Paper deep dive
Test-Time Scaling for Small VLMs on Multilingual Visual MCQ
Spiros Baxevanakis, Peng-Jian Yang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 7/13/2026, 4:29:41 AM
Summary
This paper evaluates test-time scaling (TTS) techniques for small open vision-language models on multilingual visual multiple-choice questions. It finds that parseability and per-chain token budget are the primary drivers of accuracy improvements, significantly outperforming complex search strategies or post-hoc selectors. Upgrading the base policy model (Qwen3.5-4B) and increasing the token limit to 2k tokens yield the largest gains, with the best configuration achieving 84.1% accuracy on the ImageCLEF 2026 test split.
Entities (10)
Relation Signals (6)
Qwen3.5-4b → achievesbestaccuracyon → ImageCLEF 2026
confidence 95% · Our best configuration reaches 84.1% on the held-out ImageCLEF 2026 test split, ranking first on the Visual MCQ leaderboard.
Per-chain token budget → dominates → Chain count
confidence 95% · Token budget dominates chain count. Doubling the budget from 1k to 2k tokens recovers+3.7pp... Doubling chains adds only+0.15pp.
Self-Consistency → outperforms → PRM-guided beam search
confidence 95% · PRM-guided beam search trails plain self-consistency by 0.39 pp at over eight times the cost
Test-Time Scaling → appliedto → Qwen3.5-4b
confidence 90% · We investigate whether describe-then-reason with PRM-guided search beats flat self-consistency... under a single-A40 envelope
Majority Vote → beats → Qwen-VL-PRM-7B
confidence 90% · neither a training-free generative critic nor a trained multimodal PRM beats majority vote across both policies.
Guided parse repair → eliminates → Parse failures
confidence 90% · This eliminates parse failures by construction.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Test-time scaling (TTS) reliably improves reasoning in large language models, but whether it transfers to small open vision-language models remains unclear. We examine this on EXAMS-V, a multilingual visual multiple-choice benchmark, comparing self-consistency, describe-then-reason with PRM-guided beam search, and two post-hoc selectors across Qwen2.5-VL-7B-Instruct and Qwen3.5-4B. What matters is the conditions under which TTS runs, not the search or verification machinery. The largest factor is parseability: an early prompt format left many chains reasoning correctly yet never committing to an answer letter, which a standard answer cue and a guided repair step largely remove. A larger decoding budget removes the rest: raising the per-chain token limit from 1k to 2k recovers 3.7 pp, whereas sampling more chains (8 to 16) adds only 0.15 pp. Once chains have room to finish, elaborate methods contribute little: PRM-guided beam search trails plain self-consistency by 0.39 pp at over eight times the cost, and neither a training-free generative critic nor a trained multimodal PRM beats majority vote across both policies. The largest gain comes instead from the policy model itself (+11.4 pp). Our best configuration reaches 84.1% on the held-out ImageCLEF 2026 test split, ranking first on the Visual MCQ leaderboard.
Tags
Links
- Source: https://arxiv.org/abs/2607.09438v1
- Canonical: https://arxiv.org/abs/2607.09438v1
Trouble viewing inline? Open PDF directly →
Full Text
42,729 characters extracted from source content.
Expand or collapse full text
Nika at ImageCLEF 2026 Task on Multimodal Reasoning: More Tokens, Fewer Trees — Test-Time Scaling for Small VLMs on Multilingual Visual MCQ ImageCLEF Lab at CLEF 2026 Spiros Baxevanakis 1,† , Peng-Jian Yang 1,† 1 University of Amsterdam, Science Park 904, 1098 XH Amsterdam, The Netherlands Abstract Test-time scaling (TTS) reliably improves reasoning in large language models, but whether it transfers to small open vision-language models remains unclear. We examine this on EXAMS-V, a multilingual visual multiple- choice benchmark, comparing self-consistency, describe-then-reason with PRM-guided beam search, and two post-hoc selectors across Qwen2.5-VL-7B-Instruct and Qwen3.5-4B. What matters is the conditions under which TTS runs, not the search or verification machinery. The largest factor is parseability: an early prompt format left many chains reasoning correctly yet never committing to an answer letter, which a standard answer cue and a guided repair step largely remove. A larger decoding budget removes the rest: raising the per-chain token limit from 1k to 2k recovers 3.7 p, whereas sampling more chains (8 to 16) adds only 0.15 p. Once chains have room to finish, elaborate methods contribute little: PRM-guided beam search trails plain self-consistency by 0.39 p at over eight times the cost, and neither a training-free generative critic nor a trained multimodal PRM beats majority vote across both policies. The largest gain comes instead from the policy model itself (+11.4p). Our best configuration reaches 84.1% on the held-out ImageCLEF 2026 test split, ranking first on the Visual MCQ leaderboard. Keywords test-time scaling, vision-language models, multimodal reasoning, self-consistency, process reward models, 1. Introduction Multimodal reasoning remains a challenge for vision–language models (VLMs) whenever questions combine structured visual content with multi-step inference across languages. Real exam questions are a natural stress test of this combination, and the EXAMS-V benchmark [1] spans thirteen languages and twenty school subjects rendered as text-in-image panels. The 2026 ImageCLEF multimodal reasoning shared task [2,3] fixes the practical constraints: a single A40 GPU, open-weight policies with at most 7B parameters, and a reproducibility mandate. The natural question is whether the inference-time techniques that advanced text-only language models [4,5,6] transfer to open VLMs under this budget. The evidence is cautionary: self-refinement degrades open-source VLMs at this scale [7], long sequential “thinking” often underperforms parallel sampling on small policies [8,9], and discriminative reward models trained on mathematics fail to transfer to humanities and non-English content [10,11]. A viable pipeline must externalise selection, prefer parallel sampling to long single-chain reasoning, and avoid iterative self-refinement loops. We investigate whether describe-then-reason with PRM-guided search beats flat self-consistency; how accuracy scales along two axes (chain count and per-chain token budget) under a single-A40 envelope [6]; where TTS helps and fails across subjects and languages [7]; and whether a training-free generative critic can beat majority vote on low-agreement questions, with a cross-model PRM [12] controlling for sycophantic self-scoring [13]. The latter extends generative-verifier findings [14,15,16] to a multimodal, multilingual setting under a≤7B budget not previously studied to our knowledge. CLEF 2026 Working Notes, 21 – 24 September 2026, Jena, Germany † These authors contributed equally. $ spiros.baxevanakis@student.uva.nl (S. Baxevanakis); lesterpjy@gmail.com (P. Yang) © 2026 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (C BY 4.0). Summary of findings. Across all four dimensions, the policy substrate and engineering details (parser format, decoding budget) dominate any search or selector method. Our best configuration, Qwen3.5-4B at SC-푁 =16 with a 2,048-token budget, reaches 81.6% on the full validation set. 1 2. Related Work Test-time scaling: theory and limits.Snell et al.[6]formalize TTS as a two-axis allocation (width: parallel chains; depth: per-chain revision or search), proving compute-optimal allocation can match naive Best-of-푁at less cost [17,18]. Setlur et al.[19]show the verifier advantage grows with reasoning depth, motivating an external selector; converging findings [8,9,20,7] argue for parallel sampling and against self-correction at our policy scale. Multimodal verification and PRM generalization. A process reward model (PRM) assigns each reasoning step a probability푃 (+)that the step is correct [5]; it is useful only when푃 (+)sharply separates correct from incorrect steps. Two multimodal PRMs supersede VisualPRM-8B [21]: Athena- PRM-7B [22] and Qwen-VL-PRM-7B [12], which separates perception from reasoning errors. On the text side, discriminative math-trained PRMs fail out of distribution [10,11,23] while generative verifiers transfer more reliably [14,15,16]. Prior EXAMS-V work [24,25] informs our scaffold; on the search axis, PRM-BAS [26] dominates step-level Best-of-푁 at matched compute. 3. Method We evaluate two pipelines: describe-then-reason with PRM-guided beam search and a three-way selector contrast, and flat self-consistency with a large token budget. Policy and data. We evaluate two open-weights policies:Qwen/Qwen2.5-VL-7B-Instruct(7B pa- rameters, bfloat16) for the baseline and ablation configurations, andQwen/Qwen3.5-4B(4B parameters, bfloat16), a smaller but newer-generation policy, for the headline configuration. We cap image res- olution atmax_pixels=1,003,520(≈4k vision tokens after Qwen smart-resize), balancing legibility of text-in-image exam panels against context budget: atmax_model_len=16,384, this leaves∼12k tokens for the prompt and generated reasoning. Inference uses vLLM [27] with batched parallel decoding (SamplingParams(n=N)) and automatic prefix caching. All experiments run on a single A100-40 GB GPU, which matches the shared task’s A40 memory constraint. We evaluate on the full EXAMS-V validation split (4,651 questions, 11 languages) for headline numbers and a 200-question (language, subject)-stratified subset (seed 42, minimum ten per language) for ablations whose full- validation cost is prohibitive. Describe-then-reason.Following Ahmed et al.[24], the describe-then-reason pipeline first samples 푁image-grounded descriptions of the question at푇 =0.7. For each description, with the image removed, the policy generates푀reasoning chains in the flat variant, or reasoning step-by-step via PRM-BAS-style beam annealing [26] in the search variant: an initial beam of width퐵 0 is expanded with퐵continuations per surviving beam, each step is scored by Qwen-VL-PRM-7B [12], which retains access to the original image, and the beam width is halved whenever the score range among the top beams exceeds a threshold 휏. Beams terminate on the answer cue or at a maximum depth푑. Across all (description×terminal beam) candidates, a final selector picks the answer. In our experiments the flat variant uses푁 =4,푀 =4 (20calls/q, bracketing SC-푁 =8from above to test whether structured decomposition helps at higher compute); the search variant halves descriptions to푁 =2because beam expansion is expensive, and adopts the hyperparameters reported in Hu et al. [26]: 퐵 0 =4, 퐵=2, 휏 =0.05, 푑=6 (∼13 calls/q). 1 Code available at: https://github.com/lesterpjy/tts-small-vlm. Selectors.Three selectors are applied post hoc to the same chain pool. Majority vote: most common answer letter with log-probability tie-breaking (no extra inference). A training-free generative critic inspired by Zhang et al.[15]scores each chain on three axes (Appendix F). A discriminative PRM (Qwen-VL-PRM-7B;12), a separately trained model, scores each chain in one-shot mode (Appendix E), controlling for sycophantic self-preference [13]. Both models co-reside on the A100-40 GB at bfloat16. Each selector uses a skip rule [28]: high-confidence majorities (≥5/8agreement) pass through; only the weak (≤4/8) and all-unparseable tiers are rescored. Self-consistency.The simpler pipeline samples푁 ∈8, 16chains from Qwen3.5-4B at푇 =0.7with max_new_tokens=2048, conditioned on a CoT prompt ending with the MMMU closer [29,30]. A regex extracts the answer letter (mapping Cyrillic labels to A–E for multilingual coverage). The final answer is the majority letter across parseable chains, with log-probability tie-breaking. We additionally sweep 푇 ∈0.3, 0.5, 0.7, 0.9 at full validation scale. Guided parse repair.When every chain for a question fails to produce an answer letter (2–18%of questions depending on budget), each chain’s reasoning is concatenated with the cue :and a single token is decoded under vLLM’sguided_choiceconstraint over퐴,퐵,퐶,퐷,퐸. The majority across all committed letters (original plus repaired) is the final answer. This eliminates parse failures by construction. Stratification and evaluation. We partition accuracy by subject and language; per-subject results are reported only for cells with푛≥50. Pairwise method comparisons use McNemar’s paired test with Bonferroni correction for the number of reported comparisons. 4. Experiments We evaluate seventeen configurations across two policies, three search strategies, three selectors, and two scaling axes. Table 1 summarises them; the remaining sections report findings ordered by headline result rather than by configuration. Unless noted, all headline numbers are on the full EXAMS-V validation set (푛=4,651); configurations marked “dev-200” use a 200-question stratified subset. 5. Results and Discussion We report results grouped by headline finding. 5.1. The TTS progression and the parser artifact The three-stage TTS progression (zero-shot, chain-of-thought, self-consistency) holds for both policies at full validation scale (Table 2). Each step yields a significant gain under Qwen2.5-VL-7B, and the progression is more pronounced under Qwen3.5-4B, where the full span from zero-shot to SC-푁 =8 (2k) exceeds 24 p. However, the relative contribution of each stage shifts with the policy. Under Qwen2.5, the CoT-to-SC margin is modest (+3.3p), suggesting that eight samples add limited signal once CoT prompting is in place. Under Qwen3.5, the margin widens markedly: CoT accounts for roughly half the total gain, and the remaining half comes from parallel sampling and the token-budget increase (Section 5.3). A parser artifact masked part of the CoT gain.An early prompt format left∼9%of development chains unparseable. Switching to the MMMU-standard closer [29,30] cut parse failures below2% and raised dev accuracy by∼6p; the fix corroborates at val scale (+0.82p on Q2.5 SC-푁 =8). The methodological lesson: what looks like reasoning failure can be an extraction failure, and TTS studies that do not control for parseability risk attributing engineering artifacts to the scaling method. Table 1 Experimental configurations. “Policy” is Q2.5=Qwen2.5-VL-7B-Instruct, Q3.5=Qwen3.5-4B. Calls/q is the typical end-to-end VLM forward count. Header rows group configurations by theme. All full- validation configurations use 푛=4,651. MethodPolicySearch / samplingSelectorCalls/qScale Baselines B0 zero-shotQ2.5guided_choice, 1 sample—1val-full B1 chain-of-thoughtQ2.51 sample, MMMU closerpost-hoc regex1val-full B2 self-consistencyQ2.5 푁 =8, 푇 =0.7, 1024-tokmajority vote8val-full B0 zero-shotQ3.5guided_choice, 1 sample—1val-full B1 chain-of-thoughtQ3.51 sample, MMMU closerpost-hoc regex1val-full B2 self-consistencyQ3.5 푁 =8, 푇 =0.7, 1024-tokmajority vote8val-full Search-strategy contrasts (dev-200) S-DTRQ2.5DTR 푁 =4,푀 =4majority vote20dev-200 S-PRM-BASQ2.5DTR 푁 =2 + PRM-BAS (퐵 0 =4,퐵=2,푑=6)majority vote∼13dev-200 Scaling (Q3.5, val-full) SC 1k-tokQ3.5 푁 =8, 푇 =0.7, 1024-tokmajority vote8val-full SC 2k-tokQ3.5 푁 =8, 푇 =0.7, 2048-tokmajority vote8val-full SC 푁 =16Q3.5 푁 =16, 푇 =0.7, 2048-tokmajority vote16val-full Temp. sweepQ3.5 푁 =8, 푇∈.3,.5,.7,.9, 2kmajority vote8val-full Selectors (full val, both pools) R1 generative criticQ2.5SC poolGM-PRM-style critic+3/chval-full R2 Qwen-VL-PRMQ2.5SC poolPRM 1-shot + skip+1/chval-full R1 generative criticQ3.5SC-2k poolGM-PRM-style critic+3/chval-full R2 Qwen-VL-PRMQ3.5SC-2k poolPRM 1-shot + skip+1/chval-full Best configuration S-bestQ3.5 푁 =16, 푇 =0.7, 2048-tokmajority + guided repair ≤16val-full Table 2 TTS progression on full validation (푛=4,651). All Q2.5 pairs are Bonferroni-significant at훼=0.017. Q3.5 SC rows show the effect of doubling the token budget. P.-fail: fraction of questions with all chains unparseable. Policy MethodAcc. 95% CIP.-fail C/q Q2.5Zero-shot56.83 [55.4, 58.2]0.0%1 Q2.5CoT63.10 [61.7, 64.5]2.0%1 Q2.5SC-푁 =8 (1k)66.42 [65.0, 67.8]0.2%8 Q3.5Zero-shot57.11 [55.7, 58.5]0.0%1 Q3.5CoT69.66 [68.3, 71.0]16.2%1 Q3.5SC-푁 =8 (1k)77.81 [76.6, 79.0]18.4%8 Q3.5SC-푁 =8 (2k)81.49 [80.4, 82.6]2.7%8 Q3.5SC-푁 =16 (2k) 81.64 [80.5, 82.7]1.9%16 Under Q3.5, every step is significant: CoT over zero-shot gains+12.6p (non-overlapping CIs), the 1k→2k budget increase recovers+3.7p (171additional correct answers), and the푁 =8→ 푁 =16step adds only+0.15p (7questions, within sampling noise). Q3.5 CoT itself suffers16.2%parse-fail from the same truncation that affects SC (18.4%at 1k), so the reported69.66%underestimates single-chain accuracy and the CoT-to-SC(1k) gain conflates truncation absorption with diversity benefit. 5.2. Structured search underperforms flat sampling Dev-200 CIs overlap between PRM-BAS, DTR, and flat SC-푁 =8(Table 11, Appendix H). A val-scale run of PRM-BAS under Qwen3.5-4B (푁 =2,퐵 0 =4,퐵=2,휏 =0.05,푑=6;푛=4,319,93%of validation) resolves the ranking in favour of SC: PRM-BAS reaches80.74%against SC majority at81.13%on the same questions (−0.39p). Of the14.6%of questions where PRM-BAS produces a different answer than SC,233are corrections and250are regressions, for a net loss of17questions. Two bottlenecks interact: in the DTR pipeline the reasoning stage operates on the text description alone (no image Table 3 Two-axis scaling on Qwen3.5-4B, MMMU closer. Top: dev-200 exploration. Bottom: val-full (푛=4,651) validation of the key points. Dev-200’s +2 p 푁 -scaling signal collapses at val-full. Scale Config.Acc. P.-fail 푠/q Dev-200 exploration dev푁 =8, 푇 =0.7, 1024-tok∼6518.0%∼7 dev푁 =8, 푇 =0.7, 2048-tok80.02.5%∼14 dev푁 =8, 푇 =0.7, 3072-tok81.01.5%∼21 dev푁 =16, 푇 =0.7, 2048-tok82.02.5%∼25 Val-full validation (푛=4,651) val푁 =8, 푇 =0.7, 1024-tok77.8118.4%8.8 val푁 =8, 푇 =0.7, 2048-tok81.492.4% 11.8 val푁 =16, 푇 =0.7, 2048-tok 81.641.9% 14.6 Temperature sweep (val-full, 푁 =8, 2048-tok) val푇 =0.380.842.7% 11.2 val푇 =0.581.012.6% 11.5 val푇 =0.781.492.4% 11.8 val푇 =0.981.043.0% 12.3 access), so visual detail lost at the description step cannot be recovered downstream; beam search compounds this loss because the PRM scoring each step (though it retains image access) lacks sufficient discriminative signal to steer search toward the missing detail. Diagnosis. Hu et al.[26]show that PRM-guided beam-annealing search dominates flat Best-of- 푁on multimodal math benchmarks. Our result inverts this: on multilingual visual MCQ, PRM-BAS underperforms flat SC by−0.39p at8.7×the cost. Two factors explain the discrepancy. First, the PRM’s per-step P(+) saturates (0.962on descriptions,0.849on reasoning), collapsing search diversity:71.9% of val questions have all surviving beams agreeing on the same letter (Shannon entropy0.259/2.322 bits,89%reduction from uniform); the8.7×cost yields near-zero answer diversity. Second, slicing by SC agreement tier (Appendix A) reveals that on high-confidence questions (≥7/8,91.8%SC accuracy), PRM-BAS regresses by−2.16p (48corrections vs.119regressions), dominating the overall deficit. Only on the108all-unparseable questions does PRM-BAS produce a clear positive (37correct,34.3%). The P(+) asymmetry identifies reasoning as the structural bottleneck (discussed in Section 7). 5.3. Token budget dominates chain count We sweep two scaling axes (per-chain token budget and number of sampled chains), first on the 200-question development subset, then on the full validation set (Table 3). Token budget dominates chain count.Doubling the budget from 1k to 2k tokens recovers+3.7p (Figure 1a): chains generate valid reasoning but are truncated before emitting an answer letter, producing parse failures rather than reasoning failures. Doubling chains (푁 =8→ 16) adds only+0.15p. The asymmetry is structural:∼60%of questions already reach strong agreement at푁 =8, while the weak- agreement tail (≤4/8,23%) has below-chance accuracy that additional sampling does not correct. (Appendix A). For compute-constrained TTS, ensuring chains run to completion matters more than sampling more of them. Temperature sensitivity.An inverted-U optimum at푇 =0.7replicates from dev-200 to full validation (Figure 1b), but the spread compresses to0.65p at scale, suggesting the policy’s sampling diversity is already near optimal. 12481620 VLM calls per question (log scale) 50 55 60 65 70 75 80 85 90 Accuracy (%) +3.7 p (token budget) +0.15 p (chain count) (a) Accuracy vs. compute Qwen3.5-4B (val-full) Qwen2.5-VL-7B (val-full) DTR / PRM-BAS (Q2.5, dev-200) 0.30.50.70.9 Sampling temperature T 79.0 79.5 80.0 80.5 81.0 81.5 82.0 82.5 83.0 83.5 Accuracy (%) 80.84 81.01 81.49 81.04 (b) Temperature sweep (Q3.5 SC-N=8, 2k-tok) Figure 1: (a) Accuracy vs. VLM calls per question. At푁 =8, the token-budget arrow marks+3.7p from 1k→2k tokens at constant call count; the chain-count arrow marks+0.15p from doubling푁at constant budget. DTR search points (dev-200, Q2.5) sit below flat SC at higher compute. (b) Temperature sweep on Q3.5 SC-푁 =8 (2k-tok, val-full). Error bars are 95% Wilson CIs. 5.4. Guided repair closes the parse-fail residue After the budget fix, val SC-푁 =8still has124all-unparseable questions (2.67%); the test split (푛=1,117) has82(7.34%), elevated because the test mix is44.8%English and30.6%Chinese. Guided repair drives both rates to zero at one decoded token per affected chain. The repair targets valid reasoning that was never committed to an answer letter, a failure mode that chain scaling cannot reach, recovering more than doubling 푁 (+0.15 p val) at a fraction of the compute. 5.5. Policy choice dominates search and selection Switching from Qwen2.5-VL-7B to Qwen3.5-4B at matched settings yields+11.4p (Table 2, SC- 푁 =8rows); with the budget increase the gap widens to15.1p and all eleven languages improve by≥3p (Table 10). Qwen3.5-4B is smaller; the gain reflects newer-generation training, consistent with Ahmadpour et al.[7]. At our budget scale, upgrading the policy dominates any inference-time strategy. 5.6. Selectors yield a null result on both pools We test whether a post-hoc selector improves on majority vote. Two selectors are compared: a training- free generative critic (three rubric axes; Appendix F) and a discriminative PRM (Qwen-VL-PRM-7B; one-shot scoring). Both apply a skip rule preserving high-confidence majorities. We evaluate on two pools of different baseline strength (66% Q2.5, 81% Q3.5) to control for ceiling effects (Table 4). Cross-pool replication. The null replicates across both policies: on Q2.5 (66%) the critic is an exact null and the PRM gains+0.45p (n.s.); on Q3.5 (81%) both turn negative (PRM:51corrected vs.55regressed; critic:37vs.59). As pool accuracy grows, the majority-wrong tail shrinks and a near-balanced selector’s net effect turns negative (Appendix B). Why both selectors fail. The critic exhibits self-recognition bias [13]:40%of chains receive the modal score0.75, and the correct/incorrect gap is negligible (0.713vs.0.687; Appendix B). The PRM Table 4 Selectors on full validation (푛=4,651). Both selectors are tested on two candidate pools (Q2.5 SC-푁 =8and Q3.5 SC-푁 =8at 2048 tokens). All use the skip-on-high rule; R2 uses one-shot PRM scoring. The Q2.5 pool predates the MMMU closer (Section 5.1); its majority anchor is therefore65.60%rather than the post-closer66.42%in Table 2. Pool SelectorAcc.Δ Q2.5 SC majority (anchor) 65.60%— Q2.5 R1 generative critic65.60%+0.00 p Q2.5 R2 PRM (1-shot)66.05% +0.45 p n.s. Q3.5 SC majority (anchor) 81.49%— Q3.5 R2 PRM (1-shot)81.40%−0.09 p Q3.5 R1 generative critic81.01%−0.47 p n.s. Q2.5 R2: McNemar 휒 2 =1.80, 푝≈0.18. Q3.5: both negative, not tested. Arabic Bulgarian Chinese Croatian English French German Hungarian Italian Polish Serbian 20 30 40 50 60 70 80 90 100 Accuracy (%) +8 +17 Overall best (81.5%) Q2.5 SC-N=8 Q3.5 SC-N=8 (1k-tok) Q3.5 SC-N=8 (2k-tok) Figure 2: Per-language accuracy on full validation. Red annotations mark the accuracy gain from the budget increase (1k→2k) for languages where it exceeds+5p. The dashed line is the overall best (81.5%). improves the correction rate (37.6%vs.31.8%), partially mitigating self-preference, but cannot distin- guish right from wrong overrides: the score gap between its top-ranked chain and runner-up is 0.049 on corrections and 0.048 on degradations, yielding a net effect indistinguishable from noise. 5.7. Stratified analysis Figure 2 and Table 10 (Appendix G) report per-language accuracy under both policies, with the Q3.5 column split by token budget to isolate the effect of the budget increase. The budget increase concentrates gains on English and Chinese. English gains+17.3p and Chinese+8.5p when the budget doubles; under Q2.5, over a third of English questions had all eight chains unparseable. Most other languages gain only1–2p and French (92%) is unchanged, confirming a targeted rather than general effect. Persistent below-average strata. Chinese, Polish, English, and Arabic remain below81.5%even after the budget increase. Under Q2.5, SC regresses on Chinese by2.8p: a chain-agreement audit (Appendix C) shows22.3%of Chinese questions reach unanimous8/8agreement (vs.≤15%elsewhere). SC gains power when chains err independently; when errors are highly correlated, the vote amplifies the shared mistake. Languages benefiting most from SC (Bulgarian, Croatian, Hungarian, Serbian) sit at mid-accuracy where the independence condition holds. Per-subject patterns.Per-subject accuracy (Appendix D) reveals a split: STEM subjects benefit most from TTS (Physics+14p, Mathematics+20p over zero-shot), while the weakest Q3.5 subjects (Social 43%, Islamic Studies57.8%) are text-heavy. Using subject as a proxy for visual complexity (EXAMS-V lacks content-type annotations), the pattern suggests two regimes: in STEM, reasoning errors vary across chains and majority vote aggregates them away; in the weakest subjects, the bottleneck is missing domain knowledge that no amount of sampling can compensate. 5.8. Held-out test performance Our best configuration (Qwen3.5-4B at SC-푁 =16,2,048-token budget, with guided repair) ranks first on the official Visual MCQ leaderboard at84.06%overall, and is first on each of the six test languages (Table 12, Appendix I). The test split (푛=1,117) is44.8%English and30.6%Chinese, and both exceed their validation accuracy: English rises to85.6%(from73.5%val,+12.1p) and Chinese to78.1%(from74.7%, +3.4p). These are the two languages where the token-budget and guided-repair fixes (Sections 5.3, 5.4) had the most headroom on validation, so the test result tracks with the validation diagnosis. 6. Conclusion We studied test-time scaling for open VLMs on multilingual visual MCQ, comparing structured search against flat self-consistency across two policies on the full EXAMS-V validation set. TTS improves accuracy: the full progression from zero-shot to self-consistency spans over 24 p on Qwen3.5-4B, but the gains concentrate on the per-chain token budget rather than on structured search, larger chain counts, or trained selectors: neither a generative critic nor a trained PRM beats majority vote across two policy models and two pool strengths. Stratified analysis confirms that TTS gains concentrate on mid-accuracy languages where the policy is miscalibrated rather than weak, and that persistent below-average strata (Polish, Islamic Studies) resist SC because chains converge on the same wrong answer, leaving no diversity for the majority vote to correct. The overarching lesson is that at our budget scale, the engineering substrate (parser format, decoding budget) and the policy choice dominate any search or selector method. Our best configuration, Qwen3.5- 4B at SC-푁 =16with a2,048-token budget and guided repair (Section 5.4), reaches81.6%on full validation and84.1%on the held-out ImageCLEF 2026 test split, ranking first on the Visual MCQ leaderboard (Section 5.8). Future work should investigate whether decoupling perception from reasoning via a dedicated reasoning model can address the P(+) asymmetry identified in Section 5.2, and whether better-calibrated PRMs trained on multilingual non-mathematical content would break the selector null. 7. Limitations Our findings rest on two policy models, Qwen2.5-VL-7B-Instruct and Qwen3.5-4B, both from the same family, so we cannot separate behavior typical of small open VLMs from behavior specific to Qwen models. The parseability effect in particular depends on how a model emits answer letters under a given prompt format, which may differ for other open source VLM families. The 11.4 p gain from switching policy supports our claim that the policy substrate dominates inference-time strategy, but also shows that a third model could shift the picture; whether the pattern holds across families and across the sub-7B range is untested. Ablations whose full-validation cost was prohibitive used the dev-200 subset, whose±7 p Wilson CIs preclude ranking claims, so all headline findings rest on val-full re-runs (n=4,651). The segmented PRM null appears structural, since P(+)≈0.85on reasoning steps leaves little signal once aggregated, though better per-step calibration on multilingual mixed-subject content might change this. Following the 2025 ImageCLEF winners [24], we did not fine-tune, use few-shot exemplars, or apply test-time image augmentation, all of which remain open. Acknowledgments Experiments were conducted on the Snellius national supercomputer, provided by the University of Amsterdam. Declaration on Generative AI During the preparation of this work, the authors use AI for proofreading, polishing and rephrasing sentences and paragraphs in the manuscript. References [1]R. J. Das, S. E. Hristov, H. Li, D. I. Dimitrov, I. Koychev, P. Nakov, Exams-v: A multi-discipline multilingual multimodal exam benchmark for evaluating vision language models, 2024. URL: https://arxiv.org/abs/2403.10378. arXiv:2403.10378. [2]B. Ionescu, H. Müller, D. Stanciu, A. Radu, R. Bolborici, M. Negru, A. Ene, V. Vasilescu, A.-A. Nicolae, L. Ştefan, M. Constantin, M. Dogariu, A. Andrei, H. Damm, T. M. G. Pakull, A. Ben Abacha, A. García Seco de Herrera, C. M. Friedrich, R. Brüngel, L. Reinartz, H. Schäfer, C. S. Schmidt, B. Bracke, P. Nath, B. Eryılmaz, M. Hjuler, D. Fabre, C. Lemaire, B. Lecouteux, D. Schwab, D. Dimitrov, M. S. Hee, M. Ahsan, S. Ahmad, D. Zlatkova, G. Pachov, Z. Xie, P. Nakov, I. Koychev, J. E. Heras Rivera, D. K. Low, W. Yim, J. Ruzevick, D. Child, M. Kurt, Z. Sun, F. Xia, M. Yetisgen, A. Radzhabov, Y. Prokopchuk, V. Kovalev, D. Karpenka, S. A. Hicks, S. Gautam, M. A. Riegler, V. Thambawita, P. Halvorsen, M. El Sakka, J. Mothe, A. Băicoianu, C. Florea, M. Ivanovici, Overview of imageclef 2026: Multimodal challenges in medicine, science, agritech, and security, in: Experimental IR Meets Multilinguality, Multimodality, and Interaction, Proceedings of the Seventeenth International Conference of the CLEF Association (CLEF 2026), Springer Lecture Notes in Computer Science LNCS, Jena, Germany, 2026. [3]D. Dimitrov, M. S. Hee, M. Ahsan, S. Ahmad, D. Zlatkova, G. Pachov, Z. Xie, P. Nakov, I. Koychev, Overview of the ImageCLEF 2026 Task on Multimodal Reasoning, in: CLEF 2026 Working Notes, CEUR Workshop Proceedings, CEUR-WS.org, Jena, Germany, 2026. [4] X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdhery, D. Zhou, Self-consistency improves chain of thought reasoning in language models, 2023. URL: https://arxiv.org/abs/2203. 11171. arXiv:2203.11171. [5]H. Lightman, V. Kosaraju, Y. Burda, H. Edwards, B. Baker, T. Lee, J. Leike, J. Schulman, I. Sutskever, K. Cobbe, Let’s verify step by step, 2023. URL: https://arxiv.org/abs/2305.20050.arXiv:2305.20050. [6]C. Snell, J. Lee, K. Xu, A. Kumar, Scaling llm test-time compute optimally can be more effective than scaling model parameters, 2024. URL: https://arxiv.org/abs/2408.03314. arXiv:2408.03314. [7] M. Ahmadpour, A. Meighani, P. Taebi, O. Ghahroodi, A. Izadi, M. S. Baghshah, Limits and gains of test-time scaling in vision-language reasoning, 2025. URL: https://arxiv.org/abs/2512.11109. arXiv:2512.11109. [8]S. S. Ghosal, S. Chakraborty, A. Reddy, Y. Lu, M. Wang, D. Manocha, F. Huang, M. Ghavamzadeh, A. S. Bedi, Does thinking more always help? mirage of test-time scaling in reasoning models, 2025. URL: https://arxiv.org/abs/2506.04210. arXiv:2506.04210. [9]A. P. Gema, A. Hägele, R. Chen, A. Arditi, J. Goldman-Wetzler, K. Fraser-Taliente, H. Sleight, L. Petrini, J. Michael, B. Alex, P. Minervini, Y. Chen, J. Benton, E. Perez, Inverse scaling in test-time compute, 2025. URL: https://arxiv.org/abs/2507.14417. arXiv:2507.14417. [10]T. Zeng, S. Zhang, S. Wu, C. Classen, D. Chae, E. Ewer, M. Lee, H. Kim, W. Kang, J. Kunde, Y. Fan, J. Kim, H. I. Koo, K. Ramchandran, D. Papailiopoulos, K. Lee, Versaprm: Multi-domain process reward model via synthetic reasoning data, 2025. URL: https://arxiv.org/abs/2502.06737. arXiv:2502.06737. [11]D. B. Lee, S. Lee, S. Park, M. Kang, J. Baek, D. Kim, D. Wagner, J. Jin, H. Lee, T. Bocklet, J. Wang, J. Fu, S. J. Hwang, J. Bian, L. Song, Rethinking reward models for multi-domain test-time scaling, 2025. URL: https://arxiv.org/abs/2510.00492. arXiv:2510.00492. [12]B. Ong, T. D. Pala, V. Toh, W. C. Tjhi, S. Poria, Training vision-language process reward models for test-time scaling in multimodal reasoning: Key insights and lessons learned, 2025. URL: https://arxiv.org/abs/2509.23250. arXiv:2509.23250. [13]A. Panickssery, S. R. Bowman, S. Feng, Llm evaluators recognize and favor their own generations, 2024. URL: https://arxiv.org/abs/2404.13076. arXiv:2404.13076. [14]J. Zhao, R. Liu, K. Zhang, Z. Zhou, J. Gao, D. Li, J. Lyu, Z. Qian, B. Qi, X. Li, B. Zhou, Genprm: Scaling test-time compute of process reward models via generative reasoning, 2025. URL: https: //arxiv.org/abs/2504.00891. arXiv:2504.00891. [15]J. Zhang, Y. Yan, K. Zheng, X. Zou, S. Dai, X. Hu, Gm-prm: A generative multimodal process reward model for multimodal mathematical reasoning, 2025. URL: https://arxiv.org/abs/2508.04088. arXiv:2508.04088. [16]P. Kuang, X. Wang, W. Liu, J. Dong, K. Xu, Tim-prm: Verifying multimodal reasoning with tool-integrated prm, 2025. URL: https://arxiv.org/abs/2511.22998. arXiv:2511.22998. [17]R. Liu, J. Gao, J. Zhao, K. Zhang, X. Li, B. Qi, W. Ouyang, B. Zhou, Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling, 2025. URL: https://arxiv.org/abs/2502.06703. arXiv:2502.06703. [18] Y. Wu, Z. Sun, S. Li, S. Welleck, Y. Yang, Inference scaling laws: An empirical analysis of compute- optimal inference for problem-solving with language models, 2025. URL: https://arxiv.org/abs/ 2408.00724. arXiv:2408.00724. [19]A. Setlur, N. Rajaraman, S. Levine, A. Kumar, Scaling test-time compute without verification or rl is suboptimal, 2025. URL: https://arxiv.org/abs/2502.12118. arXiv:2502.12118. [20]J. Huang, X. Chen, S. Mishra, H. S. Zheng, A. W. Yu, X. Song, D. Zhou, Large language models cannot self-correct reasoning yet, 2024. URL: https://arxiv.org/abs/2310.01798.arXiv:2310.01798. [21]W. Wang, Z. Gao, L. Chen, Z. Chen, J. Zhu, X. Zhao, Y. Liu, Y. Cao, S. Ye, X. Zhu, L. Lu, H. Duan, Y. Qiao, J. Dai, W. Wang, Visualprm: An effective process reward model for multimodal reasoning, 2025. URL: https://arxiv.org/abs/2503.10291. arXiv:2503.10291. [22]S. Wang, Z. Liu, J. Wei, X. Yin, D. Li, E. Barsoum, Athena: Enhancing multimodal reason- ing with data-efficient process reward models, 2026. URL: https://arxiv.org/abs/2506.09532. arXiv:2506.09532. [23]Z. Chen, Y. Wang, T. Xiao, R. Zhou, X. Yang, W. Wang, Z. Sui, J. Wang, From mathematical reasoning to code: Generalization of process reward models in test-time scaling, 2025. URL: https://arxiv.org/abs/2506.00027. arXiv:2506.00027. [24] S. Ahmed, M. T. Younes, A. Moustafa, A. Allam, H. Moustafa, Msa at imageclef 2025 multimodal reasoning: Multilingual multimodal reasoning with ensemble vision language models, 2025. URL: https://arxiv.org/abs/2507.11114. arXiv:2507.11114. [25]V. T. Krazheva, D. Markova, D. I. Dimitrov, I. Koychev, P. Nakov, Contextdrift at imageclef 2025 multimodal reasoning: Evaluating vlms’ multimodal, multilingual and multidomain reasoning capabilities via thinking budget variations and textual augmentation, in: CLEF 2025 Working Notes, CEUR Workshop Proceedings Vol. 4038, 2025. [26] P. Hu, Z. Zhang, Q. Chang, S. Liu, J. Ma, J. Du, J. Zhang, Q. Liu, J. Gao, F. Ma, Q. Liu, Prm- bas: Enhancing multimodal reasoning through prm-guided beam annealing search, 2025. URL: https://arxiv.org/abs/2504.10222. arXiv:2504.10222. [27] W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, I. Stoica, Efficient memory management for large language model serving with pagedattention, 2023. URL: https://arxiv.org/abs/2309.06180. arXiv:2309.06180. [28]J. Wang, K. Q. Lin, J. Cheng, M. Z. Shou, Think or not? selective reasoning via reinforcement learn- ing for vision-language models, 2025. URL: https://arxiv.org/abs/2505.16854.arXiv:2505.16854. [29]X. Yue, Y. Ni, K. Zhang, T. Zheng, R. Liu, G. Zhang, S. Stevens, D. Jiang, W. Ren, Y. Sun, C. Wei, B. Yu, R. Yuan, R. Sun, M. Yin, B. Zheng, Z. Yang, Y. Liu, W. Huang, H. Sun, Y. Su, W. Chen, Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi, 2024. URL: https://arxiv.org/abs/2311.16502. arXiv:2311.16502. [30]X. Yue, T. Zheng, Y. Ni, Y. Wang, K. Zhang, S. Tong, Y. Sun, B. Yu, G. Zhang, H. Sun, Y. Su, W. Chen, G. Neubig, Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark, 2025. URL: https://arxiv.org/abs/2409.02813. arXiv:2409.02813. A. Chain-Agreement Analysis Table 5 reports per-tier accuracy on the Q3.5 SC-푁 =8(2k-tok) pool. The agreement tier is the number of chains (out of 8) that commit the same answer letter. Table 5 Chain-agreement tiers on Q3.5 SC-푁 =8(2k-tok,푛=4,651). Weak-agreement questions (≤4/8) have accuracy below chance for 4-option MCQ. Tier푛% of valAccuracy Unanimous (8/8)143730.9%90.2% Strong (6–7/8)136429.3%70.9% Moderate (5/8)63713.7%53.5% Weak (≤4/8)108923.4%41.7% All unparseable1242.7%— The skip rule used by both selectors (Section 5.6) passes through every question with at least5/8 agreement, i.e. the unanimous, strong, and moderate tiers (73.9%of validation). Only the weak and all-unparseable tiers (26.1%) are rescored. On this scored subset the base accuracy is∼37%, leaving limited room for any selector to improve on majority vote. B. Generative Critic Score Distribution The training-free generative critic scores each chain on three axes (step intent, visual alignment, logical soundness), each rated1–5, mapped to[0, 1]via(푠−1)/4, and mean-aggregated. Table 6 shows the score distribution across 5,710 scored chains from the Q2.5 SC-푁 =8 pool. Table 6 Generative critic score distribution. The modal score0.75(4/4/4axis pattern) accounts for40%of chains. The calibration slope is nearly flat: a perfect 1.0 is correct only 40% of the time. Score% of chainsAcc. at score 0.000.1%14.3% 0.255.7%27.4% 0.505.0%29.9% 0.75 (mode)39.7%33.6% 0.838.8%30.5% 0.925.3%33.0% 1.009.8%39.7% The correct-vs-incorrect chain score gap is negligible: correct chains average0.713vs.0.687for incorrect (gap0.026), with an identical median of0.75. In59%of questions, multiple chains are tied at the top score, forcing arbitrary tie-breaking. The pattern is consistent with the self-recognition bias reported by Panickssery et al.[13]: the policy rates its own outputs uniformly high regardless of correctness. C. Per-Language Chain Agreement Table 7 breaks down chain agreement by language for the Q2.5 SC-푁 =8pool, revealing the structural differences that drive per-language selector performance. Table 7 Per-language chain agreement on Q2.5 SC-푁 =8(푛=4,651). English has34%all-unparseable questions; Arabic and Hungarian have the largest weak-agreement tails. Lang푛 Unan. Strong Weak Unparse. Arabic517 19.0%29.2% 36.4%0.4% Bulgarian 400 40.8%34.5% 13.2%0.0% Chinese600 22.3%29.3% 30.2%0.2% Croatian585 39.1%29.1% 18.5%0.0% English347 14.1%10.1% 36.0%34.0% French224 43.3%35.3% 11.6%0.0% German279 30.5%39.4% 15.8%0.4% Hungarian 535 20.4%32.0% 30.1%0.4% Italian562 45.0%28.5% 15.3%0.0% Polish100 41.0%25.0% 20.0%0.0% Serbian502 35.7%29.7% 19.3%0.0% English performs worst under Q2.5: only14%unanimous,34%all-unparseable,36%weak agreement; 70%of English questions fall in the worst two tiers, explaining why English is the only language where SC regresses against zero-shot under Q2.5. Arabic and Hungarian have the largest weak-agreement tails (36%and30%), making them the languages where selectors have the most questions to rescore and the least signal to work with. Bulgarian, French, and Italian show consistent agreement distributions (39–45% unanimous, small weak tails, 80+ % SC accuracy). D. Per-Subject Accuracy Table 8 reports accuracy for the five weakest and five strongest subjects under the best Q3.5 and Q2.5 configurations respectively. Table 8 Per-subject accuracy for selected subjects (푛≥50). Bottom 5 and top 5 under Q3.5 SC-푁 =8(2k-tok). Q2.5 columns: zero-shot (ZS), CoT, SC-푁 =8. Q2.5Q3.5 Subject푛ZS CoTSCSC 2k Bottom 5 (Q3.5) Social100 28.039.049.043.0 Professional100 47.051.052.057.0 Islamic Studies83 21.7 45.8 * 33.7 ↓ 57.8 Agriculture100 47.048.0 42.0 ↓ 59.0 Tourism76 55.363.261.867.1 Top 5 (Q2.5 SC, for contrast) Psychology81 80.291.485.2 — Politics135 76.376.380.0— Sociology190 66.872.177.4— Biology594 65.370.272.6— Mathematics100 49.060.069.0— * CoT > SC under Q2.5 (regression). ↓ SC regresses vs. ZS. “–” = Q3.5 per-subject data unavailable for non-bottom-5 subjects at time of writing. E. PRM Scoring Mode (One-shot vs Per-step) On a DTR푁 =4,푀 =4pool under Qwen2.5-VL-7B (dev-200), per-step (segmented) PRM scoring is an exact null (63.0%→ 63.0%) while one-shot scoring shows a directional+3.0p gain (63.0%→ 66.0%) and is9.4×faster (Table 9). However, the±7p Wilson CIs at푛=200overlap fully, so the difference is not significant. One-shot evaluation gives the PRM maximal context to judge chain quality holistically, avoiding the per-step error compounding diagnosed in Section 5.2. Table 9 PRM scoring mode on a DTR Q2.5 pool (dev-200). CIs overlap; the one-shot gain is directional only. SelectorAcc.95% CIΔ 푠/q DTR majority (anchor)63.0% [56.1, 69.4]— PRM segmented (per-step) 63.0% [56.1, 69.4] +0.0 p 56.3 PRM flat (one-shot)66.0% [59.2, 72.2] +3.0 p6.0 F. Generative Critic Prompt The critic scores each chain independently on three axes, one inference call per axis. The prompt template is: You are a rigorous evaluator reviewing a reasoning chain written by a student answering a multiple-choice exam question. The question and its answer options are shown in the image. Rate the chain on ONE axis: axis_name. Axis definition: axis_definition Scale: 1 (severely deficient), 2 (weak), 3 (adequate), 4 (strong), 5 (excellent). Reasoning chain to evaluate: — chain_text — Output ONLY a JSON object on a single line with exactly two fields: "score": <integer 1-5>, "reason": "<brief one-sentence justification>" Do not output anything else. The three axes are: Step intent. Does each reasoning step address the question directly and build toward a commitment to one of the answer options? Penalise irrelevant tangents, repetition, self-doubt loops, or failure to commit to a letter. Visual alignment. Does the chain’s reasoning match what is actually visible in the image? Penalise hallucinated elements, misread values, missing visual evidence, or contradictions with the options rendered in the image. Logical soundness. Does the final answer follow from the premises via valid logical, mathematical, or scientific steps? Penalise unjustified leaps, false equivalences, arithmetic errors, or a mismatch between the concluded letter and the reasoning that preceded it. Per-axis scores are mapped to [0, 1] via (푠−1)/4 and mean-aggregated to a single chain score. Table 10 Per-language accuracy on full validation (푛=4,651). Q2.5 columns: ZS, CoT, SC-푁 =8(MMMU closer). Q3.5 columns: SC-푁 =8at 1024 and 2048 tokens.Δ 1k→2k is the gain from the budget increase. Down-arrow marks a regression. Q2.5Q3.5 SC Lang푛ZS CoTSC1k2k Δ 1k→2k Arabic51740.856.759.871.8 74.5+2.7 Bulgarian 40058.876.881.092.0 93.5+1.5 Chinese600 71.8 60.8 ↓ 69.066.2 74.7+8.5 Croatian58560.767.070.484.6 86.0+1.4 English34736.338.0 30.0 ↓ 56.2 73.5 +17.3 French22469.678.181.392.0 92.0+0.0 German279 61.768.169.284.9 86.4+1.5 Hungarian 53549.457.859.377.9 80.0+2.1 Italian562 64.469.069.481.9 83.3+1.4 Polish10047.051.052.055.0 57.0+2.0 Serbian502 56.666.370.583.3 84.7+1.4 G. Per-Language Accuracy H. Dev-200 Search Comparison Table 11 Search and verification on dev-200 under Q2.5. All rows use the same pre-MMMU closer for a matched comparison; the MMMU closer was adopted later and applied at val scale (Table 2). At푛=200the±7p Wilson intervals overlap, so dev-200 alone cannot rank the more expensive methods against SC. MethodAcc. 95% CICalls/q 푠/q SC-푁 =865.5% [58.6, 71.8]87.4 DTR 푁 =2,푀 =255.0% [48.1, 61.7]6 14.1 DTR 푁 =4,푀 =461.5% [54.6, 68.0]20 28.5 DTR-푁 =2 + PRM-BAS 60.5% [53.6, 67.0]13.4 42.3 I. Official Leaderboard Table 12 Official ImageCLEF 2026 Visual MCQ leaderboard (test split,푛=1,117): overall accuracy and per- language breakdown across the six test languages. Our submission (leaderboard handlespirosbax) ranks first overall and on every language. # ParticipantOverallENBGZHHRITSR 1 spirosbax (ours)84.06 85.60 89.09 78.07 87.72 90.74 87.04 2 DS@GT79.8685.2086.3667.2585.9687.0483.33 3 FAU71.0874.8069.0963.4573.6881.4875.93 4 zhaijinghe61.5067.6061.8253.5161.4059.2657.41 5 sjaini59.2769.2050.9144.1559.6572.2266.67 6 begyed57.7460.0066.3647.0866.6770.3764.81 7 linxiaocan55.6060.4054.5551.1750.8859.2642.59 8 pratikpriyanshu50.7657.2043.6441.8156.1455.5651.85 9 wether46.7354.6056.3631.8745.6148.1548.15