acceptodds
Under review as a conference paper at ICLR 2027

DIVTHINK: Parallel Thinking via Diversified Chain-of-Thought Sampling

Abstract

Test-time reasoning has become a central paradigm for scaling Large Language Models (LLMs), but parallel Chain-of-Thought (CoT) sampling often wastes computation on redundant trajectories that repeat similar intermediate steps and converge to the same final answer. Existing alternatives either refine trajectories sequentially, as in MCMC power sampling, or require costly post-training, as in GRPO. To bridge this gap, we propose DIVTHINK, a training-free decoding method that coordinates parallel CoT trajectories through lightweight repulsive logit adjustments. At each step, DIVTHINK downweights tokens already preferred by peer trajectories, encouraging complementary reasoning paths while preserving the base model's local preferences and the throughput advantage of parallel generation. Across mathematical reasoning, code generation, and scientific question-answering benchmarks, DIVTHINK improves accuracy by 7.6 points over Base@8 and 4.0 points over Low-temperature@8 on average, with gains up to 22 points, while using far fewer generated tokens than MCMC Power Sampling and avoiding the rollout and optimization cost of GRPO-style post-training.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Parallel Thinking, Test-time Reasoning, Large Language Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NH4WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents DivThink, a test-time method that applies logit repulsion to common vocabulary across parallel trajectories, in order to increase diversity across the trajectories whose token distributions are primarily guided by the model.

Critical strengths
  • DivThink provides a test-time parallel sampling mechanism where the samples collaborate to be dissimilar to one another in every token step via logit repulsion.
  • The authors showed that controlling for eight trajectories, DivThink improves accuracy over base sampling and low-temperature sampling, and also reduces token consumption compared to MCMC Power Sampling though the benefit is small for code generation (HumanEval).
  • A good number of four benchmarks and four models were used in the main experiments.
Critical weaknesses
  • Comparison of DivThink with low-temperature is weak as low temperature favors greedy-like generation that therefore likely converges to similar reasoning modes. For example, low temperature T=0.25 on L796 is likely to be mode-seeking.
  • L364: Whilst high temperature sampling produces less-usable results, did you tune temperature between 0 and 2? There should be a comparison between DivThink and a tuned temperature; as an example, in Tables 3 and 4 relating to GPQA Diamond, when we compare Base@8 at temperature 0.50 to DivThink@8, there is no clear winner on all four metrics, more so in the absence of standard deviation.
  • Besides Base and Low-temperature sampling, are there other decoding methods targeted at improving exploration and coverage of reasoning paths?
  • L395-400: Are there analysis of generated trajectories to support the claim "... with repulsion being most useful when independent trajectories remain redundant; when the base candidate set is already strong, additional diversification offers less headroom and can slightly reduce local quality" that seems to be used as an explanation for the result "... Base@8 remains stronger on HumanEval"
  • Even with token budget controlled, how much more compute (or memory) overhead do the logit repulsions introduce? Are there efficiency analysis?
  • Table 1 and 2: What are the standard deviation of the accuracies?
  • Since DivThink diversifies parallel generations, what are the effects of DivThink on tasks such as logical reasoning and open tasks like creative writing?
  • L858: Why is the lambda value of 0.4 selected when it produces the worst accuracy changes in Figure 5 (Right)?
  • It is unclear how DivThink affects logits of various distribution profiles. Could you provide synthetic examples of token distribution profiles adjusted with logit repulsion?
  • The references contain many errors in years and page numbers, and at least an author name.
Questions for authors

Please refer to questions, clarifications and feedback in weaknesses.

3 comments

We thank the reviewer for the detailed feedback.

Q1. Comparison with temperature-tuned independent sampling

We selected temperatures separately for independent sampling and DIVTHINK using 50 MATH training problems and 30 GPQA-main problems outside Diamond. Both methods used the same validation grid (0.1, 0.25, 0.5, 0.75, 1.0, 1.25, 1.5), prompts, K=8K=8, 3,072-token cap per trajectory, answer extraction, and majority voting. DIVTHINK used λ=0.4\lambda=0.4. Each row reports mean majority-vote accuracy over five decoding seeds.

Model / benchmarkSelected TT (Independent / DIVTHINK)Independent samplingDIVTHINKGain (pp)
Qwen2.5-7B / MATH5000.25 / 0.7571.08%72.40%+1.32
Qwen2.5-7B / GPQA Diamond0.50 / 0.5035.35%37.07%+1.72
Qwen3-14B / MATH5000.25 / 0.2587.32%87.68%+0.36
Qwen3-14B / GPQA Diamond0.75 / 0.2552.42%53.74%+1.32

DIVTHINK achieved higher mean accuracy in all four comparisons after both methods were tuned on the same validation grid. In an additional Qwen2.5-7B GPQA validation check, both methods performed poorly at T=1.75T=1.75 and T=2.0T=2.0. Higher temperatures flatten the next-token distribution, making low-probability continuations more likely and potentially disrupting multi-step reasoning.

Q2. Multi-seed accuracy and uncertainty

We report mean majority-vote accuracy and 95% confidence intervals for the gain over independent sampling.

Model / benchmarkPaired seedsIndependent (%)DIVTHINK (%)Gain [95% CI], pp
Qwen2.5-7B / MATH500571.0872.40+1.32 [+0.18, +2.46]
Qwen2.5-7B / GPQA Diamond535.3537.07+1.72 [+0.24, +3.20]
Qwen2.5-14B / MATH500378.4779.20+0.73 [+0.23, +1.23]
Qwen2.5-14B / GPQA Diamond341.2541.92+0.67 [+0.17, +1.17]

DIVTHINK improved mean accuracy in all four comparisons, and the reported intervals for all four gains lie above zero.

Q3. Stronger inference-time baselines

On Qwen2.5-7B base, we applied each aggregation method to the same eight saved candidates per generator. The main comparisons average five decoding seeds, and the matched-temperature DeepConf control averages three paired seeds.

Benchmark / aggregationIndependent samplingDIVTHINK
MATH500 / Self-Consistency (majority voting)71.08%72.40%
MATH500 / Self-Certainty70.60%72.20%
MATH500 / Offline DeepConf75.84%76.40%
GPQA Diamond / Self-Consistency (majority voting)35.35%37.07%
GPQA Diamond / Self-Certainty35.45%35.86%
GPQA Diamond / Offline DeepConf35.56%36.87%

With the aggregation method held fixed, DIVTHINK achieved higher mean accuracy on both benchmarks. The highest observed MATH500 accuracy came from combining DIVTHINK with offline DeepConf. These results cover DeepConf's offline K=8K=8 component.

For prompt-based diversification, DIVTHINK@8 achieved 66.7% and the official VS_STANDARD prompt achieved 62.5% majority-vote accuracy on the same MATH subset, using Qwen3-14B at T=0.7T=0.7 under the same generation token cap.

Q4. Redundancy and headroom

DIVTHINK improved mean HumanEval pass@8 with Qwen2.5-7B under matched decoding settings for both raw completion and chat-formatted prompts. At the same candidate count, it solved more problems on average in both settings, including the chat setting where independent sampling already performed well.

In the matched Qwen3-14B evaluation, DIVTHINK reduced exact duplicates and shortened shared prefixes, alongside a modest improvement in pass@k. The trajectory metrics show less overlap among generated programs, while the pass@k gain shows that more problems had at least one test-passing candidate at the same candidate count. This connects the trajectory analysis to the outcome that matters for HumanEval: coverage of correct solutions. The gains from a strong independent baseline show that useful coverage headroom can remain even when baseline pass@k is already high.

Q5. Compute and memory overhead

On an A100 with Qwen2.5-7B, both arms decoded 256 steps from the same 64-token prefix, with identical forced tokens and forward-pass counts. Confidence and trace collection were disabled.

TrajectoriesBase median timeDIVTHINK median timePeak allocated memory, both arms
84.7426 s4.7770 s~14.39 GiB
164.9733 s5.0038 s~14.54 GiB

In this controlled decoding benchmark, DIVTHINK added less than 1% runtime at both trajectory counts, with similar peak allocated memory.

Q6. Extension to logical reasoning and creative writing

We evaluated DIVTHINK and independent sampling on propositional reasoning and constrained creative writing. Both evaluations used K=8K=8 and three decoding seeds, with matched generation settings within each task. On 36 propositional reasoning problems, the two methods achieved similar mean accuracy with Qwen2.5-7B. The labels were verified by exhaustive truth-table enumeration.

For creative writing, we used Qwen3-14B and measured compliance with explicit prompt requirements. The table reports aggregate compliance across prompts and decoding seeds.

Creative-writing metricIndependent samplingDIVTHINK
Required-object compliance31.94%34.03%
Word-budget compliance94.44%95.14%

DIVTHINK improved adherence to both writing constraints in this exploratory evaluation. We will include the results and protocols for both tasks in the revised paper.

Q7. Choice of λ\lambda and interpretation of Figure 5

The right panel of Figure 5 illustrates that sensitivity to λ\lambda can vary across tasks even for the same backbone. This pattern is consistent with the intended trade-off: moderate repulsion discourages continuations already favored by peer trajectories while preserving the base model’s local preferences. Excessive repulsion can over-penalize otherwise plausible continuations. Because the base logits and peer preferences can differ across models and tasks, the coefficient that best balances these effects can also vary.

Q8. Peer reweighting

The update lowers a token’s odds relative to alternatives that peers favor less. When trajectories share a sharp peak, repulsion can reduce the dominant token’s probability while preserving its rank. Uniform peer probabilities shift all logits by the same amount and leave the sampling distribution unchanged. A lone active trajectory receives no peer penalty.

Consider a decoding trace for a problem that gives x−y=9x-y=9 and x2+y2=153x^2+y^2=153 and asks for xyxy. At the shared prefix If y = -12, then x =, the sole active peer strongly favors -. Using the same random draw at decoding step 384 (zero-indexed), independent sampling selects -, while DIVTHINK selects y. The resulting substitutions, x = -12 + 9 and x = y + 9 = -12 + 9, are both valid.

The continuations subsequently handle the two solution branches differently. Independent sampling combines values from different branches and reports xy=−36xy=-36. DIVTHINK preserves the correct (x,y)(x,y) pairings, (−3,−12)(-3,-12) and (12,3)(12,3), and obtains xy=36xy=36. The shared-prefix comparison isolates the effect of peer reweighting on token selection at that step. The two continuations show how the resulting divergence played out in this example.

We will include the new experiments, probability examples, decoding trace, and corrections to the references in the revised manuscript.

I thank the authors for performing additional experiments and analyses that address some of my questions.

Could you provide the paired seed results for DIVTHINK vs temperature sampling on Qwen3-14B (Q1), and DIVTHINK vs temperature sampling for each aggregation strategy (Q3)?

While temperature-tuned DIVTHINK improved adherence to the two creative writing constraints, is there preference evaluation on paired prompts across the two methods? How many creative writing prompts were used? What are the paired seed differences between the two methods on the two creative writing constraints and the preference?

Could you quantify these? "DIVTHINK reduced exact duplicates and shortened shared prefixes"

Are you able to compare DIVTHINK on a larger model ~50-100B in parameter size of another model family?

On the balance of the new results and the above, I currently maintain my evaluation.

Reviewer_m46xOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DivThink is a training-free decoding method that coordinates parallel Chain-of-Thought trajectories through repulsive logit adjustments. It downweights tokens preferred by peer trajectories to reduce redundant sampling and encourage complementary reasoning paths. The paper evaluates the method on mathematical reasoning, code generation, and scientific question answering.

Critical strengths
  1. A compelling motivation and a direct, intuitive method. Redundant exploration is a meaningful concern in parallel sampling, and the proposed decoding adjustment provides a straightforward way to address it without additional training. I find both the motivation and the method easy to understand.

  2. Particularly clear and thoughtfully designed figures. I especially appreciate Figures 2 and 3: they communicate the redundancy problem and the proposed approach in a simple, direct, and accessible way. The care invested in these visualizations substantially improves the presentation.

Critical weaknesses
  1. The comparison with competitive existing methods is too limited. Although the main table includes standard sampling baselines and GRPO, the comparison with more advanced training-free inference-time methods is largely limited to MCMC Power Sampling. There is considerable related work on improving pass@k, including pass@8, and the paper should identify the relevant contemporary methods and compare against them under comparable evaluation settings. Establishing this context is the authors’ responsibility; readers should not have to collect results from other papers themselves to assess the method’s competitiveness. The current comparisons leave an important gap in the empirical evaluation.

  2. The evidence does not sufficiently establish the proposed connection between token-level repulsion and richer semantic reasoning paths. The method directly controls token-level redundancy, whereas the interpretation that it produces complementary reasoning paths relies largely on indirect metrics and selected examples. Some mechanism analyses also use only a single random seed and small samples. These results support a promising practical approach to reducing wasted parallel sampling, but the mechanism requires further validation.

The code case study in Figure 7 is particularly incomplete in this respect. It shows that independent sampling repeats an incorrect coordinate transformation, while DivThink produces a correct alternative, but it does not show how the intervention leads to that alternative. Which decoding step was affected? Which token received a penalty, which alternative token was sampled, and how did the subsequent continuation develop into a different solution? Presenting only the final programs does little to resolve this mechanism question. I want to see whether the intermediate decoding behavior actually follows the proposed explanation. As presented, the case study does not rule out the possibility that the successful alternative simply reflects a fortunate sampling outcome.

Overall, the paper only narrowly warrants my current rating. If these concerns are not adequately addressed in the rebuttal, I would lower my rating.

Questions for authors
  1. Why are comparisons with contemporary inference-time methods largely limited to MCMC Power Sampling? Which other methods targeting pass@k improvements are most relevant, and what justifies their omission from the evaluation?

  2. What evidence distinguishes increased semantic diversity in reasoning from reduced token-level overlap? Given the single-seed and small-sample mechanism analyses, which conclusions do the authors consider established, and which remain hypotheses?

  3. For Figure 7, can the authors use available decoding traces to show where the repulsive adjustment affected token selection, what alternative was sampled, and how the continuation developed into the correct program? What evidence connects this outcome to the proposed mechanism rather than a fortunate sampling outcome?

2 comments

It is a pity that I didn't have a chance to have a deeper communication with the authors. Therefore, given the weaknesses I mentioned above, I decided to lower my rating to weak rejection.

We sincerely apologize for the delay in our response and we have just completed some additional experiments. We appreciate the opportunity to address your concerns and would be grateful if you could take the new HumanEval comparisons and decoding analyses below into account when revisiting your assessment.

Q1. Inference-time baselines on HumanEval We chose MCMC Power Sampling as a key reference because it uses a frozen model and spends inference compute on iterative trajectory resampling. This provides a contrast with DIVTHINK, which coordinates parallel candidates during decoding. The original evaluation focused on this comparison and on independent sampling.

To address the gap in code-generation baselines, we added IdeaSearch and PlanSearch on the same Qwen2.5-7B backbone and all 164 HumanEval problems.

MethodPass@8Tokens/problemGPU-seconds/problem
Independent sampling (T=0.9T=0.9)92.07%2,76821.6
DIVTHINK (T=0.9T=0.9)93.29%2,75421.4
IdeaSearch87.80%10,209175.4
PlanSearch (reduced, L1/O3)85.37%40,544886.6

DIVTHINK achieved the highest observed pass@8 on the common seed, with token use and runtime close to independent sampling. Across three completed seeds, mean pass@8 was 91.67% for DIVTHINK and 91.06% for independent sampling. The search-baseline comparisons remain preliminary.

Q2. Semantic reasoning paths and token-level variation

DIVTHINK coordinates local sampling decisions across trajectories. This can improve voting by increasing the representation of a valid reasoning strategy already present in the candidate set.

In a matched-temperature MATH500 analysis with Qwen2.5-7B, K=8K=8, and three decoding seeds, we examined a geometry problem involving an 8–15–17 triangle. An independent trajectory correctly computes the total area as 60 but assumes the altitude divides it into equal areas, yielding 30. A DIVTHINK trajectory explicitly computes the projection AD=64/17AD=64/17 and the altitude CD=120/17CD=120/17, yielding the correct subtriangle area 3840/2893840/289. The two trajectories use different geometric arguments.

AnswerIndependent samplingDIVTHINK
Incorrect area, 3042
Correct area, 3840/28924
Other answers, combined22

The correct answer rose from two votes to four and became the unique plurality winner. The projection strategy already appeared under independent sampling, so this selected example shows how greater representation of an existing valid strategy can change the final prediction.

To assess semantic coverage separately from lexical overlap, we ran a method-blinded automated strategy annotation on 41 usable paired cases across two decoding seeds. The mean strategy-cluster counts were 3.463 for independent sampling and 3.146 for DIVTHINK. Whether DIVTHINK increases the number of distinct strategies overall remains open. We will report strategy counts, strategy frequencies, lexical overlap, and correctness separately in the revision.

Q3. From a token-level intervention to the subsequent reasoning

We provide a separate algebra trace that shows the requested decoding behavior step by step. It uses Qwen2.5-7B with T=0.75T=0.75, K=8K=8, and seed 29, comparing independent sampling with DivThink.

The problem gives x−y=9x-y=9 and x2+y2=153x^2+y^2=153 and asks for xyxy. The two target trajectories share the prefix If y = -12, then x = and first diverge at zero-based decoding step 384. At that step, two trajectories remain active, and the sole active peer strongly favors -.

Candidate tokenActive-peer probabilityEffect of the adjustment on sampling probabilitySelected in
-0.996094DecreasesIndependent sampling
y0IncreasesDIVTHINK

At the shared prefix, replaying the same random draw with the logit adjustment changes the sampled token from - to y. The resulting continuations handle the solution branches differently.

TrajectoryImmediate continuationSubsequent reasoning and answer
Independent samplingx = -12 + 9Finds x=−3x=-3 or 1212, then combines values from different branches and reports (−3)×12=−36(-3)\times12=-36.
DIVTHINKx = y + 9 = -12 + 9Preserves the solution pairs (−3,−12)(-3,-12) and (12,3)(12,3). Both yield xy=36xy=36.

Both immediate substitutions are valid. The replay isolates the effect of the adjustment on token selection at step 384. The later handling of the solution branches shows how that local divergence developed into a substantive reasoning difference in this trace. Estimating its effect on the probability of a correct answer would require repeated continuations from the shared prefix.

Reviewer_gsFFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DivThink, a training-free decoding method for parallel Chain-of-Thought generation that aims to reduce redundancy among concurrently sampled trajectories. At each decoding step, every trajectory’s logits are adjusted by subtracting the next-token distribution of its active peers. Empirically, the final answers are aggregated by majority voting (MATH500, AMC23, GPQA) or evaluated by pass@k (HumanEval). Across multiple model backbones, DivThink@8 demonstrates consistent gains over Base@8 and Low-temperature@8, while using fewer generated tokens and less wall-clock time than MCMC Power Sampling.

Critical strengths
  • The paper targets a practically relevant and well-motivated problem: independent parallel samples often collapse onto the same dominant (and sometimes wrong) reasoning mode, which limits the value of additional test-time compute. The method is simple, training-free, requires no extra forward passes, and is compatible with batched decoding. This makes it easy to adopt as a drop-in replacement for standard parallel sampling, and they also show the coordination step adds little overhead in most settings.

  • The evaluation spans mathematical reasoning, code generation, and scientific QA across several backbones, and compares against both inference-time and post-training baselines.

Critical weaknesses
  • The paper presents DivThink as a remedy for redundancy among parallel samples. A natural baseline would therefore be other notable training-free methods that directly target output diversity, such as Verbalized Sampling (VS) [1]. Because VS operates purely at the prompt level, requiring neither logit access nor synchronized batched decoding, it offers a cheap, model-agnostic comparison. Although VS was evaluated mainly on open-ended tasks such as creative writing, its main premise, that mode collapse limits output diversity, closely mirrors DivThink’s motivation. Including it would make the empirical comparison more comprehensive and would show whether logit-level coordination offers advantages over eliciting diversity through prompting.

  • Beyond diversity-oriented generation, a growing line of training-free parallel test-time methods improves on Self-Consistency, such as DeepConf [2] and Self-Certainty [3], could serve as additional baselines. These methods could be both competitors and potential complements. Comparing them as well as discussions showing whether DivThink can be combined with them, would help place DivThink relative to the current state of parallel test-time scaling.

  • DivThink does not appear to scale well with model capability/size. On Qwen3-14B, the only relatively larger backbone evaluated (Table 2), DivThink and the baselines perform within a few points of each other. These margins are especially small given the size of the evaluation sets; AMC23, for instance, contains only 40 problems. Moreover, the Qwen3-14B results appear to come from a single run, with no confidence intervals or significance tests, so these differences cannot be reliably distinguished from decoding noise. Taken together, the results suggest diminishing returns where test-time reasoning may matter most: enabling strong “deep think” capabilities in larger models.

--

[1] Zhang, Jiayi, et al. "Verbalized sampling: How to mitigate mode collapse and unlock llm diversity." arXiv preprint arXiv:2510.01171 (2025).

[2] Fu, Yichao, et al. "Deep think with confidence." International Conference on Learning Representations. Vol. 2026. 2026.

[3] Kang, Zhewei, Xuandong Zhao, and Dawn Song. "Scalable best-of-n selection for large language models via self-certainty." Advances in neural information processing systems 38 (2026): 19720-19745.

Questions for authors

Please refer to the weakness part.

1 comment

We sincerely thank the reviewer for the thoughtful feedback. We clarify the main points and provide additional results below.

Q1. Comparison with Verbalized Sampling

Verbalized Sampling (VS) prompts the model to generate multiple responses and their probabilities. DIVTHINK promotes diversity directly during decoding by downweighting tokens favored by peer trajectories. It preserves the original task prompt and answer format without adding model forward passes beyond standard parallel sampling.

We evaluated both methods on the same fixed 48-problem MATH subset with Qwen3-14B at T=0.7T=0.7, using the official VS_STANDARD prompt for VS.

MethodMaximum generation token budgetMajority-vote accuracy
DIVTHINK@88 × 3,072 per problem66.7%
VS24,576 per response group62.5%

Under the same generation token cap, DIVTHINK@8 outperformed VS by 4.2 percentage points on this subset. DIVTHINK directly discourages redundant continuations across trajectories before majority voting.

Q2. Comparison with Self-Certainty and DeepConf

We used Qwen2.5-7B base with eight trajectories per problem and a cap of 3,072 new tokens per trajectory. Both generators shared the task prompts, answer extraction, and evaluation. The independent-sampling baseline in this comparison uses validation-tuned temperatures. The selected temperatures for independent sampling and DIVTHINK were 0.25 and 0.75 on MATH500 and 0.5 for both on GPQA Diamond. We held each generator’s temperature fixed across aggregation methods.

We applied all three aggregation methods to the same eight saved candidates from each generator. Self-Certainty uses full-vocabulary KL scoring, and DeepConf uses the official offline minimum-window confidence-weighted vote. Both use unmodified backbone probabilities from generation without extra forward passes. The main results average five decoding seeds. The matched-temperature control averages three paired seeds. The DeepConf comparison covers its offline component at K=8K=8.

Benchmark / aggregationIndependent samplingDIVTHINK
MATH500 / Majority71.08%72.40%
MATH500 / Self-Certainty70.60%72.20%
MATH500 / DeepConf offline75.84%76.40%
GPQA Diamond / Majority35.35%37.07%
GPQA Diamond / Self-Certainty35.45%35.86%
GPQA Diamond / DeepConf offline35.56%36.87%

On MATH500, independent sampling with offline DeepConf outperformed DIVTHINK with majority voting. With either confidence-based selector held fixed, DIVTHINK achieved higher mean accuracy than independent sampling on both benchmarks. The matched-temperature DeepConf control on MATH500 showed the same direction.

Q3. Performance at 14B and 32B

To assess performance across decoding seeds and model sizes within one family, we compared DIVTHINK with independent sampling on Qwen2.5-14B and Qwen2.5-32B base. Both methods used T=0.25T=0.25. Each row reports mean majority-vote accuracy over three complete paired decoding seeds.

ModelBenchmarkIndependent samplingDIVTHINKGain (pp)95% CI for gain (pp)
Qwen2.5-14BMATH50078.47%79.20%+0.73[+0.23, +1.23]
Qwen2.5-14BGPQA Diamond41.25%41.92%+0.67[+0.17, +1.17]
Qwen2.5-32BMATH50077.80%78.60%+0.80[+0.30, +1.30]
Qwen2.5-32BGPQA Diamond38.72%40.74%+2.02[+0.53, +3.51]

DIVTHINK improved mean majority-vote accuracy on both benchmarks at both model sizes, and all four reported 95% confidence intervals exclude zero. The MATH500 gains were similar at 14B and 32B, and the largest observed gain was on GPQA Diamond at 32B.

We will include the comparison with Verbalized Sampling and its analysis in the revised paper, together with the Self-Certainty and DeepConf comparisons and the larger model experiments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.