−1.0 [−4.0, +2.0]−0.8 [−4.5, +2.9]+0.5 [−1.7, +2.7] GPQA Diamond8,192−2.4 [−5.4, +0.7]−1.9 [−4.7, +1.2]−2.9 [−5.9, +0.2]−0.5 [−2.2, +1.2] GPQA Diamond16,384+1.6 [−1.8, +5.2]+2.1 [−1.3, +5.7]+1.8 [−1.7, +5.4]+0.2 [+0.0, +0.5] MMLU-Pro512+1.7 [+0.0, +3.5]+2.4 [+0.7, +4.1]+1.9 [+0.1, +3.7]+0.1 [−0.8, +1.1] MMLU-Pro2,048+0.9 [−0.7, +2.4]+0.8 [−0.8, +2.4]+1.0 [−0.5, +2.5]+0.1 [−0.3, +0.6] MMLU-Pro8,192−0.2 [−1.5, +1.2]−0.1 [−1.5, +1.3]−0.2 [−1.5, +1.2]+0.0 [−0.2, +0.2] MMLU-Pro16,384−0.7 [−2.2, +0.6]−0.7 [−2.1, +0.7]−0.6 [−2.1, +0.7]+0.1 [+0.0, +0.3] b Source-context branch readout ModelBenchmarkBContrastnArm A acc.Arm B acc.∆ [95% CI] Qwen3-14BGPQA Diamond 512prompt–surprise19840.4%44.9%−4.5 [−9.3, 0.0] Qwen3-14BGPQA Diamond 2,048prompt–surprise19853.4%55.6%−2.2 [−5.7, +1.3] Qwen3-14BGPQA Diamond 8,192prompt–surprise19859.1%61.8%−2.7 [−5.7, +0.5] Qwen3-14BGPQA Diamond 16,384prompt–surprise19864.1%62.3%+1.9 [−1.5, +5.4] Qwen3-14BMMLU-Pro512prompt–surprise50060.7%58.7%+1.9 [+0.1, +3.8] Qwen3-14BMMLU-Pro2,048prompt–surprise50072.0%70.9%+1.1 [−0.5, +2.6] Qwen3-14BMMLU-Pro8,192prompt–surprise50075.9%76.1%−0.2 [−1.5, +1.1] Qwen3-14BMMLU-Pro16,384prompt–surprise50075.9%76.5%−0.6 [−2.1, +0.7] gpt-oss-20bGPQA Diamond 256low–high19653.4%39.6%+13.8 [+6.1, +21.4] gpt-oss-20bGPQA Diamond 512low–high19657.0%44.9%+12.1 [+4.4, +19.7] gpt-oss-20bGPQA Diamond 256medium–high19642.5%39.6%+2.9 [−2.4, +8.2] gpt-oss-20bGPQA Diamond 512medium–high19650.0%44.9%+5.1 [−0.9, +11.2] gpt-oss-20bMMLU-Pro256low–high50062.1%49.6%+12.5 [+9.1, +16.1] gpt-oss-20bMMLU-Pro512low–high50063.7%55.6%+8.1 [+4.6, +11.7] gpt-oss-20bMMLU-Pro256medium–high50056.1%49.6%+6.5 [+4.0, +9.0] gpt-oss-20bMMLU-Pro512medium–high50062.9%55.6%+7.3 [+4.7, +9.9] gpt-oss-120bGPQA Diamond 256low–high19851.0%40.6%+10.4 [+3.0, +17.8] gpt-oss-120bGPQA Diamond 512low–high19861.1%42.9%+18.2 [+10.9, +25.4] gpt-oss-120bGPQA Diamond 256medium–high19843.8%40.6%+3.2 [−1.2, +7.6] gpt-oss-120bGPQA Diamond 512medium–high19853.0%42.9%+10.1 [+4.5, +15.7] gpt-oss-120bMMLU-Pro256low–high49969.9%58.1%+11.8 [+7.9, +15.8] gpt-oss-120bMMLU-Pro512low–high49972.5%65.3%+7.1 [+3.3, +10.8] gpt-oss-120bMMLU-Pro256medium–high50065.9%58.1%+7.8 [+5.3, +10.4] gpt-oss-120bMMLU-Pro512medium–high50072.0%65.3%+6.7 [+3.9, +9.4] 41 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A20: Candidate-logit and generated-answer readouts both show higher accuracy for completed lower-effort reasoning, with complete or near-complete coverage. Each row uses the primary token-exact matched gpt-oss prefixes. Panel a gives absolute candidate-logit arm means under the common medium-effort answer context; every pair eligible under the matched-horizon rule is replayable. Panel b gives absolute greedy-continuation arm means after averaging the three observed replicates within item. Complete/eligible reports pairs with both continuations relative to runnable pairs, and parse coverage requires both answers to parse. a Candidate-logit readout, all replayable pairs ModelBenchmarkTerminal effortTerminal acc.High-prefix acc.Replayable / eligibleNonreplayable gpt-oss-20bGPQA Diamond low55.9%41.4%594/5940 gpt-oss-20bGPQA Diamond medium64.9%47.5%593/5930 gpt-oss-20bMMLU-Prolow64.2%43.3%1,500/1,5000 gpt-oss-20bMMLU-Promedium72.4%54.4%1,500/1,5000 gpt-oss-120bGPQA Diamond low65.0%41.4%594/5940 gpt-oss-120bGPQA Diamond medium68.5%44.9%594/5940 gpt-oss-120bMMLU-Prolow75.5%49.2%1,500/1,5000 gpt-oss-120bMMLU-Promedium79.1%54.9%1,500/1,5000 b Greedy generated-answer continuation ModelBenchmarkTerminal effortTerminal acc.High-prefix acc.Complete / eligibleParse coverage gpt-oss-20bGPQA Diamond low55.9%40.4%594/59499.7% gpt-oss-20bGPQA Diamond medium64.7%47.3%593/59398.8% gpt-oss-20bMMLU-Prolow64.1%43.7%1,500/1,50099.3% gpt-oss-20bMMLU-Promedium72.3%54.1%1,500/1,50099.1% gpt-oss-120bGPQA Diamond low65.0%41.9%594/59499.8% gpt-oss-120bGPQA Diamond medium68.5%45.5%594/594100.0% gpt-oss-120bMMLU-Prolow75.5%49.5%1,500/1,50099.7% gpt-oss-120bMMLU-Promedium79.1%55.0%1,500/1,50099.5% 42 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A21: Correct-option probabilities are markedly lower among incorrect completed lower-effort predictions than among incorrect matched-horizon high-effort predictions. Rows condition separately within each arm on an incorrect candidate-readout answer among all replayable token-exact pairs. Mean conditional log loss and the share with option-normalizedq gold < 10 −4 describe the lower tail of correct-option probability. Counts are arm-specific incorrect item–replicate probes and need not match across arms. Means first average observed replicates within item and then average items. ModelBenchmarkEffortArmIncorrect probesMean log loss q gold < 10 −4 gpt-oss-20bGPQA Diamond lowTerminal26212.74580.1% High prefix3481.9360.3% gpt-oss-20bGPQA Diamond mediumTerminal20816.55898.4% High prefix3122.9523.8% gpt-oss-20bMMLU-ProlowTerminal53714.65788.9% High prefix8512.8620.8% gpt-oss-20bMMLU-PromediumTerminal41418.31599.1% High prefix6844.46011.1% gpt-oss-120bGPQA Diamond lowTerminal20825.28999.6% High prefix3485.23312.4% gpt-oss-120bGPQA Diamond mediumTerminal18726.860100.0% High prefix3275.50214.5% gpt-oss-120bMMLU-ProlowTerminal36825.34399.4% High prefix7625.48615.4% gpt-oss-120bMMLU-PromediumTerminal31426.839100.0% High prefix6776.65824.8% 43 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A22: In single-rollout Qwen3-8B and Omni-MATH-2 comparisons, Qwen3-8B shows no broad prompt gain and every gpt-oss fixed-checkpoint difference is positive. Panel a compares one Qwen3-8B numeric/concision-prompt rollout with one separately generated surprise-stop rollout; the accuracy denominator is the number of unique paired items. For MMLU-Pro atB = 2,048, the length medians use 499 paired item–trace records although the accuracy contrast contains 498 items. Panel b starts from a difficulty-stratified 500-item Omni-MATH-2 draw; the analyzed model-specific paired cohorts contain 475 items for gpt-oss-20b and 490 for gpt-oss-120b. The available data do not record why the remaining 25 and 10 sampled items, respectively, are absent. Panel b uses one rollout per effort setting, with GPT-5 mini judging free-form answers at common source-token checkpoints. Intervals bootstrap items. These estimates characterize the sampled trajectories and do not quantify variation across generated replicates. a Qwen3-8B single-rollout comparison BenchmarkB nPrompt acc.Surprise acc.∆ [95% CI]Sign-flip pMedian prompt LMedian surprise L GPQA Diamond 51219634.2%36.7% −2.6 [−6.1, +1.0]0.2695,0225,929 GPQA Diamond 2,04819843.4%42.9%+0.5 [−4.5, +5.6]1.0005,0965,957 MMLU-Pro51249853.4%52.6%+0.8 [−1.8, +3.2]0.6381,2861,623 MMLU-Pro2,04849867.7%65.1%+2.6 [−0.2, +5.2]0.0901,2471,639 b gpt-oss fixed-checkpoint Omni-MATH-2 readout ModelContrastB n Lower acc. High acc. ∆ [95% CI] gpt-oss-20blow–high512 47529.9%18.3% +11.6 [+8.4, +14.9] gpt-oss-20blow–high1,024 47538.1%21.7% +16.4 [+12.6, +20.2] gpt-oss-20blow–high2,048 47545.9%29.1% +16.8 [+12.6, +20.8] gpt-oss-20blow–high4,096 47548.2%34.9% +13.3 [+9.3, +17.3] gpt-oss-20bmedium–high512 47521.7%18.3% +3.4 [+0.8, +6.1] gpt-oss-20bmedium–high 1,024 47525.9%21.7% +4.2 [+1.3, +7.4] gpt-oss-20bmedium–high 2,048 47538.7%29.1% +9.7 [+5.9, +13.5] gpt-oss-20bmedium–high 4,096 47546.1%34.9% +11.2 [+7.4, +14.9] gpt-oss-120blow–high512 49039.0%26.9% +12.0 [+8.2, +15.9] gpt-oss-120blow–high1,024 49050.4%30.6% +19.8 [+15.9, +23.9] gpt-oss-120blow–high2,048 49059.0%41.0% +18.0 [+13.7, +22.2] gpt-oss-120blow–high4,096 49060.6%52.7% +8.0 [+3.9, +12.0] gpt-oss-120bmedium–high512 49030.0%26.9% +3.1 [+0.0, +6.1] gpt-oss-120bmedium–high 1,024 49039.6%30.6% +9.0 [+5.5, +12.4] gpt-oss-120bmedium–high 2,048 49053.9%41.0% +12.9 [+9.2, +16.5] gpt-oss-120bmedium–high 4,096 49065.5%52.7% +12.9 [+9.0, +16.7] 44 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A23: The Qwen MMLU-ProB = 512contrast varies by category, and a text-reconstructed four-horizon concise/early-answer grid shows no broad advantage. Panel a pools the primary 500-item subset and an independent 500-item subset atB = 512. Accuracies average three rollout-level candidate-logit readouts within item and then average items within category; no category-level intervals are available. In panel a, prompt denotes the numeric/concision prompt and∆is prompt minus surprise stop. Panel b compares the concise/early-answer prompt with independently generated surprise-stop and numeric/concision-prompt cohorts; its deltas are concise/early-answer minus the named comparator. Both panels use cohorts without stored completion token IDs or verified generation seeds. Panel a describes category variation across two subsets and panel b extends the prompt comparison across four horizons, but neither contributes to the token-exact primary estimates. Panel b intervals average three nominal runs within item and bootstrap items. Deltas are reported in percentage points. a MMLU-Pro category decomposition at B = 512 CategorynSurprise acc.Prompt acc.∆ (p) Biology6077.8%80.0%+2.2 Business6660.1%63.1%+3.0 Chemistry9451.8%57.4%+5.7 Computer science3464.7%72.5%+7.8 Economics7072.4%76.7%+4.3 Engineering8043.8%42.9% −0.8 Health6868.6%66.7% −2.0 History3269.8%67.7% −2.1 Law9231.5%37.0%+5.4 Math11258.9%60.4%+1.5 Other7654.4%53.9% −0.4 Philosophy4261.1%54.0% −7.1 Physics10850.9%55.2%+4.3 Psychology6672.2%74.7%+2.5 Pooled total1,00057.6%59.8%+2.1 b Four-horizon concise/early-answer prompt grid BenchmarkBConcise acc.Surprise acc.∆ vs. surprise [95% CI] Numeric-prompt acc.∆ vs. numeric prompt [95% CI] GPQA Diamond 51243.1%42.9%+0.2 [−4.0, +4.4] 43.9% −0.8 [−3.5, +1.9] GPQA Diamond 2,04856.1%54.9%+1.2 [−2.5, +4.7] 53.0%+3.0 [+0.0, +6.2] GPQA Diamond 8,19259.8%61.8% −2.0 [−5.4, +1.3] 58.9%+0.8 [−1.9, +3.5] GPQA Diamond 16,38460.8%62.3% −1.5 [−4.9, +1.7] 64.1% −3.4 [−6.7,−0.2] MMLU-Pro51263.5%58.7%+4.9 [+2.8, +7.0] 60.9%+2.7 [+0.9, +4.5] MMLU-Pro2,04871.8%70.8%+1.0 [−0.7, +2.7] 71.9% −0.1 [−1.7, +1.5] MMLU-Pro8,19274.1%76.1% −2.1 [−3.5,−0.7] 75.9% −1.9 [−3.3,−0.4] MMLU-Pro16,38473.9%76.5% −2.7 [−4.1,−1.3] 75.9% −2.0 [−3.4,−0.6] 45 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A24: An independent MMLU-Pro subset preserves the small numeric-prompt gain at 512 tokens but not at 2,048 tokens. The original-subset row uses the 500-item MMLU-Pro sample corresponding to the design in Table A6; it remains a separate replication rather than part of the token-exact primary analysis. The independent row uses a second category-stratified 500-item subset that excludes the original items, and the pooled row combines both subsets. At B = 2,048, the numeric/concision and concise/early-answer instructions are compared with the same independent- subset surprise-stop baseline under the same three-run procedure. Accuracy and survival use the candidate-logit readout over three nominal Qwen3-14B runs. Intervals are paired two-way item×nominal-run bootstrap intervals. Seeds and completion token IDs are unavailable for these cohorts, so the intervals do not support generalization across generation seeds. Only aggregate estimates and intervals are available for the twoB = 2,048rows, so they cannot be regenerated independently and are treated as descriptive evidence. Sample and instructionBnPrompt acc.Surprise acc.∆ [95% CI]P (L > B) prompt P (L > B) surpriseMedian prompt LMedian surprise L Original 500, numeric/concision 51250060.4%58.7%+1.7 [−0.6, +4.0] 75.9%81.3%8761,184 Independent 500, numeric/concision 51250059.1%56.6%+2.5 [+0.4, +4.8] 77.1%81.5%1,0481,314 Pooled 1,000, numeric/concision 5121,00059.8%57.6%+2.1 [+0.6, +3.6] 76.5%81.4%9551,214 Independent 500, numeric/concision 2,04850071.5%71.7% −0.2 [−2.6, +2.2] 26.3%36.0%1,0431,314 Independent 500, concise/early-answer 2,04850074.4%71.7%+2.7 [+0.5, +5.0] 19.9%36.0%8121,314 46 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A25: Qwen prompt controls change reasoning length more consistently than accuracy at the 512-token stop. Each row compares a Qwen3-14B instruction with the same surprise-stop baseline atB = 512, using three nominal runs averaged within item. The accuracy columns report terminal-aware policy accuracy, with candidate-logit scoring for unfinished prefixes, and the displayed median is the full reasoning length under the instruction. Completion token IDs and verified generation seeds are unavailable for these cohorts, which therefore characterize prompt patterns. The token-exact paired experiment in Table A7 supplies the primary estimate. Accuracy differences are instruction minus surprise stop in percentage points, with paired item-cluster bootstrap intervals. The four-horizon concise/early-answer results and direct prompt-to-prompt contrasts appear in Table A23. Full survival, survivor-accuracy, terminal-aware accuracy, and ∆ logL results are included in the companion data release. InstructionBenchmarknInstruction acc.Surprise acc.∆ (p)95% CIMedian instruction L Numeric/concisionGPQA Diamond 19843.9%42.9%+1.0[−2.5, +4.5]3,641 Limited, no numberGPQA Diamond 19845.3%42.9%+2.4[−1.7, +6.4]3,654 Number onlyGPQA Diamond 19843.8%42.9%+0.8[−3.2, +4.9]3,690 Advance exact-stop noticeGPQA Diamond 19843.6%42.9%+0.7[−3.9, +5.1]3,690 Approximate budgetGPQA Diamond 19841.4%42.9% −1.5[−5.4, +2.2]3,904 Irrelevant numberGPQA Diamond 19842.4%42.9% −0.5[−4.7, +3.5]4,222 Concise/early-answerGPQA Diamond 19843.1%42.9%+0.2[−4.0, +4.4]3,080 Thorough semanticGPQA Diamond 19842.6%42.9% −0.3[−4.4, +3.7]4,847 Numeric/concisionMMLU-Pro50060.9%58.7%+2.2[+0.3, +4.1]876 Limited, no numberMMLU-Pro50061.0%58.7%+2.3[+0.5, +4.1]892 Number onlyMMLU-Pro50060.9%58.7%+2.3[+0.5, +4.0]912 Advance exact-stop noticeMMLU-Pro50059.9%58.7%+1.3[−0.5, +3.0]880 Approximate budgetMMLU-Pro50060.9%58.7%+2.2[+0.5, +4.0]938 Irrelevant numberMMLU-Pro50059.9%58.7%+1.2[−0.5, +3.0]1,065 Concise/early-answerMMLU-Pro50063.5%58.7%+4.9[+2.8, +7.0]710 Thorough semanticMMLU-Pro50059.7%58.7%+1.0[−0.7, +2.7]1,299 47 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A26: Continuing an incorrect Qwen prefix can rescue its answer, but anchoring remains substantial. Each cell reports the rescue rate and anchoring rate in percent after replaying the firstBtokens of an incorrect Qwen3-14B natural trace. The analysis uses the returned-top-20forced-answer readout rather than full-vocabulary candidate-logit scoring. “Answer now” closes reasoning immediately. The continuation conditions requestR = 32,768additional reasoning tokens before the forced answer, with the remaining budget either hidden or stated in the prompt. The effective maximum is limited by the context window in every displayed continuation cell, with medians from approximately 24,100 to 31,900 tokens. BenchmarkB nRescue / anchoring rate (%) Answer nowContinue, hidden RContinue, visible R GPQA Diamond51212330.1/40.738.2/37.437.4/35.8 GPQA Diamond2,04811224.1/45.533.9/42.927.7/50.0 GPQA Diamond8,1923420.6/55.926.5/47.126.5/47.1 MMLU-Pro51217626.1/43.231.2/51.131.2/51.7 MMLU-Pro2,0489520.0/55.829.5/48.429.5/55.8 MMLU-Pro8,1923026.7/46.733.3/50.030.0/50.0 48 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A27: The prompted budget is decodable at the prompt boundary but mostly falls to label-shuffle levels after reasoning begins. The table reports the best-layer activation readouts for Qwen3-14B at representative token positionsk. “Bundled prompt?” is a binary readout over surprise-stop and numeric/concision-prompt traces. “Which budget?” is a four-class readout over numeric/concision-prompt traces only, withB ∈ 512, 2,048, 8,192, 16,384. Balanced accuracy uses item-grouped cross-validation. The shuffle column is the 95th percentile over 20 label shuffles. ReadoutBenchmarkkBest layerItemsBalanced acc.Shuffle q95 Bundled prompt?GPQA Diamond0101971.0000.546 Bundled prompt?GPQA Diamond64391970.6210.533 Bundled prompt?GPQA Diamond512391970.7000.526 Bundled prompt?GPQA Diamond2,048391630.6230.539 Bundled prompt?MMLU-Pro0102001.0000.537 Bundled prompt?MMLU-Pro64392000.6270.540 Bundled prompt?MMLU-Pro512391780.6890.548 Bundled prompt?MMLU-Pro2,04839750.5880.580 Which budget?GPQA Diamond0391970.9380.283 Which budget?GPQA Diamond64101970.2580.268 Which budget?GPQA Diamond512301970.2780.269 Which budget?GPQA Diamond2,048391560.2620.286 Which budget?MMLU-Pro0392000.9050.278 Which budget?MMLU-Pro6402000.2610.279 Which budget?MMLU-Pro512201710.2590.278 Which budget?MMLU-Pro2,0480690.3430.301 49 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A28: Adding numeric wording to gpt-oss effort produces no broad advantage at the 512-token stop. Each row compares a numeric-budget instruction crossed with the indicated effort setting against the effort-only fixed-checkpoint baseline at the sameB = 512stop. Differences are numeric-instruction minus effort-only accuracy in percentage points, with paired bootstrap intervals over items. Longer announced-budget contrasts lack a directly matched effort-only baseline because the effort-only comparison covers only B = 512. ModelBenchmarkEffortnNumeric acc.Effort-only acc.∆ (p)95% CI gpt-oss-20bGPQA Diamond low19855.6%56.1% −0.5[−7.6, +6.1] gpt-oss-20bGPQA Diamond medium19844.4%50.0% −5.6[−10.6,−0.5] gpt-oss-20bGPQA Diamond high19644.4%44.4%+0.0[−6.1, +6.1] gpt-oss-20bMMLU-Prolow49866.5%63.7%+2.9[−0.5, +6.5] gpt-oss-20bMMLU-Promedium50063.0%63.0%+0.0[−2.7, +2.7] gpt-oss-20bMMLU-Prohigh49958.9%56.2%+2.7[+0.0, +5.5] gpt-oss-120bGPQA Diamond low19864.1%61.1%+3.0[−3.5, +9.6] gpt-oss-120bGPQA Diamond medium19853.5%53.0%+0.5[−4.2, +5.2] gpt-oss-120bGPQA Diamond high19841.4%43.9% −2.5[−7.1, +2.0] gpt-oss-120bMMLU-Prolow49973.7%73.1%+0.6[−2.2, +3.4] gpt-oss-120bMMLU-Promedium50073.2%71.8%+1.4[−0.8, +3.7] gpt-oss-120bMMLU-Prohigh50066.0%66.7% −0.7[−3.0, +1.6] 50 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A29: The scripted positive control detects a known early-answer policy. We construct 96 balanced synthetic four-choice items whose questions contain no answer information. The long policy uses a neutral reasoning prefix whose candidate answer appears after 64 tokens, whereas the short policy states the candidate answer immediately. Both policies reveal the candidate answer byB = 512. Accuracy uses the Qwen3-14B candidate-logit readout. Differences are short minus long in percentage points, with paired bootstrap intervals over items. The reportedpvalues use a two-sided paired item-level sign-flip randomization test with 10,000 draws. B nLong acc.Short acc.∆ (p)95% CILong revealShort revealSign-flip p 649625.0%100.0%+75.0[+65.6, +83.3]0.0%100.0%0.0001 51296100.0%100.0%+0.0[+0.0, +0.0]100.0%100.0%1.0000 51 Shorter Reasoning, Earlier Answers? An Evaluation of Reasoning Interfaces Table A30: None of the 2,788 terminal gpt-oss replay prefixes in a separate single-run cohort contains final- channel answer content. This terminal-replay validation does not exhaustively cover the three-replicate primary cohort. It covers terminal low- and medium-effort prefixes in the single-run common-context readout. Answer assertions inside the reasoning text are part of the measured behavior and are not counted as leaks. A final-channel leak is a replay prefix containing a final-channel delimiter. A final-reply suffix leak is the parsed natural final reply appended to the replay prefix when the reply contains at least three non-whitespace characters. ModelBenchmarkPrefix sourcePrefixesFinal-channel leaksReply checksFinal-reply suffix leaks gpt-oss-120bGPQA Diamondlow terminal198000 gpt-oss-120b GPQA Diamondmedium terminal198000 gpt-oss-120bMMLU-Prolow terminal500000 gpt-oss-120bMMLU-Promedium terminal500000 gpt-oss-20bGPQA Diamondlow terminal196000 gpt-oss-20b GPQA Diamondmedium terminal196000 gpt-oss-20bMMLU-Prolow terminal500010 gpt-oss-20bMMLU-Promedium terminal500010 Total2,788020 52

Instant research discovery