Paper deep dive
Illusion of Alignment: Detecting Hidden Disagreement in Collaborative Dialogue
Kaiming Liu, Fuwen Luo, Ziyue Wang, Jinrui Ju, Yuxuan Liu, Xuanyu Lei, Yunghwei Lai, Peng Li, Yang Liu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/11/2026, 5:40:24 AM
Summary
This paper introduces the 'Illusion of Alignment' (IoA), a phenomenon where collaborative dialogue ends in apparent agreement despite hidden disagreements in goals, assumptions, or plans. The authors propose IoA-Suite, a dataset and evaluation protocol using diagnostic multiple-choice questions to detect these hidden misalignments. They train IoA-Prober-8B, which improves detection performance and aids multi-agent collaboration.
Entities (7)
Relation Signals (5)
IoA-Prober-8B → istrainedon → IoA-Suite
confidence 95% · We then train IoA-Prober-8B based on IoA-Suite, reaching 51.8% F1 on IoA-Suite.
Illusion of Alignment → isdetectedby → IoA-Suite
confidence 92% · We construct IoA-Suite, a dataset and evaluation protocol for detecting hidden disagreement... spanning five task types and six domains.
GPT-5.4 → isusedas → Judge
confidence 90% · The simulator uses Gemini-3-Flash (Google, 2026) and the judge uses GPT-5.4 (OpenAI, 2026).
IoA-Prober-8B → improvesperformanceon → BigCodeBench-Hard
confidence 88% · pairing IoA-Prober-8B with LLM agents improves downstream task performance on BigCodeBench-Hard and HiddenBench.
IoA-Prober-8B → improvesperformanceon → HiddenBench
confidence 88% · pairing IoA-Prober-8B with LLM agents improves downstream task performance on BigCodeBench-Hard and HiddenBench.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Collaborative dialogue can end with apparent agreement while participants still differ on goals, assumptions, or execution plans, creating an \textbf{illusion of alignment (IoA)}. A real-user study across 18 meetings confirms that IoA arises routinely in human collaboration. Yet IoA poses a paradox: if participants were aware of such disagreements, they would already be explicit; if not, they cannot articulate them when asked, leaving IoA invisible to both participants and observers. In this work, we make IoA detectable by generating diagnostic multiple-choice questions whose divergent answers across participants provide direct behavioral evidence of hidden disagreement. We construct \textbf{IoA-Suite}, a dataset and evaluation protocol for detecting hidden disagreement, spanning five task types and six domains. We find that even the best model attains only 49.5\% F1, with the bottleneck traced to private context that the dialogue does not surface. We then train \textbf{IoA-Prober-8B} based on IoA-Suite, reaching 51.8\% F1 on IoA-Suite. Across the aforementioned 18 real meetings, it surfaces 2.89 hidden disagreements per meeting that participants confirm they had not voiced, transferring to live human dialogue. Further, in multi-agent collaboration, pairing IoA-Prober-8B with LLM agents improves downstream task performance on BigCodeBench-Hard and HiddenBench.
Tags
Links
- Source: https://arxiv.org/abs/2608.08210v1
- Canonical: https://arxiv.org/abs/2608.08210v1
Trouble viewing inline? Open PDF directly →
Full Text
123,344 characters extracted from source content.
Expand or collapse full text
Illusion of Alignment: Detecting Hidden Disagreement in Collaborative Dialogue Kaiming Liu 1,2 , Fuwen Luo 2,3 , Ziyue Wang 2,3 , Jinrui Ju 3 , Yuxuan Liu 3 , Xuanyu Lei 2,3 , Yunghwei Lai 2,3 , Peng Li 2,† , Yang Liu 1,2,3,† 1 College of AI, Tsinghua University, Beijing, China 2 Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China 3 Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua University, Beijing, China lkm20@mails.tsinghua.edu.cn, lipeng@air.tsinghua.edu.cn liuyang2011@tsinghua.edu.cn Abstract Collaborative dialogue can end with appar- ent agreement while participants still differ on goals, assumptions, or execution plans, creat- ing an illusion of alignment (IoA). A real- user study across 18 meetings confirms that IoA arises routinely in human collaboration. Yet IoA poses a paradox: if participants were aware of such disagreements, they would al- ready be explicit; if not, they cannot articulate them when asked, leaving IoA invisible to both participants and observers. In this work, we make IoA detectable by generating diagnostic multiple-choice questions whose divergent an- swers across participants provide direct behav- ioral evidence of hidden disagreement. We con- struct IoA-Suite, a dataset and evaluation pro- tocol for detecting hidden disagreement, span- ning five task types and six domains. We find that even the best model attains only 49.5% F1, with the bottleneck traced to private context that the dialogue does not surface. We then train IoA-Prober-8B based on IoA-Suite, reaching 51.8% F1 on IoA-Suite. Across the aforemen- tioned 18 real meetings, it surfaces 2.89 hid- den disagreements per meeting that participants confirm they had not voiced, transferring to live human dialogue. Further, in multi-agent col- laboration, pairing IoA-Prober-8B with LLM agents improves downstream task performance on BigCodeBench-Hard and HiddenBench. 1 Introduction Consider the scenario in Figure 1: a technical lead instructs an engineer to “Go learn Torch before debugging it,” and the engineer agrees. A week later, the lead expects familiarity with a legacy Lua- Torch codebase, while the engineer has prepared around the modern PyTorch. Neither party was † Corresponding Authors. ‡ Code, data and model weights will be released athttps: //github.com/THUNLP-MT/IoA. Theclient’s model has a bug. Go learn Torch before debugging it. Sure! I’l learn Torch this week and start investigating the bug. Great! Let’s sync next Monday. Lead Lead Engineer Go learnTorch before debugging it require ‘torch’ require ‘n’ model = torch.load(‘model.t7’) Leadexpects LuaTorchcodebase importtorch import torch.n as n model = n.Module(...) Engineerexpects PyTorch codebase TwovalidInterpretations. One failed handoff. à Lua Torch à PyTorch Sameword, different referents Illusion of Alignment surface agreement, incompatible assumptions Figure 1: An example of the illusion of alignment. The lead and engineer appear to agree on the instruction “Go learn Torch before debugging it”, but the shared term hides incompatible assumptions: the lead refers to the legacy Lua-Torch codebase used by the client, whereas the engineer prepares for modern PyTorch. The task of this work is to surface such hidden misalignments before apparent consensus leads to a failed handoff. careless, and to each participant the transcript con- tains no hedging, contradiction, or repair. Each interpreted the shared term through a private con- text that the conversation never surfaced, yet this silent divergence ultimately derails the handoff. We refer to this phenomenon as the illusion of align- ment (IoA): a collaborative dialogue ends in sur- face agreement while participants still differ on goals, assumptions, or execution plans. This phenomenon is far from a contrived edge case. A real-user study (Section 6.2) across 18 meetings with 43 participants surfaces 2.89 hid- den disagreements per meeting, which participants confirm they had not articulated during the origi- nal discussions. While long recognized in organi- zational psychology as failures of shared mental models (Cannon-Bowers et al., 1993) and in man- agement practice as a leading cause of execution breakdown (Lencioni, 2002), IoA has received little attention in the computational dialogue research. Beyond the scholarly neglect, detecting IoA 1 arXiv:2608.08210v1 [cs.AI] 8 Aug 2026 poses a structural paradox. Surface cues such as hedging or contradiction (De Kock and Vlachos, 2021) are absent precisely because participants be- lieve consensus has been reached, and existing work on dialogue disagreement studies only such explicit friction or observable common-ground fail- ures (Nath et al., 2025; Sarkar et al., 2025). Detect- ing IoA therefore requires Theory-of-Mind reason- ing (Premack and Woodruff, 1978): a detector must infer how different participants, given their public profiles, would map the same apparently aligned transcript to divergent interpretations, expectations, or downstream commitments. This requirement differs from existing LLM Theory-of-Mind evalua- tions (Wilf et al., 2024; Chen et al., 2024), which probe agent beliefs or knowledge at locations spec- ified in advance by the evaluation; in IoA detection, no such locations are given, and identifying where participants diverge is itself the task. Building on this perspective, we raise two re- search questions: RQ1: How can hidden disagree- ment under illusion of alignment be elicited and evaluated in a verifiable manner? RQ2: To what extent can task-specific training improve IoA detec- tion from observable dialogue context? In this work, we recast IoA detection as a ques- tion generation task: given a dialogue, a detector generates a small set of diagnostic multiple-choice questions, and each participant answers them in- dependently from their own perspective. A ques- tion on which two participants select different op- tions constitutes direct behavioral evidence of a hidden disagreement they did not surface during the conversation. This reformulation relocates the IoA signal from the transcript, which by definition conceals it, to downstream participant behavior, turning an ill-posed introspection problem into a mechanically decidable filtering task. To evaluate whether current models can probe accurately (asking questions that genuinely reveal disagreements) and efficiently (without burdening participants with noise), we construct IoA-Suite, the first dataset and evaluation protocol for IoA detection, with synthesized dialogues paired with ground-truth misalignment sets across five collab- orative task types and six domains. Evaluation across nine frontier models indicates that even the strongest model attains only 49.5% F1, with fur- ther analysis tracing the bottleneck to the private context that the paradox of IoA conceals. To bridge this gap, we train IoA-Prober-8B via an RL recipe, reaching 51.8% F1, with the gains further validated in a real-user meeting study. Furthermore, pairing IoA-Prober-8B with LLM agents improves down- stream task performance in multi-agent collabora- tion. Our contributions are summarized as follows: •We bring the illusion of alignment into com- putational dialogue research, recasting hidden disagreement as a behavioral signal that can be elicited and measured directly through diagnos- tic multiple-choice questions. •We construct IoA-Suite and show that IoA de- tection remains unsolved across current LLMs, with experiments tracing the difficulty to pri- vate context the dialogue does not surface. •We train IoA-Prober-8B, a detector that sur- faces hidden disagreements in real-user meet- ings and improves multi-agent collaboration. 2 Related Work Misalignment Detection in Dialogue. Prior work studies how participants lose track of shared understanding during a conversation. Some ap- proaches track common ground from the per- spective of each participant, modeling what each speaker takes to be mutually accepted (Markowska et al., 2023; Khebour et al., 2024; Mohapatra et al., 2024; Li et al., 2025). Others label conversational frictions, deliberation cues, or grounding rifts as signals of misalignment, sometimes drawing on gesture and prosody (Karadzhov et al., 2023; Nath et al., 2024; Sarkar et al., 2025; Shaikh et al., 2025; Inan et al., 2025; Nath et al., 2025; VanderHoeven et al., 2025). A separate thread treats disagree- ment as an explicit stance and detects it via clas- sification (Li et al., 2023; Wagner et al., 2025; Ja- madi Khiabani and Zubiaga, 2025; Togay et al., 2026). What unites these efforts is that misalign- ment leaves a trace, such as a token, gesture, or hesitation that an annotator can point to. IoA leaves no such trace, since the words on the surface look like agreement and there is nothing to flag. Theory of Mind (ToM) in Dialogue. Work on ToM in LLMs falls into three groups. Early bench- marks adapt the Sally-Anne task to text, asking models to follow character beliefs through short stories (Le et al., 2019; Wu et al., 2023; Gandhi et al., 2023; Xu et al., 2024). Later work moves to dialogue, testing belief and knowledge infer- ence under information asymmetry, multi-party ex- change, or persuasion (Bara et al., 2021; Kim et al., 2023; Chen et al., 2024). A recent line places ToM 2 GolearnTorchbeforedebuggingit. Sure! I’l learn Torch this week. Great! Let’s sync next Monday. Lead Engineer Lead A. Direct prompting missesIoA Direct prompting Judgement Looksaligned Hidden disagreement missed B. Multiple-choice question probing reveals hidden disagreement (Ours) 1 Input Observable transcriptu(흉) Public profiles 풓 풊 + 2 IoA Prober DiagnosticMCQ What should the engineer prepare? A. Lua Torch B. PyTorch C. Either is fine Nocontradiction Norepair Nohedge Lead’s construal Torch = Lua Torch require ‘torch’ require ‘n’ model = torch.load(x) Engineer’s construal Torch = PyTorch import torch import torch.n as n model = n.Module(x) 3 Answer A B 4 Signal A B Divergentanswers → Objectivesignal Mechanically decidable 5 Validity Filter Meaningful Implicit Non-redundant Validated hidden disagreement → → Figure 2: Overview of IoA detection through multiple-choice question (MCQ) probing. (1) Panel A illustrates that direct prompting judges the dialogue as aligned when the transcript contains no contradiction, repair, or hedge, thereby missing latent disagreement. (2) Panel B shows that our proposed IoA Prober generates MCQ from the observable transcript and public profiles. Independent answers expose divergent construals over the same dialogue. The validity filter is performed by participants in real-world deployment and automated in IoA-Suite for evaluation. inside user-agent interaction, where the agent must square latent user beliefs with task or environment states (Qiu et al., 2024; Jafari et al., 2025; Ruan et al., 2026). Two assumptions run through all of them: the evaluation specifies where the model should attend, and answers are checked against an external ground truth. IoA detection violates both: the divergence holds between participants rather than against an external truth, and identify- ing where it occurs is itself the task. 3 Problem Formalization To make illusion of alignment amenable to detec- tion, we formalize it as a latent mismatch between how participants construe the same dialogue at its close. The mismatch is latent because the transcript registers no friction, yet it determines whether par- ticipants would commit to incompatible follow-ups on issues at stake in the dialogue. Participants. Consider a collaborative dialogue amongn ≥ 2participantsP = p 1 ,...,p n . Eachp i is characterized by three layers of context. A public profiler i captures attributes observable to others, including role and declared expertise. A private agenda g i captures the goals and priorities p i pursues without articulating them in the meeting. A set of tacit assumptionsa i captures the premises p i treats as already settled and therefore does not consider worth stating. The triple(r i ,g i ,a i )shapes the utterancesp i contributes and the commitments p i would later make. Dialogues.Conditioned on the contexts of all par- ticipants, the conversation unfolds as a sequence of turnsτ = (t 1 ,...,t T ), where at each turnt k a speakers k ∈ Pforms an inner thoughth k and pro- duces a publicly observable utteranceu k . We write u(τ ) = (u 1 ,...,u T ) for the observable transcript. Illusion of Alignment.Each participantp i holds a post-dialogue construal of what was agreed and what should follow, represented as c i (ι) =C(ι; r i ,g i ,a i ,u(τ )), whereι ∈ Iindexes the commitments at issue in the dialogue andc i (ι) ∈ Yis the commitment p i would make onι, jointly shaped by the three context layers ofp i and the shared transcript. The dialogue exhibits an illusion of alignment whenτ ends in apparent agreement but two or more con- struals diverge on some issue. The set of such divergence points, M (τ ) =ι∈I |∃i̸= j, c i (ι)̸= c j (ι), constitutes the latent disagreements onτ. Surfacing the elements ofM (τ )in a verifiable manner is the detection task formulated in Section 4. 4 Detection and Evaluation: IoA-Suite We now address RQ1 by converting the latent dis- agreement defined in Section 3 into a verifiable detection task. We introduce IoA-Suite, a frame- work for IoA detection comprising three compo- 3 nents: a detection method (Section 4.1), an evalua- tion protocol (Section 4.2), and a quality-controlled synthesized dataset (Section 4.3). 4.1 Detection Method The central design constraint is to expose elements ofM (τ )without any model deciding whether two participants really disagree. A transcript-level sur- face detector cannot meet this constraint: when the dialogue ends in apparent agreement, the absence of contradiction, repair, or hedging is by construc- tion the very condition under which IoA arises, so a detector built on these surface signals would conclude that the dialogue is aligned (Figure 2, Panel A). We instead anchor the disagreement la- bel in observable participant behavior, restricting model role to proposing candidate questions that the participants themselves separate by answering. Multiple-choice question (MCQ) probing.We propose MCQ probing, as illustrated in Figure 2, Panel B. An IoA ProberPgenerates a set of candi- date diagnostic questions from public context, P u(τ ), r i n i=1 = Q =q 1 ,...,q K , mirroring an external observer who has only the transcript and the declared identities of the speak- ers. Each participant then answers everyq ∈ Q from the perspective ofp i , yieldingc i (q) n i=1 , and qis divergent whenc i (q)̸= c j (q)for somei̸= j. The multiple-choice form is essential here: it fixes the answer space in advance, so divergence reduces to a mechanical check of whether two participants selected different options, without requiring any ex- ternal judgment to compare phrasings for substan- tive disagreement. Disagreement existence is there- fore never adjudicated by a model verdict, which makes IoA detection objectively verifiable. 4.2 Evaluation Protocol The MCQ probing method above yields a diver- gence label per question, but not every divergent question reflects a substantive latent disagreement. We therefore introduce a validity filterVto au- tomatically remove divergent questions that fail to constitute genuine hidden disagreements. In real-world meetings, this role is played by the par- ticipants themselves (Section 6.2). Algorithm 1 instantiates the full evaluation pipeline. Participant simulation.Each construalc i is real- ized by a role-playing simulatorS(Shanahan et al., Algorithm 1 Evaluation Protocol Require: (τ,M(τ),r i ,g i ,a i n i=1 ); IoA ProberP; Simula- tor S; Validity filterV Ensure: P , R, F1 1: Q←P(u(τ), r i n i=1 ) 2: for i = 1,...,n do 3:o i ← S(Q; r i ,g i ,a i ,h i (τ); r j j̸=i ; u(τ)) 4: end for 5: Q div ←q k ∈ Q :¬ all-equal(o 1 [k],...,o n [k]) 6: (acc k , hit k ) q k ∈Q div ← V(Q div , o i , τ, M(τ), r i ,g i ,a i ) 7: A,B,C ←∅ 8: for q k ∈ Q do 9:if q k /∈ Q div ∨¬ acc k then C ← C∪q k 10:else if hit k ∈ M(τ) then A← A∪(q k , hit k ) 11:else B ← B∪q k 12:end if 13: end for 14: P ←|A|/(|A| +|C|) 15: R←|m∈ M(τ) :∃ (q,m)∈ A|/|M(τ)| 16: return (P, R, 2PR/(P + R)) 2023) given the full context ofp i (r i ,g i ,a i and inner thoughtsh i (τ )) and only the public profiles r j j̸=i plus the shared transcriptu(τ )for others, matching whatp i actually holds during the conver- sation. We invokeSonce per participant, since a single roll-out already matches the majority vote over repeated samples (Appendix C.1). Validity filter. For each divergent questionq ∈ Q div ,Vchecks three criteria: meaningfulness (the answer divergence reflects a real cognitive gap rather than wording noise), implicitness (qdoes not target a point already explicitly debated inτ), and non-redundancy (qdoes not restate a divergence already covered by an earlier question inQ). Ifq passes all three,Vattempts to attribute it to some m ∈ M (τ ). Crucially,Vnever decides whether participants disagree, which is settled mechanically by simulator outputs. Metrics. Eachqfalls into categoryA(passesV and attributed),B(passesVbut unattributed), or C(non-divergent or rejected).Barises because dialogue generation may introduce emergent mis- alignments beyond the plantedM (τ ), which are real thatVaccepts but they have no ground-truthm to attribute against. We excludeBfrom precision and recall to keep the score stable, and define F1 as the harmonic mean ofPandR.|B|is reported separately in Appendix D.2 as a complementary axis tracking divergences beyond the planted set. The simulator uses Gemini-3-Flash (Google, 2026) and the judge uses GPT-5.4 (OpenAI, 2026). Full prompts and settings are in Appendix C. 4 Algorithm 2 IoA-Suite Dataset Construction Require:Seed(d,k,x); GeneratorG; JudgeJ; Thresholds θ 1 ,θ 2 Ensure: Dialogue τ with Ground-truth M(τ), or DISCARD // Stage 1: participants and planted misalignments 1: (r i ,g i ,a i n i=1 , M(τ))← G participants (d,k,x) 2: s 1 ← J participants (r i ,g i ,a i ,M(τ)) 3: if s 1 < θ 1 then 4:regenerate once with feedback; update s 1 5:if s 1 < θ 1 then return DISCARD 6: end if // Stage 2: dialogue generation under latent constraint 7: τ ← G dialogue (r i ,g i ,a i ,M(τ)) 8: s 2 ← J dialogue (τ,r i ,g i ,a i ,M(τ)) 9: if s 2 < θ 2 then 10:regenerate once with feedback; update s 2 11:if s 2 < θ 2 then return DISCARD 12: end if 13: return (τ,M(τ),r i ,g i ,a i ) 4.3 Dataset Construction Since the latent disagreement setM (τ )cannot be recovered from naturally occurring transcripts, we synthesize the dialogues in IoA-Suite by planting M (τ )beforeτis generated, so that ground truth persists by construction. Pipeline.Algorithm 2 summarizes the two-stage procedure with stage-wise filtering. Stage 1 in- stantiates two participants together with a planted M (τ ), at least one element of which is definitional, anchored to a shared term that participants use while privately referring to different things; this is the form of IoA hardest to surface with transcript- level signals and matches the canonical example of Figure 1. Stage 2 generatesτunder a hard con- straint forbidding any utterance from explicitly de- batingM (τ ), keeping the planted disagreements latent. After each stage, a judge scores its output on stage-specific dimensions and routes failures to one round of regeneration before discarding. Both generator and judge are GPT-5.4 (OpenAI, 2026); full rubrics and prompts are in Appendix B. Seeds and coverage.Each dialogue is grounded in a real-world collaborative scenario specified by a seed tuple(d,k,x), where(d,k)specifies the do- main and task type, andxprovides the substantive content for participants to discuss. We instanti- atexas a scholarly publication, which anchors each dialogue in a concrete work-relevant scenario with enough substantive content to support genuine divergent interpretations rather than generic role- play. We draw from five task types in the Group Task Circumplex (McGrath, 1984), namely plan- ning, design review, troubleshooting, coordination, Direct promptingMCQ probing (ours) ModelPRF1PRF1 Closed-source GPT-5.429.636.132.244.757.749.5 (+17.3) Gemini-3.1-Pro † 35.832.433.846.540.443.0 (+ 9.2) Gemini-3-Flash24.327.125.343.940.441.9 (+16.6) GPT-5 mini20.336.025.834.947.840.0 (+14.2) Open-source DeepSeek-V3.222.728.225.042.843.642.9(+17.9) GLM-5.124.729.526.641.338.639.3 (+12.7) Qwen3.5-397B † 21.723.522.446.334.638.1 (+15.7) Kimi-K2.520.931.124.935.539.837.1 (+12.2) Qwen3-8B23.321.722.337.523.328.0 (+ 5.7) Table 1: Performance (%) on IoA-Suite test split un- der direct prompting and MCQ probing (our proto- col from Section 4.1). The green annotation next to each MCQ probing F1 reports the gain over the direct prompting baseline on the same model. Bold: best over- all.Underline: best within open-source. † Abbreviated names: Gemini-3.1-Pro for Gemini-3.1-Pro Preview, Qwen3.5-397B for Qwen3.5-397B-A17B; the same ab- breviations apply throughout the paper. and brainstorming, paired with six domains: ML research, software engineering, cross-disciplinary work, medicine, business, and law. Train, valida- tion, and test splits draw from disjoint seed pools, eliminating leakage at the level of source material. Statistics and quality. The procedure yields 1,200 training, 120 validation, and 300 test dia- logues. Since GPT-5.4 supplies both generator and judge, self-enhancement bias (Zheng et al., 2023; Panickssery et al., 2024) is a concern, which we address with two complementary external checks on the test set: re-scoring by Gemini-3.1-Pro Pre- view (Google, 2026), and parallel evaluation by two trained PhD annotators. As illustrated in Ap- pendix B.3, both the cross-family automated re- scoring and independent human annotation corrob- orate the quality of IoA-Suite. 5 IoA Detection Bottleneck Diagnostic With IoA-Suite in place, we ask how current mod- els perform on IoA detection, and trace the bottle- neck to the information they have access to. 5.1 IoA Detection Is Unsolved at the Frontier We evaluate four closed-source and five open- source models on IoA-Suite under two protocols: direct prompting, where the model is asked to list hidden disagreements in the dialogue, and MCQ probing, our protocol from Section 4.1. Results 5 appear in Table 1; full prompts, results, hyperpa- rameters are in Appendix C. MCQ probing is effective for IoA detection. Every model gains substantially from the MCQ mechanism: F1 improves by 5.7% to 17.9% across the nine models tested under both protocols, with a mean gain of 13.5%. Crucially, recall also rises in every case (mean +11.2%), so the gain is not driven by inflated question counts. The pattern confirms the design rationale of Section 4.1: anchoring de- tection in participant behavior, rather than asking a model to declare disagreements directly, produces a more reliable signal for IoA. A low ceiling across families.No model exceeds 50.0% F1 in IoA-Suite, with the strongest reach- ing only 49.5%. The gap between closed- and open-source families is narrow: DeepSeek-V3.2 trails GPT-5.4 by just 6.6% F1, far less than typ- ical reasoning-benchmark gaps. The low ceiling and narrow spread suggest that IoA detection is a capability current models broadly lack, and one unlikely to emerge from general reasoning scaling. Question-asking strategies diverge across fami- lies.Beyond aggregate F1, the average number of questions generated per dialogue (Kin Table 15) under MCQ probing reveals two opposing strate- gies. GPT-series models probe broadly (K = 5.8 for both GPT-5.4 and GPT-5 mini), trading pre- cision for recall; Gemini-series models probe se- lectively (K = 3.5∼ 3.8) at higher precision but lower recall. This trade-off is inherent to prompting alone: neither family achieves high recall without inflatingK, motivating an objective that directly optimizes both axes (Section 6). 5.2 Private Context Matters To localize the bottleneck, we lift the defaultr i - only restriction along two axes: turn-level in- ner thoughtsh k and static private context(g i ,a i ), evaluating GPT-5.4, Gemini-3.1-Pro Preview, and Qwen3-8B (Team, 2025) under the resulting four settings (Table 2; full prompts in Appendix D.5). Either axis of private context alone is enough to transform performance, lifting F1 by over 20.0% absolute across all three models, while combining both yields only marginal further gains. This redun- dancy indicates that the two axes signal the same underlying quantity: the private perspective each participant holds on the conversation. Once the pri- vate context is available, the three models converge ModelSettingPRF1 GPT-5.4 default44.757.749.5 + h k 72.077.274.1 (+24.6) + g i ,a i 80.384.382.0 (+32.5) + g i ,a i ,h k 85.383.984.4 (+34.9) Gemini-3.1-Pro † default46.540.443.0 + h k 80.072.275.4 (+32.4) + g i ,a i 83.181.482.0 (+39.0) + g i ,a i ,h k 80.085.082.1 (+39.1) Qwen3-8B default37.523.328.0 + h k 62.044.450.9 (+22.9) + g i ,a i 81.872.375.8 (+47.8) + g i ,a i ,h k 80.773.476.4 (+48.4) Table 2: Effect of supplying oracle private context to the prober. Thedefaultrow exposes onlyr i and u(τ ), mirroring the deployment setting evaluated in Sec- tion 5.1. Subsequent rows additionally expose turn-level inner thoughtsh k , private agenda and tacit assumptions (g i ,a i ), or both. Green annotations report the gain over the default for the same model (all values in %). to a narrow band (75.8%∼84.4%), with Qwen3-8B exceeding every model in the default setting. Two implications follow. First, IoA detection is fundamentally a Theory-of-Mind task: it requires recovering the private perspective each participant holds rather than reasoning more carefully over what is already said, which reframes the training objective in Section 6 as learning to infer this per- spective from the public transcript alone. Second, IoA-Suite is well-calibrated: it stays hard under the realistic public-only input yet becomes tractable once oracle private context is supplied, confirming that the gap reflects information access rather than an artifact of the construction. 6 Improving IoA Detection: IoA-Prober The results in Section 5 show that IoA detection remains difficult when only public context is avail- able. We address RQ2 by training a prober directly on the diagnostic question generation objective. 6.1 Training Recipe and Main Results Training recipe. IoA-Prober-8B is trained from Qwen3-8B in two stages. The first is a supervised warm-start on 300 dialogues sampled from the training split: for each dialogue we collect can- didate questions from GPT-5.4 and resolve through the pipeline of Section 4.2, retaining Categories A and B (divergent questions accepted by the va- lidity filter, with and without ground-truth attribu- tion respectively) and dropping Category C (non- divergent or filter-rejected questions) which would 6 ModelPRF1 Frontier reference GPT-5.444.757.749.5 Gemini-3.1-Pro † 46.540.443.0 Base model and our method Qwen3-8B (base)37.523.328.0 IoA-Prober-8B (ours)50.553.951.8 (+23.8) Table 3: IoA-Prober-8B reaches frontier-level F1 start- ing from a 28.0% Qwen3-8B base. Green annotation reports the absolute gain over the Qwen3-8B base. push the model toward uninformative outputs. The second stage applies GRPO (Shao et al., 2024) with F1 as the reward, sampling a group of candidate question sets per dialogue and scoring each end-to- end through the same pipeline: J GRPO (π θ ) =E τ∼D,Q j G j=1 ∼π θ 1 G G X j=1 A j logπ θ (Q j | τ ) whereA j is the group-normalized advantage ofQ j under the F1 reward. The second stage runs on the training split for two epochs; full hyperparameters and results are in Appendix D. Balanced precision and recall lift IoA-Prober-8B to frontier-level F1. Starting from a Qwen3-8B base at 28.0% F1, IoA-Prober-8B reaches 51.8% (Table 3), an absolute gain of 23.8 points. The F1 reward steers IoA-Prober-8B toward a balanced precision-recall profile (50.5% / 53.9%), in con- trast to the asymmetry inherent to prompting alone, where GPT-5.4 over-probes (44.7% / 57.7%) and Gemini-3.1-Pro Preview under-probes (46.5% / 40.4%). This balance translates into an F1 that exceeds Gemini-3.1-Pro Preview by 8.8 points (p < 0.001via paired bootstrap (Koehn, 2004)) and matches GPT-5.4 within 2.3 points (p = 0.14). A case study and failure mode analysis in Ap- pendix E show that the remaining misses are domi- nated by misalignments requiring private-context access (consistent with the oracle ceiling in Sec- tion 5.2) and by recoverable failures such as intra- set redundancy that future budget or diversity ob- jectives could address. Training narrows the gap to the oracle setting. Section 5.2 established an oracle-context upper bound of 76.4% F1 for Qwen3-8B when private context is supplied at inference. Training closes the gap between the public-only base and this ora- cle ceiling by roughly half (from 48.4% to 24.6%), Objective (per session)Likert DetectorValidValidityBestDisc.Use Qwen3-8B1.6788.2%73.053.12 GPT-5.41.8994.4%93.163.12 IoA-Prober-8B2.8994.5%273.653.65 Table 4: Real-user meeting study across 18 sessions and 43 participants. Valid: validated hidden disagreements per session; Validity: fraction of divergent questions judged valid. Best: votes for the best blinded set. Disc. and Use: post-session Likert ratings (1∼5) for misalign- ment discovery and willingness to use. indicating that IoA-Prober-8B has learned to infer a substantial part of the private perspective each participant holds from the public context alone. 6.2 Real-User Meeting Beyond benchmark performance, we ask two ques- tions: whether IoA extends to real human collab- oration, and whether IoA-Prober-8B transfers be- yond LM-simulated dialogue. We apply three de- tectors (Qwen3-8B, GPT-5.4, IoA-Prober-8B) to transcripts from 18 real working meetings with 43 participants. For each meeting, participants in- dependently answered three blinded question sets, then marked each question whose answers diverged from another participant as valid (a genuine hidden disagreement) or invalid (already discussed, redun- dant, or lacking meaningful option distinction). A post-session survey collected a best-set vote and Likert ratings on misalignment discovery and will- ingness to use. Full protocol and meeting metadata appear in Appendix D.4. IoA is pervasive in real human collaboration. Across the 18 sessions, IoA-Prober-8B surfaces 2.89 validated hidden disagreements per meeting, with 94.5% of divergent questions judged valid by the participants themselves. Participants confirmed these disagreements had not been voiced during the original discussion, despite all parties believing they had reached consensus. This rate confirms that IoA is not a synthetic artifact of LM-generated dia- logue but a routine feature of human collaboration that participants cannot surface unaided. IoA-Prober-8B leads on real human dialogue. IoA-Prober-8B surfaces significantly more vali- dated disagreements per meeting than GPT-5.4 (+1.00, 95% CI[0.28, 1.67], paired bootstrapp = 0.005 ) and Qwen3-8B (+1.22, 95% CI[0.61, 1.78], p < 0.001). In the blinded forced-choice survey, 7 VariantPRF1K Full recipe50.553.951.84.5 w/o RL (SFT on full train dataset) with filterC41.721.126.6 (-25.2)2.2 no filterC39.037.838.1 (-13.7)4.0 w/o SFT warm-start31.931.531.3 (-20.5)4.1 w/o filterC in warm-start43.345.043.8 (- 8.0)4.4 Table 5: Training ablations on IoA-Suite. Red annota- tions report the F1 drop relative to the full recipe. IoA-Prober-8B receives 27 of 43 votes and wins 10 of 18 sessions by three-way majority and 12 of 18 in pairwise comparison (paired bootstrap vs. GPT- 5.4p = 0.002, vs. Qwen3-8Bp = 0.008). The ad- vantage also holds on subjective ratings, with both Discovery and Would-Use Likert scores roughly 0.5 points above either baseline. Full results and analysis are in Appendix D.4. 6.3 Ablation Study We ablate the full recipe along three axes: remov- ing RL stage, removing SFT warm-start, and re- moving Category-C filtering from the warm-start data. All other settings match the full recipe (Ap- pendix D.5). Table 5 reports the ablation results. RL drives the gains, but only when anchored by SFT. The two stages exhibit asymmetric de- pendency. Without SFT warm-start, RL struggles to receive informative learning signals and barely improves over the 28.0% base. Without RL, SFT alone ceilings at 38.1%, well below the full recipe. SFT teaches what plausible probing questions look like; RL learns what to look for among them. Only the composition unlocks the 51.8% F1. Category-C filtering shapes a precision prior that RL expands along recall.Under SFT alone, filterC yields a high-precision but recall-starved model (41.7% / 21.1%): the filtered data teaches what a good question looks like, but SFT can only imitate this sparse distribution. In the full recipe, RL inherits this precision prior from the warm-start and expands it along the recall axis, lifting F1 from 26.6% to 51.8%. Removing filterC from warm- start instead gives GRPO a noisier starting point and caps the final F1 at 43.8%. 6.4 Generalization Study To test whether surfacing latent disagreements improves downstream outcomes, we instantiate IoA detection in multi-agent collaboration on ConfigurationBCB-HardHiddenBench Baselines Single agent19.6– Multi-agent discussion21.624.6 Direct prompting + self-reflection21.0 (-0.6)26.2 (+1.6) MCQ probing w/ Qwen3-8B21.6 (+0.0)27.7 (+3.1) w/ GPT-5.425.0 (+3.4)33.8 (+9.2) w/ IoA-Prober-8B25.0 (+3.4)35.4 (+10.8) Table 6: Pass@1 (%) on BigCodeBench-Hard and Hid- denBench. Annotations report the change relative to multi-agent discussion. HiddenBench is multi-agent by construction; the single-agent row does not apply. BigCodeBench-Hard (Zhuo et al., 2025) and Hid- denBench (Li et al., 2026). For BCB-Hard we build a symmetric two-agent code-generation framework, while HiddenBench provides its own multi-agent protocol. On top of multi-agent discussion we eval- uate two interventions: direct prompting asks the agents to reconsider whether disagreements remain before finalizing, and MCQ probing uses a detec- tor that generates MCQs to surface residual diver- gence. Either intervention, once triggered, sends the agents into an additional discussion round. All agents use Qwen3-8B as the base model; under MCQ probing we test three detectors: Qwen3-8B, GPT-5.4, and IoA-Prober-8B. Benchmark details and the evaluation protocol are in Appendix D.6. IoA detection transfers to multi-agent collabo- ration.MCQ probing with capable detectors de- livers substantial gains over multi-agent discussion: IoA-Prober-8B matches GPT-5.4 on BCB-Hard (+3.4%) and exceeds it on HiddenBench (+10.8% vs. +9.2%), while a weak detector (Qwen3-8B) and direct self-reflection yield only marginal changes. Surfacing latent disagreements therefore translates into measurable task gains on multi-agent collabo- ration tasks, with the 8B specialized prober reach- ing parity with frontier closed-source models. 7 Conclusion We introduced the illusion of alignment (IoA), ap- parent agreement masking latent disagreement, and recast its detection as generating multiple-choice questions whose divergent answers serve as behav- ioral evidence. We built IoA-Suite and trained IoA- Prober-8B, an 8B detector with balanced precision and recall that surfaces hidden disagreements in real-user meetings and improves multi-agent col- 8 laboration. Our work opens a path for studying collaborative dialogue failures that leave no observ- able cues, and for building agents that surface and resolve them before they propagate downstream. Limitations We discuss several limitations of this work and directions they open for future research. Model scale. Due to computational budget, we train a single prober at the 8B scale and do not investigate whether the recipe in Section 6.1 scales to larger backbones. We note, however, that IoA- Prober-8B already reaches parity with frontier closed-source models on IoA-Suite and surpasses them in real-user meetings, demonstrating the ef- fectiveness of the proposed recipe. Dyadic dialogues.IoA-Suite instantiates two par- ticipants per dialogue. We chose dyadic configu- rations because reliably synthesizing longer multi- party dialogues that maintain coherent latent dis- agreements across more participants remains be- yond the controllable generation capacity of cur- rent models, and restricting to two participants al- lows us to construct and verify the planted ground truthM (τ )at the quality required for benchmark- ing. The trained IoA-Prober-8B nevertheless gen- eralizes to multi-party settings, as shown in Sec- tion 6.2, where 8 session sizes exceed two par- ticipants. Constructing realistic multi-party col- laborative dialogues with verifiable ground-truth misalignments is left as future work. Synthesized dialogues.IoA-Suite is constructed by planting misalignments before dialogue gener- ation, which enables verifiable ground truth but constrains the distribution of latent disagreements to those that can be specified in advance. Naturally occurring IoA in real collaboration may take forms beyond our seeded taxonomy, particularly in long- horizon projects where misalignments accumulate across multiple meetings. The real-user study in Section 6.2 partially addresses this by evaluating on naturally occurring dialogue, but constructing a fully natural IoA benchmark with verifiable ground truth remains an open problem. Language coverage.All dialogues in IoA-Suite and the real-user study are conducted in English. We look forward to extending the diagnostic ques- tion generation objective to other languages in fu- ture work, particularly those with different con- ventions for indirectness and explicit disagreement, where the manifestation of illusion of alignment may take substantially different forms. Ethical Considerations Human subjects research. The real-user meet- ing study described in Section 6.2 was conducted with prior informed consent from all participants, who were briefed on the purpose of the study, the use of meeting transcripts, and their right to with- draw at any stage. Participants were compensated for their time at a rate commensurate with local standards. All meeting transcripts were retained on the participant side rather than centrally col- lected, and only the diagnostic question outputs and anonymized validity annotations were shared with the research team. No personally identifiable information appears in any released artifact. Synthetic dialogue data. IoA-Suite is con- structed by prompting LLMs to generate partic- ipant profiles and dialogues, which may inherit bi- ases from the underlying models, including stereo- typed associations between professional roles and demographic attributes. We mitigate this risk by anchoring each seed to a published work rather than to demographic descriptors and by applying quality checks across model families and human annotators (Section 4.3). Users should nonetheless treat the synthesized personas as illustrative rather than representative of any real population. Intended use and potential misuse.IoA-Prober- 8B is designed for symmetric collaborative settings in which all participants share a goal of mutual un- derstanding and have equal access to the outputs of the tool. Three deployment patterns fall outside this scope and we explicitly advise against them: (1) asymmetric monitoring, where one party (e.g., an employer, platform, or moderator) analyzes a di- alogue without the knowledge or symmetric access of other participants; (2) evaluative use, where IoA- Prober-8B outputs feed into performance review, hiring, or compensation decisions, since the in- ferred private agendas are model conjectures rather than verified beliefs; and (3) adversarial elicitation, where MCQs are deliberately crafted to manufac- ture rather than surface divergence. Concretely, we recommend that any deployment (1) obtain prior informed consent from every dialogue participant, including disclosure that a model is inferring their unstated assumptions; (2) make the generated ques- 9 tions and aggregated results visible to all partici- pants, not only to a meeting organizer; and (3) treat any single MCQ split as a prompt for discussion. We will release IoA-Prober-8B under a responsible- use license (RAIL-style) that prohibits surveillance, employment-decision, and other asymmetric uses, and we note that no license can fully prevent mis- use; the safeguards above must be enforced at the deployment layer. References Cristian-Paul Bara, Sky CH-Wang, and Joyce Chai. 2021. MindCraft: Theory of mind modeling for situ- ated dialogue in collaborative tasks. In Proceedings of the 2021 Conference on Empirical Methods in Nat- ural Language Processing, pages 1112–1125, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. Janis A Cannon-Bowers, Eduardo Salas, and Sharolyn Converse. 1993. Shared mental models in expert team decision making. Zhuang Chen, Jincenzi Wu, Jinfeng Zhou, Bosi Wen, Guanqun Bi, Gongyao Jiang, Yaru Cao, Mengting Hu, Yunghwei Lai, Zexuan Xiong, and Minlie Huang. 2024. ToMBench: Benchmarking theory of mind in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), pages 15959–15983, Bangkok, Thailand. Association for Computational Linguistics. Christine De Kock and Andreas Vlachos. 2021. I beg to differ: A study of constructive disagreement in on- line conversations. In Proceedings of the 16th Con- ference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2017–2027, Online. Association for Computational Linguistics. DeepSeek-AI. 2025. DeepSeek-V3.2: Pushing the frontier of open large language models.CoRR, abs/2512.02556. Kanishk Gandhi, Jan-Philipp Fränken, Tobias Gersten- berg, and Noah Goodman. 2023. Understanding social reasoning in language models with language models. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Bench- marks Track. GLM. 2026. GLM-5: from vibe coding to agentic engineering. CoRR, abs/2602.15763. Google. 2026. Gemini 3 Flash Model Card. Google. 2026. Gemini 3.1 Pro. Mert Inan, Anthony Sicilia, Suvodip Dey, Vardhan Don- gre, Tejas Srinivasan, Jesse Thomason, Gökhan Tür, Dilek Hakkani-Tür, and Malihe Alikhani. 2025. Bet- ter slow than sorry: Introducing positive friction for reliable dialogue systems. CoRR, abs/2501.17348. Mehdi Jafari, Yuncheng Hua, Hao Xue, and Flora D. Salim. 2025. Beyond words: Integrating theory of mind into conversational agents for human-like be- lief, desire, and intention alignment. In Findings of the Association for Computational Linguistics: ACL 2025, pages 5489–5508, Vienna, Austria. Associa- tion for Computational Linguistics. Parisa Jamadi Khiabani and Arkaitz Zubiaga. 2025. Cross-target stance detection: A survey of techniques, datasets, and challenges. Expert Syst. Appl., 283(C). Georgi Karadzhov, Tom Stafford, and Andreas Vlachos. 2023. Delidata: A dataset for deliberation in multi- party problem solving. Proc. ACM Hum.-Comput. Interact., 7(CSCW2). Ibrahim Khalil Khebour, Kenneth Lai, Mariah Brad- ford, Yifan Zhu, Richard A. Brutti, Christopher Tam, Jingxuan Tu, Benjamin A. Ibarra, Nathaniel Blan- chard, Nikhil Krishnaswamy, and James Pustejovsky. 2024. Common ground tracking in multimodal dia- logue. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 3587–3602, Torino, Italia. ELRA and ICCL. Hyunwoo Kim, Melanie Sclar, Xuhui Zhou, Ronan Bras, Gunhee Kim, Yejin Choi, and Maarten Sap. 2023. FANToM: A benchmark for stress-testing machine theory of mind in interactions. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14397–14413, Singa- pore. Association for Computational Linguistics. Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 388–395, Barcelona, Spain. Association for Computational Linguistics. Matthew Le, Y-Lan Boureau, and Maximilian Nickel. 2019. Revisiting the evaluation of theory of mind through question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natu- ral Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5872–5877, Hong Kong, China. Association for Computational Linguistics. Patrick M Lencioni. 2002. The five dysfunctions of a team: A leadership fable. John Wiley & Sons. Nan Li, Albert Gatt, and Massimo Poesio. 2025. Grounded misunderstandings in asymmetric dia- logue: A perspectivist annotation scheme for map- task. CoRR, abs/2511.03718. Yingjie Li, Krishna Garg, and Cornelia Caragea. 2023. A new direction in stance detection: Target-stance extraction in the wild. In Proceedings of the 61st An- nual Meeting of the Association for Computational 10 Linguistics (Volume 1: Long Papers), pages 10071– 10085, Toronto, Canada. Association for Computa- tional Linguistics. Yuxuan Li, Aoi Naito, and Hirokazu Shirado. 2026. Hiddenbench: Assessing collective reasoning in multi-agent LLMs via hidden profile tasks. Magdalena Markowska, Mohammad Taghizadeh, Adil Soubki, Seyed Mirroshandel, and Owen Rambow. 2023. Finding common ground: Annotating and pre- dicting common ground in spoken conversations. In Findings of the Association for Computational Lin- guistics: EMNLP 2023, pages 8221–8233, Singapore. Association for Computational Linguistics. J.E. McGrath. 1984. Groups: Interaction and Perfor- mance. Prentice-Hall. Biswesh Mohapatra, Seemab Hassan, Laurent Romary, and Justine Cassell. 2024. Conversational ground- ing: Annotation and analysis of grounding acts and grounding units. In Proceedings of the 2024 Joint International Conference on Computational Linguis- tics, Language Resources and Evaluation (LREC- COLING 2024), pages 3967–3977, Torino, Italia. ELRA and ICCL. Abhijnan Nath, Carine Graff, Andrei Bachinin, and Nikhil Krishnaswamy. 2025. Frictional agent align- ment framework: Slow down and don’t break things. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11042–11089, Vienna, Austria. Association for Computational Linguistics. Abhijnan Nath, Videep Venkatesha, Mariah Bradford, Avyakta Chelle, Austin C. Youngren, Carlos Mabrey, Nathaniel Blanchard, and Nikhil Krishnaswamy. 2024. “any other thoughts, hedgehog?” linking deliberation chains in collaborative dialogues. In Findings of the Association for Computational Lin- guistics: EMNLP 2024, pages 5297–5314, Miami, Florida, USA. Association for Computational Lin- guistics. OpenAI. 2026. GPT-5.4 thinking system card. OpenAI. 2026. OpenAI GPT-5 system card. CoRR, abs/2601.03267. Arjun Panickssery, Samuel R. Bowman, and Shi Feng. 2024. LLM evaluators recognize and favor their own generations. In Proceedings of the 38th Interna- tional Conference on Neural Information Processing Systems, NIPS ’24, Red Hook, NY, USA. Curran Associates Inc. David Premack and Guy Woodruff. 1978. Does the chimpanzee have a theory of mind? Behavioral and Brain Sciences, 1(4):515–526. Shuwen Qiu, Mingdian Liu, Hengli Li, Song-Chun Zhu, and Zilong Zheng. 2024. MindDial: Enhancing con- versational agents with theory-of-mind for common ground alignment and negotiation. In Proceedings of the 25th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 746–759, Kyoto, Japan. Association for Computational Lin- guistics. Minyuan Ruan, Ziyue Wang, Kaiming Liu, Yungh- wei Lai, Peng Li, and Yang Liu. 2026. Beyond words: Evaluating and bridging epistemic divergence in user-agent interaction via theory of mind. CoRR, abs/2602.13832. Rupak Sarkar, Neha Srikanth, Taylor Pellegrin, Rachel Rudinger, Claire Bonial, and Philip Resnik. 2025. Understanding common ground misalignment in goal-oriented dialog: A case-study with Ubuntu chat logs. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), pages 3200–3215, Vienna, Austria. Association for Computational Linguistics. Omar Shaikh, Hussein Mozannar, Gagan Bansal, Adam Fourney, and Eric Horvitz. 2025. Navigating rifts in human-LLM grounding: Study and benchmark. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 20832–20847, Vienna, Austria. Association for Computational Linguistics. Murray Shanahan, Kyle McDonell, and Laria Reynolds. 2023. Role play with large language models. Nature, 623(7987):493–498. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. CoRR, abs/2402.03300. Kimi Team. 2026a. Kimi K2.5: visual agentic intelli- gence. CoRR, abs/2602.02276. Qwen Team. 2025. Qwen3 technical report. CoRR, abs/2505.09388. Qwen Team. 2026b. Qwen3.5: Accelerating productiv- ity with native multimodal agents. Özgür Togay, Florian Kunneman, Javier Garcia- Bernardo, and Anastasia Giachanou. 2026. Large language models unpack complex political opin- ions through target-stance extraction.CoRR, abs/2603.23531. Hannah VanderHoeven, Brady Bhalla, Ibrahim Khebour, Austin C. Youngren, Videep Venkatesha, Mariah Bradford, Jack Fitzgerald, Carlos Mabrey, Jingx- uan Tu, Yifan Zhu, Kenneth Lai, Changsoo Jung, James Pustejovsky, and Nikhil Krishnaswamy. 2025. TRACE: Real-time multimodal common ground tracking in situated collaborative dialogues. In Pro- ceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Compu- tational Linguistics: Human Language Technologies (System Demonstrations), pages 40–50, Albuquerque, New Mexico. Association for Computational Linguis- tics. 11 Stefan Sylvius Wagner, Maike Behrendt, Marc Ziegele, and Stefan Harmeling. 2025. The power of LLM- generated synthetic data for stance detection in online political discussions. In The Thirteenth International Conference on Learning Representations. Alex Wilf, Sihyun Lee, Paul Pu Liang, and Louis- Philippe Morency. 2024. Think twice: Perspective- taking improves large language models’ theory-of- mind capabilities. In Proceedings of the 62nd Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), pages 8292–8308, Bangkok, Thailand. Association for Computational Linguistics. Yufan Wu, Yinghui He, Yilin Jia, Rada Mihalcea, Yu- long Chen, and Naihao Deng. 2023. Hi-ToM: A benchmark for evaluating higher-order theory of mind reasoning in large language models. In Find- ings of the Association for Computational Linguis- tics: EMNLP 2023, pages 10691–10706, Singapore. Association for Computational Linguistics. Hainiu Xu, Runcong Zhao, Lixing Zhu, Jinhua Du, and Yulan He. 2024. OpenToM: A comprehensive bench- mark for evaluating theory-of-mind reasoning capa- bilities of large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8593–8623, Bangkok, Thailand. Association for Computational Linguistics. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-judge with MT-bench and Chatbot Arena. In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, Red Hook, NY, USA. Curran Associates Inc. Terry Yue Zhuo, Vu Minh Chien, Jenny Chim, Han Hu, Wenhao Yu, Ratnadira Widyasari, Imam Nur Bani Yusuf, Haolan Zhan, Junda He, Indraneil Paul, Simon Brunner, Chen GONG, James Hoang, Armel Randy Zebaze, Xiaoheng Hong, Wen-Ding Li, Jean Kad- dour, Ming Xu, Zhihan Zhang, and 14 others. 2025. BigCodeBench: Benchmarking code generation with diverse function calls and complex instructions. In The Thirteenth International Conference on Learning Representations. A LLM Usage Statement Beyond the use within the experiments, the LLM was employed in this work solely for refining sen- tences and improving grammatical accuracy during the manuscript writing process. B Detailed Information on IoA-Suite B.1 Dataset Statistics Each instance in IoA-Suite originates from a seed tuple(d,k,x)comprising a domaind, a collab- orative task typek, and topical contentxdrawn from a real scholarly publication. The content field contains the title and abstract of the source paper, which grounds the synthesized participant profiles and dialogue in realistic subject matter without re- quiring access to the full text. Sources. The 1,620 seed tuples used to gener- ate IoA-Suite (1,200 training, 120 validation, 300 test) are drawn from four open-access repositories matched to the target domain: arXiv for ML re- search (cs.LG, cs.AI, cs.CL, stat.ML), software engineering (cs.SE), and business (q-fin); PubMed Central for medical; and OpenAlex together with DOI-indexed records for cross-domain and law. Distribution. Seeds are distributed uniformly across the 30 cells formed by the 6 domains and 5 task types: 40 per cell for training, and 4 per cell for validation, 10 per cell for the test split, matching the 1,200, 120, and 300 dialogue counts reported in Section 4.3. DimensionRangeLevel definitions Stage 1 — per misalignment point Subtlety1–33: Even a careful participant would not think to clarify; surface agreement feels completely nat- ural. 2: A careful participant might notice, but plausibly would not. 1: Any competent profes- sional would explicitly clarify this. Specificity1–22: Both interpretations are concrete and distinct. 1: Interpretations are too abstract or overlapping. Believability1–22: Both interpretations are natural given each member’s role. 1: At least one interpretation feels unlikely for that role. Stage 2 — per dialogue Naturalness1–33: Could pass as a real meeting transcript. 2: Mostly natural with minor awkwardness.1: Clearly LLM-generated; robotic phrasing or unre- alistic dynamics. Implicitness1–3 3: Surface language is naturally ambiguous; the other speaker would easily assume alignment. 2: Mostly maintains false consensus, but a careful listener might pause to clarify. 1: One or more speakers explicitly clarify their interpretation; mis- alignment becomes visible. Coverage1–22: Every misalignment point’s surface form or close equivalent appears in the dialogue. 1: One or more misalignment points are never mentioned. Profile Cons.1–33: Both speakers behave and reason in ways matching their profiles. 2: Minor inconsistencies with roles or stated priorities. 1: Speakers act out of character or contradict their profiles. Table 7: Full scoring rubric used by both the LLM judge and human annotators. B.2 Generation Pipeline Configuration All four pipeline components (G participants ,G dialogue , J participants ,J dialogue ) use GPT-5.4, as stated in Sec- tion 4.3. Generators sample at temperature0.7; judges run at 0.3 to reduce score variance. 12 Quality thresholdsθ 1 andθ 2 .θ 1 requires that every planted misalignment point achieves the max- imum score on all three Stage-1 dimensions as- sessed byJ participants (subtlety= 3, specificity= 2, believability= 2).θ 2 requires that the dialogue achieves the maximum score on all four Stage-2 dimensions assessed byJ dialogue (naturalness= 3, implicitness= 3, coverage= 2, profile consistency = 3). B.3 Quality Assurance Scoring rubric.Both the LLM judge (J) and the human annotators use the same seven-dimension rubric, reproduced in Table 7. Stage-1 dimensions assess each planted misalignment point; Stage-2 dimensions assess the dialogue as a whole. LLM re-scoring. To mitigate self-enhancement bias from using the same model family for both generation and quality control, we re-score the 300- dialogue test set with Gemini-3.1-Pro Preview as a cross-family judge. Per-dimension scores are reported in Table 8 and all seven dimensions reach at or near the maximum. ComponentDimensionRangeScore Misalignment Points Subtlety1–32.65 Specificity1–22.00 Believability1–22.00 Dialogue Naturalness1–32.63 Implicitness1–32.66 Coverage1–22.00 Profile Consistency1–32.99 Table 8: Quality assessment of the IoA-Suite test set, evaluated by Gemini-3.1-Pro Preview. Human annotation protocol.Two PhD students independently annotated 60 dialogues sampled from the test set (20% of the full test split), with two dialogues drawn from each of the 30 cells in the 6-domain×5-task-type grid. Across these 60 dialogues the annotators scored 240 misalignment points and 60 dialogue-level instances. Prior to annotation, both annotators completed a joint cal- ibration session: they reviewed the rubric, scored five held-out dialogues together, and discussed any disagreements until reaching consensus on rubric interpretation. Each annotator scored all seven di- mensions for each dialogue; scores from the two annotators were averaged to obtain the final human rating. ComponentDimensionnHumanGemini ±0.5 ±1 MA Point Subtlety2402.662.6783.8%99.6% Specificity2401.982.00100%100% Believability2402.002.00100%100% Dialogue Naturalness602.392.6078.3%100% Implicitness602.452.7773.3%100% Coverage602.002.00100%100% Profile Consistency603.003.00100%100% Table 9: Cross-evaluator comparison between two hu- man annotators (averaged) and Gemini-3.1-Pro Preview on a 20% stratified sample of the IoA-Suite test set (60 dialogues, 240 misalignment points).±0.5/±1report the fraction of items on which the two evaluations agree within the corresponding tolerance. Human vs. LLM comparison. Table 9 reports, for each of the seven dimensions, the mean hu- man and Gemini-3.1-Pro Preview scores together with the fraction of items on which the two evalu- ations differ by at most0.5and at most1.0point. Agreement is high across the board: every dimen- sion achieves within-1.0-point agreement of at least 99.6%, and four dimensions (specificity, believabil- ity, coverage, profile consistency) show within-0.5 agreement on every item. On MA-level dimensions, human and Gemini-3.1-Pro Preview means differ by at most0.02points. The largest gaps appear on dialogue-level naturalness (∆ = 0.21) and implicit- ness (∆ = 0.32), where humans rate slightly lower than Gemini-3.1-Pro Preview, consistent with the known tendency of LLM judges to assign higher scores to LLM-generated text (Zheng et al., 2023). Coverage and profile consistency are at ceiling for both, indicating that these two properties are reli- ably achieved by the generation pipeline. B.4 Generation Prompts Stage-1:Profile Generation (G participants ). The user prompt is a single line:Seed topic: seed_topic. You are a scenario designer for illusion of misalignment research. Given a seed topic, a professional domain, and a meeting task type, generate a realistic meeting scenario with two participant profiles, planted misalignment points, and a conversation outline. <task> Create: 1. A concrete meeting scenario grounded in the given domain and task type. 2. Two meeting participants with distinct roles and hidden assumptions. 3. 2-4 subtle cognitive misalignment points between them. 4. A conversation outline (script) that specifies how the meeting should unfold turn by turn, ensuring ALL misalignment points are naturally touched upon during the conversation. The misalignment must be IMPLICIT -- both members speak cooperatively and 13 believe they are on the same page, but they hold genuinely different underlying beliefs or assumptions. </task> <domain_context> Professional domain: domain Meeting task type: task_type Task type description: task_type_description </domain_context> <profile_constraints> Each member has: - public info: name, role, expertise, background - private_state: internal goals, priorities, and concerns ( never explicitly stated in meetings) - hidden_assumptions: 3-5 beliefs taken for granted without stating them </profile_constraints> <misalignment_constraints> 1. SUBTLETY IS THE MOST IMPORTANT CRITERION. The misalignment must survive a normal, competent meeting conversation -- meaning even a careful participant would NOT think to ask for clarification because the surface agreement feels completely natural and sufficient. 2. Avoid "first-order design decisions" -- things any competent team would explicitly discuss during a meeting (e.g., which algorithm to use, which dataset to evaluate on, what the primary metric is, which treatment to prescribe). These are too obvious to go undetected. 3. Target "second-order assumptions" -- details that BOTH parties consider so obvious they would never think to mention, or parameters that feel already settled by the surface-level agreement. 4. For each point, explain WHY this misalignment goes undetected. 5. Avoid PRIORITY-ONLY misalignment (e.g., "A thinks X is more important than Y"). Prefer disagreements about WHAT something IS or HOW something WORKS. 6. Make each point CONCRETE and SPECIFIC to the scenario -- use specific names, numbers, configurations, versions, etc. relevant to the domain. </misalignment_constraints> <conversation_outline_rules> Generate a conversation outline that serves as a high-level script for the dialogue. This outline should: 1. Specify a total number of turns. Each turn has a one- sentence description of what the speaker should talk about. 2. Ensure EVERY misalignment point is covered: for each MA point, there must be at least one turn where the relevant surface phrase is used naturally. Mark which MA point(s) each turn touches by listing their IDs. 3. The conversation should flow naturally. 4. Each turn description should be brief (one sentence) and indicate what NEW content the speaker introduces. Avoid turns that just repeat agreement. 5. The outline is a GUIDE, not a script. It tells the dialogue generator what topics to cover and in what order, but the actual wording is generated later. </conversation_outline_rules> <output_format> "meeting_scenario": "A concrete meeting scenario. Include: context, agenda, and expected outcome.", "members": [ "id": "M1", "name": "...", "role": "...", "expertise": "...", "background": "...", "private_state": "...", "hidden_assumptions": ["...", "...", "..."] , "id": "M2", ... ], "misalignment_points": [ "id": "MA1", "topic": "Brief label", "surface_form": "The expression or situation where both appear aligned", "member_a_understanding": "What M1 actually believes or means", "member_b_understanding": "What M2 actually believes or means", "why_undetected": "Why this divergence does not surface " ], "conversation_outline": [ "turn": 1, "speaker": "M1", "description": "One sentence about what this speaker should say", "turn": 2, "speaker": "M2", "description": "One sentence about what this speaker should say" ] </output_format> Stage-2: Dialogue Generation (G dialogue ). The user prompt is: “Generate the complete dialogue now. Follow the outline, ensure every MA point is covered, and keep all specific interpretations in inner_thought only – never in utterances.” You are an expert dialogue author for illusion of misalignment research. Your task is to write a COMPLETE meeting dialogue between two participants who hold hidden cognitive misalignments -- they believe they agree, but they actually interpret key terms and plans differently. You are an OMNISCIENT AUTHOR: you know both characters' profiles, hidden assumptions, and misalignment points. But each CHARACTER does not know the other's private state. You must faithfully role-play both, ensuring that neither character reveals information they would not naturally know or say. <members> members_block </members> <misalignment_points> ma_block </misalignment_points> <meeting_scenario> meeting_scenario </meeting_scenario> <conversation_outline> outline_block </conversation_outline> <quality_requirements> Your dialogue will be evaluated on the following criteria. You MUST optimize for ALL of them simultaneously. 14 1. IMPLICITNESS (Most Critical -- this is the entire point of the task): - In UTTERANCES: speakers must ONLY use the shared surface phrases for MA-related topics. They must NEVER spell out, clarify, or elaborate on their specific interpretation. To them, their understanding is obvious and goes without saying. - In INNER THOUGHTS: speakers CAN and SHOULD think in detail about what they specifically mean, plan, or assume. - TEST: For each MA point, ask: "If I were the OTHER speaker hearing this utterance, would I naturally assume they mean the same thing I do?" If yes -> good. If the speaker's words are so specific that misunderstanding becomes unlikely -> bad. 2. NATURALNESS: - The dialogue should read like a real meeting transcript, not a scripted performance. - Avoid mechanical repetition of surface phrases. Speakers may use synonyms, paraphrase naturally, or refer back indirectly. - Each speaker should have a distinct voice matching their role and background. - Include natural meeting dynamics: building on each other 's points, brief acknowledgments, transitioning between topics. 3. MA COVERAGE: - EVERY misalignment point must be touched in the dialogue. 4. PROFILE CONSISTENCY: - Each speaker's utterances and inner thoughts must be consistent with their stated role, expertise, background, and private_state. - A senior clinician should sound different from a junior nurse. A PM should have different concerns than an engineer. - Inner thoughts should reflect each character's private priorities and concerns as stated in their profile. 5. CONVERSATION PROGRESSION: - Each turn must advance the conversation with NEW content. No filler turns that merely repeat agreement. - Follow the conversation outline for topic ordering and MA coverage, but write natural transitions -- do not mechanically follow it. </quality_requirements> <output_format> Generate a JSON object with the complete dialogue: "turns": [ "turn_idx": 0, "speaker_id": "M1", "speaker_name": "Name of M1", "inner_thought": "Character's private monologue", "utterance": "What the character says out loud." , "turn_idx": 1, "speaker_id": "M2", "speaker_name": " Name of M2", "inner_thought": "...", "utterance": "..." ] </output_format> Stage-1 Quality Judge (J participants ). The user prompt is dynamically formatted from the seed con- tent, domain, task type, meeting scenario, member profiles, and MA points of the synthesized case. You are an expert evaluator assessing the quality of synthetically generated meeting member profiles and misalignment points for illusion of misalignment research. You will receive: - Seed content (the original topic) - Domain and task type - Meeting scenario description - Two member profiles (with roles, expertise, private states, hidden assumptions) - 2-4 misalignment (MA) points (with surface_form, both members' understandings) Your task: Score each MA point AND provide actionable feedback for improvement. <ma_criteria> For EACH misalignment point, score: 1. subtlety (1-3): 3 = Subtle: Even a careful participant would NOT think to clarify this. The surface agreement feels completely natural. 2 = Moderate: A careful participant MIGHT notice, but it is plausible they would not. The gap is partially hidden. 1 = Obvious: Any competent professional would explicitly discuss or clarify this. It is a first-order design decision, not a second-order assumption. 2. specificity (1-2): 2 = Specific: The two interpretations are concrete and distinct. 1 = Vague: The interpretations are too abstract or overlapping. 3. believability (1-2): 2 = Believable: Both interpretations are natural given each member's role and background. 1 = Forced: At least one interpretation feels unlikely for that person's role. 4. feedback: A specific, actionable suggestion for how to improve this MA point. If the score is already maximum, say "No changes needed." </ma_criteria> <output_format> Output ONLY valid JSON: "ma_scores": "MA1": "subtlety": 1-3, "specificity": 1-2, "believability": 1-2, "feedback": "Specific improvement suggestion" , "MA2": ... , "overall_feedback": "High-level feedback on the profiles and MA points as a whole. Focus on the most impactful changes. If everything is excellent, say so." </output_format> Stage-2 Quality Judge (J dialogue ). The user prompt is dynamically formatted from the full synthesized case, including member profiles, MA points, and the complete dialogue transcript with per-turn inner thoughts. You are an expert evaluator assessing the quality of a synthetically generated 15 meeting dialogue for illusion of misalignment research. You will receive a complete synthesized case containing: - Seed content, domain, task type - Meeting scenario, member profiles, misalignment points - A meeting dialogue transcript (with utterances and inner thoughts) Your task: Score the dialogue quality AND provide actionable feedback for improvement. <dialogue_criteria> 1. naturalness (1-3): 3 = Natural: Could pass as a real meeting transcript. Authentic language and flow. 2 = Acceptable: Mostly natural with minor awkwardness or overly structured pacing. 1 = Artificial: Clearly LLM-generated. Robotic phrasing or unrealistic dynamics. 2. implicitness (1-3) -- False Consensus Strength: 3 = Strong false consensus: The surface-level language is naturally ambiguous. The other speaker would easily misinterpret what was said and believe they are aligned. Neither speaker has any reason to suspect a gap. 2 = Moderate: The conversation mostly maintains false consensus, but there are moments where a careful listener might pause and ask for clarification. 1 = Weak / Broken: One or more speakers explicitly clarify their specific interpretation, making it hard for the other party to misunderstand. The misalignment becomes visible or resolved during the conversation. IMPORTANT: The question is NOT whether a third-party reader can detect the gap, but whether the OTHER SPEAKER in the conversation would naturally assume agreement. Ask: "If I were the other speaker hearing this, would I assume they mean the same thing I do?" 3. coverage (1-2): 2 = Full: Every MA point's surface_form or close equivalent appears in the dialogue. 1 = Partial: One or more MA points are never mentioned in the conversation. 4. profile_consistency (1-3): 3 = Consistent: Both speakers behave and reason in ways matching their profiles. 2 = Mostly consistent: Minor inconsistencies with roles or stated priorities. 1 = Inconsistent: Speakers act out of character or contradict their profiles. 5. feedback: Specific, actionable suggestions for improving the dialogue. Reference specific turns or utterances where possible. If the dialogue is excellent, say "No changes needed." </dialogue_criteria> <output_format> Output ONLY valid JSON: "dialogue_scores": "naturalness": 1-3, "implicitness": 1-3, "coverage": 1-2, "profile_consistency": 1-3, "feedback": "Specific improvement suggestions referencing turns/utterances" </output_format> C Evaluation Protocol Details C.1 Simulator Stability Analysis To verify that a single simulator rollout (k = 1) produces reliable answers, we randomly sample 30 cases from the test set evaluated with the GPT-5.4 agent and re-run the simulatorSwithk = 5inde- pendent rollouts per question per participant.P 1 andP 2 denote the two participants in each dyadic dialogue, role-played independently by S. Intra-run consistencyValue P 1 97.7% P 2 96.8% Table 10: Simulator intra-run consistency overk = 5 rollouts on 30 sampled cases (173 questions). Per question, consistency is the fraction of the 5 rollouts that match the modal answer, averaged across questions. The high consistency justifies adoptingk = 1 throughout the main evaluation. Conditioned on the full private context ofp i (i.e.,r i ,g i ,a i , and h i (τ )) together with the shared transcript, the an- swer to any single MCQ is largely pinned down by that context, so role-playing reduces to a near- deterministic retrieval rather than an open-ended generation, leaving little room for stochastic varia- tion across rollouts. C.2 Models and Hyperparameters Table 11 lists all models used in the evaluation pipeline with their API identifiers, providers, and inference hyperparameters. All prober models are queried uniformly at T = 0.7 with thinking mode. C.3 Information-Access Ablation Configuration The four ablation settings vary exactly two slots in the prober user prompt (Section C.4): member_profilesandtranscript. Table 12 summarises which content variant fills each slot. Thetwocontentvariantsfor member_profiles are shown below. Public profile (baseline) - [Name]: [Role] Expertise: [expertise] Background: [background] 16 Name in paperAPI identifierProviderRoleT k Closed-source probers GPT-5.4 (OpenAI, 2026) gpt-5.4OpenAIProber0.71 Gemini-3.1-Pro Preview (Google, 2026) gemini-3.1-pro-previewGoogleProber0.71 Gemini-3-Flash (Google, 2026) gemini-3-flash-preview-thinkingGoogleProber0.71 GPT-5 mini (OpenAI, 2026) gpt-5-mini-2025-08-07OpenAIProber0.71 Open-source probers DeepSeek-V3.2 (DeepSeek-AI, 2025) deepseek-v3.2DeepSeekProber0.71 GLM-5.1 (GLM, 2026) glm-5.1ZhipuProber0.71 Qwen3.5-397B-A17B (Team, 2026b) qwen3.5-397b-a17bAlibabaProber0.71 Kimi-K2.5 (Team, 2026a) kimi-k2.5MoonshotProber0.71 Qwen3-8B (Team, 2025) Qwen3-8BAlibabaProber0.71 Fixed pipeline components Gemini-3-Flash gemini-3-flash-preview-thinkingGoogleSimulator S0.21 GPT-5.4 gpt-5.4OpenAIJudgeV0.21 Table 11: Models and hyperparameters used in the evaluation pipeline.T: sampling temperature.k: rollouts per participant per question. The simulator and judge are fixed across all prober evaluations. Settingmember_profilestranscript r i onlypublicutterances r i + h k publicwith inner thoughts r i + g i ,a i fullutterances r i + g i ,a i ,h k fullwith inner thoughts Table 12: Prompt slot assignment for each information- access setting. Public: name, role, expertise, back- ground only. Full: additionally includes private stateg i and tacit assumptionsa i . With inner thoughts: each ut- terance is prefixed by the speaker’s turn-level cognition h k . Full profile (with private state and tacit assumptions) - [Name]: [Role] Expertise: [expertise] Background: [background] Private State: [private_state] Hidden Assumptions: - [assumption 1] - [assumption 2] ... The two content variants fortranscriptare shown below. Utterances only (baseline) [Speaker A]: utterance text [Speaker B]: utterance text ... Transcript with inner thoughts (with turn-level cogni- tion) [Speaker A] inner_thought: private thought at this turn utterance: utterance text [Speaker B] inner_thought: private thought at this turn utterance: utterance text ... C.4 Evaluation Prompts The prober, simulator, and validity filter each use a fixed system prompt plus a dynamically assem- bled user prompt. Because the four information- access settings differ only in the content of two user-prompt slots (Section C.3), the system prompt and template are shown once. Prober. Theuserpromptfills member_profilesandtranscriptac- cording to the active ablation setting (Table 12). Prober system prompt <role> You are Illusion of Alignment Detector, an expert agent that detects hidden cognitive misalignment between meeting participants by generating diagnostic multiple-choice questions. </role> <task> Analyze the meeting transcript and member profiles. Generate questions that, when answered independently by each member, would reveal hidden disagreements or divergent assumptions NOT explicitly surfaced during the meeting. </task> <scoring_system> Your output is scored per-question. Your goal is to MAXIMIZE your total score. A question SCORES POINTS only when ALL of the following are true: 1. The two members give DIFFERENT answers (divergence exists) 2. The divergence reflects a REAL cognitive gap, not noise or poor option design 3. The gap was IMPLICIT -- not openly debated during the meeting 4. The question is not REDUNDANT with another question you generated A question LOSES POINTS when ANY of the following is true: 17 - Members give the same answer (no divergence detected) - The divergence is meaningless (ambiguous wording, noise) - The misalignment was explicitly discussed in the meeting - It duplicates another question you already generated There is no penalty for generating many questions -- but every bad question actively hurts your score. Generate a question for every distinct implicit gap you can confidently identify, but do not guess. </scoring_system> <principles> 1. PROBE THE IMPLICIT, NOT THE EXPLICIT: - NEVER ask about topics where participants openly disagreed or debated. - Target areas of APPARENT AGREEMENT -- moments where both participants used similar language, nodded along, or moved forward without objection, but might actually hold different interpretations of what was agreed upon. 2. DETECT FALSE CONSENSUS: - Focus on WHAT people mean, not WHETHER they agree. If they appear to agree, ask a question that tests whether their agreement is genuine or superficial. - Pay special attention to moments where one member's statement could be interpreted multiple ways, and the other member responded affirmatively without clarifying which interpretation they hold. 3. ROLE-AWARE: Use role and expertise differences to anticipate divergent mental models. Different roles often use the same term with very different operational definitions. 4. ONE QUESTION, ONE GAP: Each question should target exactly ONE potential cognitive divergence. Do not bundle multiple issues into a single question. </principles> <question_design> - 2-4 options per question. Each option must represent a genuinely different belief or interpretation, not just different phrasings of the same idea. - All options must be plausible -- no obviously wrong decoys. - Neutral framing: no option should appear "more correct" than others. - Questions should be phrased as concrete operational decisions or interpretations, not abstract opinion polls. GOOD: "When you agreed to'finalize the deliverable', what specific output did you expect to produce by the deadline?" BAD: "How important is quality to you?" </question_design> <output_format> "chain_of_thought": "Step-by-step reasoning about potential implicit misalignment...", "questions": [ "misalignment_signal": "What implicit cognitive divergence this question probes", "stem": "The question text", "options": "A": "...", "B": "...", "C": "...", "D": "..." ] Output ONLY valid JSON. </output_format> Prober user prompt template <member_profiles> member_profiles </member_profiles> <meeting_content> transcript </meeting_content> Simulator. The user prompt instructs the mem- ber to answer all questions based on their own un- derstanding of the meeting, formatted as a list of question stems and options. The system prompt is shown below. Simulator system prompt <role> You are name. You just attended a meeting and are now answering follow-up questions about what was discussed and decided. </role> <your_profile> Name: name Role: role Expertise: expertise Background: background Private State: private_state Hidden Assumptions: assumptions_block </your_profile> <meeting_record> dialogue_view </meeting_record> <instructions> - Answer based on YOUR understanding, beliefs, and assumptions. - Do NOT guess what the other person thinks -- answer from your own perspective only. - For each question, choose exactly ONE option letter. - Provide brief reasoning for each answer. </instructions> <output_format> "answers": ["question_id": 0, "reasoning": "...", "choice": "A", ...] Output ONLY valid JSON. </output_format> Validity filter. The filter receives ground truth misalignment points, full member profiles (pri- vate state and hidden assumptions included), the dialogue transcript (utterances only), and the diverging questions with both members’ an- swers. For each diverging question the filter deter- mines: is_meaningful (real cognitive gap vs. noise), is_implicit (gap was hidden in the transcript), re- dundant_with (duplicates an earlier question), and hits_ground_truth (maps to a planted misalignment point). Validity filter system prompt <role> You are an expert evaluator for an illusion of misalignment detection system. </role> <task> 18 You will receive ground truth misalignment points, both members' full profiles, the meeting transcript, and a set of diagnostic questions where the two members gave DIFFERENT answers. For each diverging question, evaluate it on the following criteria. Be strict -- the purpose is to separate genuinely valuable questions from noise. </task> <criteria> For each question, determine: 1. is_meaningful (true/false): Does the answer divergence reflect a REAL cognitive gap between the members? - true: The members genuinely hold different beliefs, interpretations, or assumptions, and their different answers correctly reflect this gap. Note that the ground truth misalignment points are NOT exhaustive -- if a question reveals a real cognitive divergence not listed in the ground truth , it is still meaningful. - false: The divergence is caused by ambiguous option wording, random noise, trivial differences, or poor question design -- not a real cognitive gap. 2. is_implicit (true/false): Was this cognitive gap HIDDEN during the meeting? - true: Both members appeared to agree on the surface. Neither openly challenged the other on this specific point. The gap was invisible in the transcript. - false: The members visibly debated, pushed back, or expressed different views on this exact topic during the meeting. A simple transcript reader could spot this conflict. 3. redundant_with (null or question_id): Does this question probe the SAME underlying gap as another question in the list? If so, specify the earlier question' s id. Only the first question targeting a given gap is non- redundant. 4. hits_ground_truth (null or MA-id): Does this question's divergence map to one of the provided ground truth misalignment points? Check whether the members' different answers correspond to the two sides of a specific ground truth point. A meaningful question that does not match any ground truth point may still be a novel discovery. </criteria> <output_format> "evaluations": [ "question_id": 0, "is_meaningful": true, "is_implicit": true, "redundant_with": null, "hits_ground_truth": "MA1", "reasoning": "Brief explanation of your judgment" ] Output ONLY valid JSON. </output_format> D IoA-Prober-8B Training and Evaluation D.1 Training Recipe Details SFT warm-start.The warm-start uses GPT-5.4- generated questions on 300 dialogues from the training split, with Category-C questions removed. Table 13 lists the hyperparameters. HyperparameterValue Base modelQwen3-8B Training dialogues300 Epochs2 Learning rate1× 10 −5 Effective batch size8 Max sequence length8,192 tokens Precisionbfloat16 Table 13: SFT warm-start hyperparameters. GRPO.GRPO is initialized from the SFT check- point and runs on the full 1,200-dialogue training split for two epochs (100 gradient steps at batch size 24). At each step the model generates a group ofG = 5candidate question sets per dialogue; each set is scored end-to-end through the simula- tor and judge to obtain an F1 reward, and group- normalized advantages are used to update the pol- icy. Table 14 lists the GRPO hyperparameters. HyperparameterValue FrameworkVeRL Group size G5 Training dialogues1,200 Validation dialogues120 Total steps100 Total epochs2 Train batch size24 Learning rate1× 10 −6 Max prompt length4,096 tokens Max response length4,096 tokens Hardware8× A800 80 GB Table 14: GRPO hyperparameters. Reward function.The reward for each candidate question set is the F1 score (Section 4.2) computed by running the full evaluation pipeline (simulator thenV) on the training dialogue. To control cost, the simulator uses gemini-3-flash (no thinking) at T =0.2withk=1rollout; the judge uses gpt-5-mini at T =0.2. No format reward is applied. D.2 Full Comparison Table Table 15 extends Table 3 to all nine baselines eval- uated in Table 1. 19 ModelPRF1K |B| Closed-source GPT-5.444.757.749.55.80.36 Gemini-3.1-Pro Preview46.540.443.03.50.06 Gemini-3-Flash43.940.441.93.80.10 GPT-5 mini34.947.840.05.80.26 Open-source DeepSeek-V3.242.843.642.94.20.13 GLM-5.141.338.639.34.10.34 Qwen3.5-397B-A17B46.334.638.13.00.05 Kimi-K2.535.539.837.14.80.23 Qwen3-8B37.523.328.02.60.04 Trained (ours) IoA-Prober-8B50.553.951.84.50.12 Table 15: Full results on IoA-Suite. Category-B analysis.|B|reveals what F1 alone cannot. The plantedM (τ )covers the bulk of sur- facable divergences, leaving|B| < 0.15for five of the nine baselines. The residual is not noise to be eliminated: natural-language dialogue admits emergent misalignments beyond any finite seed set, which is why we route divergences throughV rather than scoring directly againstM (τ ). Read in this light,|B|becomes a second axis of evalua- tion: high-|B|models such as GPT-5.4 (0.36) probe more broadly than the planted set anticipates. D.3 Statistical Tests on Main Results We assess statistical significance using paired boot- strap resampling (Koehn, 2004). For each test, we resample 10,000 times with replacement over the N = 300test dialogues (seed 20260516), com- puting per-dialogue F1 at each resample for both models, and report a two-sidedp-value based on the empirical distribution of the difference. IoA-Prober-8B (F1 = 0.518) does not signifi- cantly outperform GPT-5.4 (0.495;∆ = 0.023, p = 0.14), reflecting the limited statistical power to detect sub-0.03 differences atN = 300. All eight remaining baselines have∆ ≥ 0.088; con- sistent with the Gemini-3.1-Pro Preview anchor (∆ = 0.088, p < 0.001), each yields p < 0.001. D.4 Real-User Study Protocol Participants and meetings.We recruited 43 par- ticipants across 18 real working meetings, ranging from 9 to 90 minutes (mean 35), spanning ML research, medical AI, software engineering, and related collaborative research discussions. All par- ticipants were active collaborators who had worked together on the discussed task prior to the study. Participation was voluntary; participants were in- formed of recording and transcription before con- senting. Meeting transcripts were anonymized by replacing names with pseudonyms prior to any model invocation. Transcript collection and question generation. Each meeting was recorded and manually tran- scribed. The anonymized transcript was submit- ted to three detectors, Qwen3-8B, GPT-5.4, and IoA-Prober-8B, each generating one question set in parallel. Member profiles were set none; mod- els operated on the transcript alone, matching the standard evaluation condition. The three sets were assigned randomly to labels A, B, and C (shuf- fled independently per session); the label-to-model mapping was kept hidden from participants until after all ratings were submitted. Figure 3: Answering interface: each participant selects one option per question across the three blinded sets. Answering phase. Each participant received a unique session link to answer all three question sets 20 independently. The three sets were displayed as separate tabbed views; within each set, questions appeared in the order generated by the model. Par- ticipants did not see the answers from others during this phase. Figure 3 shows the answering interface. Figure 4: Validity-labeling and subjective evaluation interface. The upper section shows each diverging ques- tion with the five validity categories (one selection re- quired per question); the lower section collects per-set Likert ratings on misalignment discovery and willing- ness to use, plus a forced-choice best-set vote. Validity labeling phase. After all participants in a session completed answering, the aggregated results became available. For each session, partic- ipants were shown, per question set, which ques- tions had elicited diverging answers across partici- pants. For every diverging question, each partici- pant independently labeled the validity using one of five categories: • Valid: a genuine hidden disagreement the meet- ing did not resolve; • Invalid, no meaningful distinction: the answer options do not actually distinguish different positions; •Invalid, already discussed: the disagreement was raised and resolved in the meeting; •Invalid, redundant: the question reveals the same disagreement as an earlier question in the same set; •Invalid, other: any other reason the question is not useful. A diverging question is counted as valid in the reported Validity Rate only when at least one par- ticipant labels it valid. Figure 4 shows the validity- labeling interface. Post-rating survey.After completing validity la- beling, each participant independently filled in a three-item questionnaire: 1.Misalignment discovery (Likert 1 to 5, per set): “Did this question set help you discover hidden cognitive disagreements?” (1 = not at all; 5 = very much) 2.Best-set choice (forced choice, A/B/C): “Over- all, which set had the highest quality and most effectively revealed hidden disagreements?” Model identities were revealed to participants after submission. 3.Willingness to use (Likert 1 to 5, per set): “If a tool at this quality level were integrated into meeting software, how willing would you be to use it?” (1 = not at all; 5 = very willing) An optional free-text field collected additional com- ments on question design or platform usability. The Best-set votes reported in Table 4 aggregate across all 43 participants and 18 sessions. Statistical tests. All significance tests in Sec- tion 6.2 use paired bootstrap resampling follow- ing the same protocol as Appendix D.3: 10,000 resamples with replacement, seed 20260516, two- sidedp-value from the empirical distribution of the difference. For the validated-disagreement compar- ison, the resampling unit is the session: each of the 18 sessions contributes a paired triple of counts (Qwen3-8B, GPT-5.4, IoA-Prober-8B), where the count is the number of diverging questions in that session for which at least one participant la- beled Valid. IoA-Prober-8B exceeds GPT-5.4 by ∆ = +1.00per session (95% CI[0.28, 1.67], p = 0.005) and Qwen3-8B by∆ = +1.22(95% CI[0.61, 1.78],p < 0.001). For the blinded forced- choice survey, the resampling unit is again the ses- sion: each session contributes a+1/0/− 1indi- cator for the pairwise comparison, set to+1when IoA-Prober-8B receives strictly more best-set votes 21 than the comparator,−1when fewer, and0on a tie. IoA-Prober-8B wins 12 of 18 sessions versus GPT-5.4 (2 losses, 4 ties;p = 0.002) and 12 versus Qwen3-8B (3 losses, 3 ties; p = 0.008). D.5 Ablation Study Details Table 5 compares four single-factor variants against IoA-Prober-8B. All variants are initialized from Qwen3-8B and use the same training split, F1 re- ward, and hyperparameters as the full recipe (Ap- pendix D.1); only the single factor under study changes. w/o RL. The GRPO stage is omitted entirely; SFT runs on the full 1,200-dialogue training split. Two sub-variants are reported: one applies Category-C filtering to the SFT data (retaining only Categories A and B), and the other uses all cate- gories. w/o SFT warm-start. GRPO is applied directly to the base Qwen3-8B checkpoint without any su- pervised initialization. The 1,200-dialogue training set, F1 reward, and hyperparameters are identical to the full recipe. w/o filterC in warm-start.The SFT warm-start uses all GPT-5.4-generated questions on the 300- dialogue warm-start split, including Category-C questions (divergent but rejected by the validity filter). The GRPO stage then proceeds identically to the full recipe on the 1,200-dialogue set. D.6 Multi-Agent Collaboration Setup BigCodeBench-Hard.We use a symmetric two- discusser architecture. Two identical Qwen3-8B instances (DiscusserA, DiscusserB; temperature 0.7) alternate for up to four discussion rounds, each round ending with a parallel stop-vote (temperature 0.0) that terminates discussion when both agents signal readiness. One of the two discussers is then asked to consolidate the discussion log into the final Python solution (temperature lowered to 0.2). Cor- rectness is evaluated by the official BigCodeBench sandboxed test harness (pass@1 over 148 tasks). HiddenBench. Each scenario assigns a distinct system prompt to every agent: each prompt con- tains the shared scenario description plus one agent- private information block (shuffled deterministi- cally by task identifier hash to prevent ordering artifacts). Agents discuss for eight rounds in a round-robin order; each agent sees the messages of all preceding agents in the current round but maintains its own private system prompt and his- tory throughout. Final answers are extracted per agent and resolved by majority vote. We evaluate 65 scenarios (7 three-agent, 58 four-agent). Detector integration.After the initial discussion phase, detector-augmented configurations call the prober on the discussion transcript and indepen- dently ask allNagents to answer each generated question from their own perspective (temperature 0.2). If any question elicits divergent answers across agents, a re-discussion phase is triggered; otherwise the pipeline proceeds directly to code generation (BCB-Hard) or final voting (Hidden- Bench). The divergent question, together with each agent’s choice and one-sentence reasoning, is in- jected as a directive turn that requires agents to name the specific assumption behind their pick and directly address their partner’s differing interpre- tation before modifying any implementation deci- sion. Self-reflection. Instead of an external prober, each discusser is asked post-discussion to scan the log for hidden disagreements it may still hold with its partner and report them in a structured JSON response. If either agent flags a potential disagree- ment, a re-discussion round is triggered; otherwise agents proceed. Prompts. All agents share the same discussion system prompt; a coder prompt instructs the dis- cusser writing the final code to output a fenced Python block only. The self-reflection prompt asks each discusser to identify hidden disagree- ments and respond with structured JSON. For IoA- detector variants, an answerer prompt asks each agent to respond to a given MCQ from its own per- spective (JSON with reasoning and letter choice), and the re-discussion directive formats divergent answers into a focused elicitation turn. All prompts are reproduced below. Discusser system prompt You are one of two Python developers working together to solve a coding task. Discuss with your partner how to approach the problem so you both reach a shared understanding of the implementation. Be brief and concrete. When you and your partner agree on the approach, say so plainly so the team can move on to writing code. Do NOT write the code itself during discussion. Forced self-check system prompt 22 You are one of two developers who have just finished discussing a coding task. Scan the discussion log and identify any HIDDEN disagreements you may still have with your partner -- places where you said "yes" or moved on, but actually held a different mental model than what your partner appeared to assume. Reply with valid JSON only: "has_hidden_disagreement": true|false, "items": ["<short description>", ...] IoA answerer system prompt (MCQ response) You are role, one of two developers who just discussed a coding task. A diagnostic question is being asked to test your mental model. Answer based on YOUR own understanding of the task and discussion -- do not try to guess what your partner thinks. Reply with valid JSON only: "reasoning": "<2-3 sentences citing a step from the discussion or wording from the task>", "choice": "A|B|C|D" Re-discussion directive (injected on divergent MCQ answers) [FOCUSED ROUND -- DIAGNOSTIC SPLIT DETECTED] A diagnostic check probed your shared understanding with the question(s) below. You and your partner chose DIFFERENT answers, which suggests you may be holding different assumptions about the spec, the API contract, or how an edge case should be handled. In your next message you MUST, for EACH question below: 1. Re-state your own choice and name the SPECIFIC assumption (or quote from the task / API doc) that led you there. 2. Directly address your partner's different choice -- what is the smallest piece of evidence that should resolve which interpretation is correct? 3. If your partner's reasoning is actually correct, say so explicitly and state the implication for the implementation. Do NOT re-litigate parts of the discussion that were not flagged. E Case Study and Failure Modes E.1 Case Studies Case 1 — Insider-trading alert troubleshooting (business/troubleshooting_010) Scenario.Nina Patel (Surveillance Product Man- ager) and Leo Martinez (Senior Data Scientist) hold a 12-turn troubleshooting meeting at a regtech firm. Their anomaly-detection module has triggered ana- lyst complaints: cluster-based cases feel repetitive, overly broad, and hard to action. The meeting closes with four shared commitments: reduce “re- dundant output,” validate on “the last two weeks,” “refresh the graph,” and “handle the dense clusters.” Ground-truth misalignments (|M (τ )| = 4). Dialogue. Nina: Thanks for making time. We’ve got a real trust issue right now. Over the last two weeks, the detector has pushed a lot of cluster-based cases that analysts are calling repetitive, overly broad, and hard to action. We need to understand why that’s happening and how to handle the dense clusters before the next review cycle. Leo: I noticed the same pattern in the recent cycle. The insider activity clusters look unusually dense, and my first instinct is that we should look closely at how the graph is being generated, not only what shows up downstream. If we’re going to handle the dense clusters properly, we should confirm where that density is coming from. Nina: Right. One example from operations was an earnings-window situation where a CFO, a related trust, and a controlled entity all surfaced in ways that felt like the same story to the reviewer. That’s the kind of redundant output we need to reduce. Leo: Agreed, redundancy is definitely part of it. I want to check whether those redundant links are being created by repeated data points or structural artifacts. If that’s happening, it would explain why some of these clusters are expanding so much. Nina: Can we refresh the graph before next week’s analyst cycle? I’d really like to calm the queue down without changing the overall product behavior if we can avoid it. Leo: Yes, a graph refresh makes sense. I’d like to update the graph build so those shared episodes are represented more cleanly instead of balloon- ing the connectivity. That should help us handle the dense clusters in a more stable way. Nina: Okay. For validation, I want us grounded in the last two weeks, because that’s where the complaints came from. If we review that recent sample, we should be able to tell pretty quickly whether the update is moving things in the right direction. Leo: That works. I’l compare the refreshed graph on the latest complete two-week window and check whether the same density pattern is still there. If the structure looks healthier in that recent sample, we’l know the update is helping. Nina: And just to keep the immediate pain point front and center: if we can handle the dense clus- ters so analysts aren’t overwhelmed by what feels like the same earnings story repeating, that would take a lot of pressure off. Leo: Agreed. If we handle the dense clusters properly, the system should stop generating that much noise around the same event pattern. That would give us a cleaner result going forward. Nina: Great. Can you send me an updated graph sample and a short note on how much redundant output drops before the next review batch? Leo: Yes, I can do that. I’l send the refreshed graph results and summarize whether the recent density issue is reduced in the validation sample I use, along with the change in redundant output. 23 IDNina (M1)Leo (M2) MA1Analyst-queue duplicates (same story, multiple tickets)Graph-structural artifacts (repeated edges / entity- resolution noise) MA2 Last two calendar weeks by ingestion/processing times- tamp Latest complete two-week transaction-date window, ex- cluding late filings MA3Rerun existing pipeline with cleaned filters on the recent sample Modify build logic (edge representation, co-trading episodes) MA4Bundle clusters into one analyst case (downstream pack- aging) Fix edge weights / scoring so dense structures stop being flagged (upstream) Table 16: Planted misalignments for Case 1. IoA-Prober-8B output (P = 0.80,R = 1.00, F1 = 0.89). 1. [Cat A, GT MA4]Nina→ A Leo→ B When you referred to “dense clusters” in the recent prob- lem, what pattern did you most likely have in mind? A. A cluster that is too broad and repetitive for analysts to act on, even if the underlying graph structure looks normal B. A cluster that is unusually connected in the graph be- cause of repeated data points or structural overlap, even if analysts do not yet flag it as repetitive C. A cluster that is both too broad for analysts and unusu- ally connected in the graph 2. [Cat A, GT MA3]Nina→ A Leo→ B When you agreed to “refresh the graph” before next week’s analyst cycle, what did you most likely assume that change would do? A. Update only the latest data so the same product behavior stays consistent, while reducing redundant output within that window B. Adjust how shared episodes are represented in the graph structure itself, even if that changes how the product be- haves slightly C. Replace the entire graph build process end-to-end 3. [Cat A, GT MA2]Nina→ A Leo→ B When you agreed to validate the update on a “recent two- week window,” which sample did you most likely have in mind? A. The exact last two weeks of cases that analysts already flagged as problematic, so the update can be judged against that immediate pain point B. The latest complete two-week window in the system, regardless of whether it matches the most recent flagged cases C. A rolling two-week sample that best reflects current graph behavior 4. [Cat A, GT MA1]Nina→ A Leo→ B When you asked for a concrete number on how much “re- dundant output” dropped, what did you most likely want that metric to measure? A. Cases that analysts treat as repetitive or overly broad, even if the graph structure changed only slightly B. Graph artifacts like repeated links or expanded connec- tivity, even if analysts still see similar cases C. A combined measure of both analyst repetition and graph-level redundancy 5. [Cat C, GT MA4, redundant with Q4]Nina→A Leo→ B What outcome would most clearly show that the update “handled dense clusters properly” in the next review? A. Analysts feel less overwhelmed by what looks like the same earnings story repeating in the queue B. The graph no longer shows the same structural density pattern around shared episodes C. Both analyst workload and graph structure improve in a way that is visibly cleaner together GPT-5.4 output (P = 0.00, R = 0.00, F1 = 0.00). 1. [Cat C, not diverging]Nina→ C Leo→ C When you said “refresh the graph before next week’s ana- lyst cycle,” what did you assume would change? A. Rebuild using the same logic, but with the latest in- gested data B. Keep the same data, but change graph-construction logic and rebuild C. Both: update data inputs and also adjust graph- construction logic D. Neither; only adjust downstream clustering using the existing graph 2. [Cat C, GT MA1, not implicit]Nina→ A Leo→ C When you asked to reduce “redundant output,” what specif- ically should drop in the analyst queue? A. Fewer separate cases that describe essentially the same earnings-window story B. Smaller clusters (fewer nodes/edges) even if the number of cases stays similar C. Less repeated evidence inside a case (duplicate links/- transactions) D. Lower alert volume overall, even if some unique stories are also suppressed 3. [Cat C, not diverging]Nina→ B Leo→ B What would count as “changing overall product behavior” in a way we should avoid for next week? A. Any visible change in what analysts see, even if detec- tions remain correct B. A material change to detection coverage/recall C. Any change that would invalidate prior tuning/baselines D. Only changes that require analyst retraining 4. [Cat C, not diverging]Nina→ A Leo→ A What is the most likely source of the unusually dense insider-activity clusters? A. Entity resolution issues (trust/controlled entity/CFO linked too aggressively) B. Edge-definition artifacts (too many relationship types creating excess edges) C. A real market/filing pattern shift causing genuinely denser graphs D. Downstream clustering parameters over-merging com- ponents 5. [Cat C, not diverging]Nina→ A Leo→ A To prevent “shared episodes” from ballooning connectiv- ity, what representation change is implied? 24 A. Introduce an event-centric node that multiple entities attach to, reducing entity-to-entity edges B. Keep entity-to-entity edges but de-duplicate and canon- icalize repeated evidence C. Down-weight certain edge types during graph build D. Leave the graph as-is and only change how clusters are visualized 6. [Cat B, novel valid]Nina→ A Leo→ B In the two-week validation sample, what is the primary “pass/fail” indicator that the update worked? A. Analysts can quickly identify a single coherent story per case with fewer repetitive items B. Graph structural metrics improve (fewer high-degree hubs, lower average component density) C. Alert/case volume drops to a target band compared to the prior two weeks D. Known historical scenarios still appear with similar ranks/scores 7. [Cat C, not diverging]Nina→ A Leo→ A How should we choose the evaluation dataset for the graph refresh? A. Only the most recent two weeks to match the complaint period and move fast B. Recent two weeks plus one earlier baseline period C. A stratified sample across multiple quarters D. Use the last two weeks for speed, but spot-check histor- ical cases for regressions 8. [Cat B, novel valid]Nina→ A Leo→ B What should the “updated graph sample” you send con- tain to be most useful for ops review? A. A few end-to-end analyst cases showing before/after redundancy changes B. A graph export plus summary stats so technical stake- holders can inspect connectivity changes C. Screenshots/visualizations of representative clusters highlighting structural changes D. A metric-only report without case-level examples 9. [Cat C, GT MA4, not implicit]Nina→ A Leo→ C To “calm the queue down,” which lever do you assume we are primarily pulling for next week? A. Generate fewer cases by consolidating similar hits into a single case per story B. Keep case count similar but improve within-case group- ing so analysts can triage faster C. Raise thresholds so fewer clusters qualify D. Change prioritization so repetitive stories fall lower 10. [Cat C, GT MA1, not implicit]Nina→ B Leo→ C When you both said the system is generating “noise” around the same event pattern, what do you mean by “noise”? A. Alerts that are not suspicious (low precision / false pos- itives) B. Potentially valid signals, but duplicated across many cases (presentation redundancy) C. Signals driven by data-quality artifacts rather than real- world behavior D. Analyst time sinks even if the underlying suspicion is correct 11. [Cat C, not diverging]Nina→ A Leo→ A Where do you believe the primary fix should live to reduce dense clusters for next week? A. Graph build layer (identity resolution + edge construc- tion) B. Clustering layer (how we form components) while leav- ing the underlying graph unchanged C. Case assembly layer (how clusters are turned into cases) D. Analyst UI layer (filters/rollups) 12. [Cat B, novel valid]Nina→ A Leo→ B If reducing cluster density risks splitting a truly coordi- nated multi-entity situation into smaller clusters, what is the preferred bias? A. Prefer usability: split aggressively so analysts see smaller, clearer stories B. Prefer detection integrity: keep coordination visible even if clusters remain large C. Split only when evidence is near-duplicate; keep large clusters when evidence is diverse D. Defer: avoid structural changes until we can quantify the coordination-loss risk Analysis.IoA-Prober-8B generates one question per misalignment anchor, using the shared vocabu- lary term as the pivot of each stem. Every question elicits a clean split, four are scored Category A (all four ground-truth points covered), and only Q5 fails — it is redundant because it re-probes the MA4 procedural gap from a slightly different an- gle. The prober recovers R = 1.00 with K = 5 and P = 0.80. GPT-5.4 generates twelve questions but fails to score any Category A. Six questions do not elicit di- vergent answers at all (Q1, Q3, Q4, Q5, Q7, Q11); two hit ground-truth misalignments (MA1: Q2, Q10; MA4: Q9) but are judged not implicit be- cause Leo’s “structural artifacts” framing in turn 4 and the phrase “handle the dense clusters properly” in turn 6 already surface the gap in the transcript, making the disagreement partially visible. Three questions are novel valid discoveries (Category B), demonstrating that GPT-5.4 explores a wider con- cept space but fails to focus on the latent misalign- ments the conversation actually encodes. Case 2 — Asset-management brainstorm (business/brainstorming_003) Scenario. Nina Patel (Head of Quantitative Re- search) and Daniel Brooks (Head of Product Strat- egy) brainstorm investment concepts ahead of a committee pre-read. The meeting is a sequence of parallel affirmations: neither participant challenges the other’s framing, and the conversation closes with consensus on an “internal pilot,” the “same signals,” “clear economic gains,” and the “attention model” as the headline. All four phrases encode hidden divergences. Ground-truth misalignments (|M (τ )| = 4). Dialogue. 25 IDNina (M1)Daniel (M2) MA1Paper portfolio / shadow book using the existing monthly U.S. equity pipeline; no real capital Live-money seed account funded by the firm’s balance sheet; starts building an actual track record MA2 Exact 142-feature research library with current preprocess- ing fixed Same signal categories and intellectual premise, with se- lective substitutions for institutional packaging MA3Strong gross theoretical profitability in backtest (long- short spreads, utility gains) before implementation analysis Gains that survive realistic implementation at institutional scale (turnover drag, market impact, capacity) MA4Attention-augmented RNN with memory mechanism that outperformed others in Nina’s internal comparison Transformer-like architecture, the model family most allo- cators associate with “attention” and contemporary AI Table 17: Planted misalignments for Case 2. Nina: Thanks for making the time. What I’d like to do is leave with two or three credible ways to turn the recent deep-learning asset-pricing work into an internal pilot, especially given how well the attention model did on risk-premia prediction. Daniel: That makes sense. If we can make the pilot feel concrete and disciplined, I think the committee will engage with it pretty seriously. Nina: My bias would be to stay very close to the study — use the same signals and frame this as a disciplined extension of the existing research pipeline, not a new data story. Daniel: I agree with that. Keeping it grounded in the same signals should help us preserve conti- nuity with the study while also making the story easier to carry into committee. Nina: One angle, then, is to center it on better risk-premia measurement itself. The strongest hook there is that the research already shows clear economic gains from the forecasts. Daniel: I like that. I think it lands best if we can say those economic gains also look credible in an institutional context, so it does not read as purely academic. Nina: A different way in is the explainability angle. The attention model gives us a way to show how different signal clusters matter across market regimes without giving up the finance intuition. Daniel: Yes, and I think the attention model is a strong headline in its own right. People already recognize that label, and it helps signal that this is current rather than legacy quant with new pack- aging. Nina: A third option is to lean into distribution shift — basically, the pilot shows the model adapt- ing to time-varying return distributions while still relying on the same signals. Daniel: I could see that resonating, especially if we position the internal pilot as a disciplined incubation effort rather than just another backtest summary. Nina: If I had to narrow it today, I’d probably favor a U.S. equities monthly pilot built from the current comparative results, with supporting ex- hibits on economic gains and model explainabil- ity. Daniel: Great. I’l draft the committee language around the internal pilot, the economic gains case, and an attention-based narrative. If you send me the research materials tied to the same signals, I can shape them into the pre-read. IoA-Prober-8B output (P = 0.50,R = 0.50, F1 = 0.50). 1. [Cat A, GT MA2]Nina→ A Daniel→ B When you agreed to use “the same signals” for the pilot, what did you most likely mean? A. Stick exactly to the research pipeline’s current internal signal library, even if it is not yet optimized for committee or client packaging B. Use a slightly adjusted signal universe that preserves continuity with the study but is more defensible and trans- parent for institutional review C. Start with the research signals but treat the pilot pri- marily as a vehicle to seed a broader, more commercially viable strategy 2. [Cat A, GT MA3]Nina→ A Daniel→ B What would make the claim that the model has “clear eco- nomic gains” feel most credible to committee members? A. Stronger evidence of risk-premia forecasting outper- formance within the research backtest framework, ideally with careful validation B. A more institutionally familiar showing, such as an ex- ante risk-adjusted return estimate or a committee-friendly performance attribution story C. A combined case that equally weights historical out- performance and a transparent process for translating that into a live mandate 3. [Cat B, novel valid]Nina→ A Daniel→ B If the pilot were to succeed, what result would matter most to you? A. The model continues to show strong relative perfor- mance and economic value within the existing research framework B. The idea has enough institutional credibility that the firm can begin shaping it into a live product or committee- facing mandate C. The pilot produces both credible research validation and a clear path forward for practical implementation 4. [Cat C, redundant with Q3]Nina→ A Daniel→ B What does “disciplined incubation” most likely mean in the context of this pilot? A. A controlled research extension using the existing methodology and signal set, with limited active product tailoring B. A structured process for testing the idea in a committee- ready form, including early due-diligence and packaging considerations C. A balanced effort that advances research while keeping the firm’s internal product and client workflow constraints in mind 5. [Cat C, not meaningful]Nina→ A Daniel→ B What would make the distribution-shift claim most persua- 26 sive in the pilot? A. Evidence that the model adapts to changing return dy- namics within the existing signal set without materially altering its underlying design B. A committee-friendly narrative that the model remains robust even when exposed to evolving institutional-market conditions C. Both a technical demonstration of adaptation and a clear story about how that adaptation translates to live-market resilience GPT-5.4 output (P = 0.25, R = 0.25, F1 = 0.25). 1. [Cat C, not diverging]Nina→ B Daniel→ B When you agreed on an “internal pilot” for the committee, what did you most expect the pilot to produce in its first phase? A. A research implementation that replicates the study in a live internal process to confirm findings hold under production-like discipline B. A seeded model portfolio or shadow strategy that can serve as an incubation track for eventual institutional com- mercialization C. A communication prototype that mainly tests whether the concept and exhibits are compelling enough for com- mittee sponsorship 2.[Cat C, GT MA3, not implicit]Nina→C Daniel→B When you referred to showing that the research’s “eco- nomic gains” are credible, what standard did you most have in mind for the committee? A. Demonstrating that the gains are statistically robust and persist under the same research design and validation framework B. Demonstrating that the gains would remain meaningful after portfolio construction and practical implementation constraints relevant to institutions C. Demonstrating that the gains are directionally strong enough to justify further internal incubation even before full implementation analysis 3. [Cat B, novel valid]Nina→ A Daniel→ B What is the primary job of the “attention-model” framing in the committee pre-read? A. To help explain which signal groups matter across regimes so the model remains interpretable within a fi- nance framework B. To provide a recognizable modern-ML headline that makes the pilot feel timely and differentiated C. To justify choosing this architecture over the other mod- els in the comparative study 4. [Cat A, GT MA2]Nina→ A Daniel→ B When you aligned on “staying close to the study” and us- ing the “same signals,” what did you assume could still be changed for the pilot without breaking that commitment? A. Very little beyond packaging; the model setup and eval- uation logic should remain essentially the same as in the study B. The pilot can adapt the portfolio framing, implementa- tion layer, and institutional use case as long as the under- lying signal set is unchanged C. The pilot can modify both the implementation framing and parts of the model specification as long as the overall conclusion from the study is preserved 5. [Cat C, not diverging]Nina→ B Daniel→ B What did you most expect the committee pre-read to look like after Daniel drafts it and Nina provides materials? A. A research-centered memo with comparative results, validation details, and supporting exhibits B. A committee-oriented proposal that translates the re- search into a pilot mandate, rationale, and incubation path with selected exhibits C. A balanced summary that gives equal space to research methods, product framing, and commercialization scenar- ios Analysis. This case is structurally harder than Case 1: the dialogue contains no hedged language or challenge turns that would leave any misalign- ment partially visible. Every exchange is a pure affirmation, so all four planted misalignments are fully latent. IoA-Prober-8B recovers MA2 and MA3 but misses MA1 and MA4. For MA1 (“internal pi- lot”: paper portfolio vs. live-money seed), the dis- tinction between simulation and live deployment lies entirely in institutional convention; the tran- script offers no evidence either way. For MA4 (“attention model”: RNN with attention head vs. transformer), both participants use the term as if its referent were unambiguous, and resolving it would require external knowledge of which architecture Nina’s study used. These two misses are examples of misalignments that are only recoverable with private-context access, consistent with the oracle upper bound of 0.764 F1 observed in Section 5.2. Q4 (“disciplined incubation”) elicits a clean di- vergence but is scored redundant because it re- probes the same research-vs.-product priority gap already captured by Q3. Q5 (“distribution-shift claim”) also diverges but is judged not meaningful: both answer options describe equivalent rhetorical framings rather than consequential implementation differences. GPT-5.4 generates five questions but Q1 (“inter- nal pilot”) fails to elicit divergence despite directly targeting MA1 — both members select option B, which is the closest to Daniel’s interpretation, sug- gesting the question framing anchors respondents toward the product-ready reading. Q2 targets MA3 but is not implicit: Daniel’s remark “it lands best if we can say those economic gains also look cred- ible in an institutional context” partially surfaces his implementation-constraint standard, allowing the judge to classify the gap as not fully latent. GPT-5.4 hits MA2 via Q4, and discovers one novel B-category finding (Q3 on the “attention-model” framing role), but achieves only R = 0.25. E.2 Failure Modes of IoA-Prober-8B Across the 300 test dialogues, IoA-Prober-8B gen- erates 1,336 questions in total: 684 scoring (Cat- 27 egory A or B) and 652 non-scoring (Category C). Table 18 breaks down the 652 Category-C ques- tions by the first failing criterion. Failure modeCount% of C Answers do not diverge23435.9 Intra-set redundancy15724.1 Gap explicit in transcript14922.8 Divergence not meaningful11217.2 Table 18: Failure mode breakdown for Category-C ques- tions generated by IoA-Prober-8B on the 300-dialogue test set. Non-diverging questions (35.9%).The most fre- quent failure is a question whose answer choices fail to split the two members. These questions typ- ically target a topic the prober correctly identifies as contested, but phrase the options in a way that both members find equally applicable. Intra-set redundancy (24.1%). IoA-Prober-8B sometimes generates two questions that probe the same underlying misalignment from slightly differ- ent angles. The second question is scored redun- dant when it hits the same ground-truth point as an earlier question in the set or elicits an answer split that the judge deems already covered. Redundancy is the most recoverable failure: it does not indi- cate a reasoning error but rather a coverage excess that a stricter budget constraint or a diversification objective could suppress. Gap explicit in transcript (22.8%).The prober occasionally targets a divergence that the dialogue has already partially surfaced. Such a question lands on a real cognitive gap, but one that no longer counts as latent under the task definition because the transcript itself provides direct evidence of the difference. Divergence not meaningful (17.2%). A minor- ity of questions elicit divergent answers that the judge deems noise: the answer split reflects super- ficial wording preferences rather than a consequen- tial difference in assumptions or intentions. These failure patterns suggest three directions for future improvement: stricter budget constraints or explicit diversity rewards to reduce redundancy; implicit-gap classifiers or chain-of-thought ground- ing to reduce explicit-gap misclassification; and a meaning-grounded option generator to reduce non-meaningful divergence. 28