acceptodds
Under review as a conference paper at ICLR 2027

Self-Consistency via Marginal Sharpening

Abstract

Test-time compute has become a standard way to improve language-model reasoning without changing model weights, by allocating additional computation to sampling, search, or refinement before committing to an answer. A recent line of inference-time power-sampling methods pushes this idea by amplifying high-probability model outputs, typically doing so by sharpening complete generated sequences. We argue that this is the wrong object for reasoning models: a complete output entangles a latent reasoning trace with a final answer, even though many distinct traces may support the same answer. We introduce *Marginal Sharpening*, which instead sharpens the answer marginal induced by aggregating over latent reasoning traces. This gives an answer-level analogue of self-consistency, but as an inference target rather than a post-hoc voting rule. For integer sharpening strengths, the target has an exact multi-trace decomposition. We use this identity to derive a simple parallel auto-regressive approximation that samples multiple traces and generates one answer whose tokens are jointly supported across them, avoiding both complete-sequence search and exact string aggregation. Across math and code benchmarks, Marginal Sharpening improves over standard sampling and is particularly strong for code generation, where semantically equivalent solutions often differ in surface form. Compared with power sampling, Marginal Sharpening delivers stronger performance with substantially lower runtime, supporting answer-level sharpening as a more effective test-time objective.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Test Time Compute, Self Consistency, LLM Reasoning, Sharpening

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TN6gOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a test-time scaling method for reasoning models that sharpens the marginal probability of the final answer rather than the joint probability of the reasoning trace and answer. Using the identity that the K-th power of the answer marginal m(a|x) equals the expected product of answer probabilities across K independent reasoning traces, the authors build a sequential importance sampler. The algorithm samples K reasoning traces in parallel up to a delimiter token, decodes a single answer from the normalized product of the K next-token distributions, and weights P such particles by the product of their per-token normalizers. Because these weights are bounded by 1, the authors can derive a finite-particle total variation bound relative to the sharpened marginal target. They test the method on math and code benchmarks across six small-to-medium language models and compare against temperature sampling, MCMC power sampling, and majority voting.

Critical strengths

(1) The multi-trace decomposition gives a fast, one-pass alternative to MCMC power sampling that scales to long reasoning traces. Decoding from the per-token product of K trace-conditioned distributions makes the incremental weight rho_t <= 1, so the sampler runs as one batched autoregressive pass with no lookahead. On MATH-500 with Qwen3-4B at T=8192 this cuts 75th-pct runtime from ~100 min per problem for power sampling to under 8 and raises accuracy from 69% to 84%.

(2) The particle scaling ablation in Figure 2 shows that the importance weights add real signal beyond the proposal. At fixed K=4, uniform selection over P particles would match P=1, so the 4 to 9 point gains from P=1 to P=8 come directly from resampling by ∏tρt\prod_t \rho_t

Critical weaknesses

(1) The exact-string marginal target in Eq 4 does not solve surface-form fragmentation on code, and the actual decoding step is very close to prior logit-aggregation methods that are not compared. Raising m(a|x) to a power K still concentrates on the single most likely string rather than pooling equivalent programs. What helps on code is the P=1 proposal, which sums next-token log-probs across K traces like Integrative Decoding [1] and ThinkMerge [2]. In Fig 3, K=32, P=1 with no importance weights is within 1 point of the best setup on four of five tasks, and the TV bound in Prop 3.1 exceeds 1 at K=8,P=4 and K=16,P=2 unless one string already has >80% base mass.

(2) The code baselines in Table 1 collapse on MBPP+, and an unstated detail on trace truncation in Algorithm 1 may confound the comparison. On Qwen2.5-7B, majority voting over 32 samples drops to ~22% versus 35% for a single sample. Voting on raw code strings is also an odd baseline here since the prompt includes public tests that could be used to filter or cluster programs by execution [3].

Also, Algorithm 1 samples each trace until the delimiter and then runs a seperate answer phase, but never says what happens if a trace loops until T as in App. C.2. If the delimiter is force-appended, marginal sharpening gets a budget-forcing rescue [4] while looping baseline runs fail extraction.

ref: [1] Cheng et al., Integrative Decoding: Improving Factuality via Implicit Self-Consistency, ICLR 2025.

[2] Wang et al., Think in Parallel, Answer as One: Logit Averaging for Open-Ended Reasoning, ICLR 2026...

[3] Shi et al., Natural Language to Code Translation with Execution, EMNLP 2022.

[4] Muennighoff et al., s1: Simple test-time scaling, arXiv:2501.19393, 2025.

Questions for authors

(1) In Algorithm 1, what happens when a sampled reasoning trace hits the length limit T without emitting the answer delimiter? Do you force-append the delimiter and still run the answer decoding loop, and how often does this happen on MBPP+ for Qwen2.5?

(2) How concentrated are the normalized particle weights for K=8, P=4 and K=16, P=2, eg effective sample size or how often the top particle takes >90% of the weight? If one particle almost always dominates, how does the method compare to verifier-free best-of-N rerankers like self-certainty [5]?

(3) Are Qwen2.5 7B and 14B in Table 1 the base or Instruct checkpoints? Please also compare against Integrative Decoding [1] or ThinkMerge [2] at K=32 and an execution-filtered vote using the public assert statements in the prompt.

ref: [5] Kang, Zhao, Song, Scalable Best-of-N Selection for Large Language Models via Self-Certainty, arXiv:2502.18581, 2025.

3 comments

We thank the reviewer for recognizing the efficiency of the multi-trace construction and the evidence from particle scaling. Below we address the main questions:

1) Traces without an answer delimiter.

We do not force-append the answer delimiter. Instead, completions without a delimiter are excluded from the usable trace pool. Then each particle is constructed from K differnt traces that all contain a delimiter. In the MBPP+ source pools, 21.94% of Qwen2.5-7B completions and 8.55% of Qwen2.5-14B completions lacked the delimiter. If less then K traces contain a delimiter we reduced K as a fallback, but this only happened in 0.12% of the MBPP+ problems.

This matches the provided code, and we will clarify the procedure in the appendix.

2) P=1 proposal and ThinkMerge.

We would like to argue that the P=1 proposal is not the best method, as also mentioned in the critical strengths you identified. In Figure 3, K=8, P=4 is substantially better than K=32, P=1 on SxH. The importance weights are crucial for this improvement. At fixed K=4, Figure 2 shows gains of 4–9 percentage points as P increases from 1 to 8, highlighting the importance of the target and its importance weights beyond a strong proposal alone.

We compare MS with ThinkMerge on HumanEval+ and MBPP+ using Qwen2.5-14B and the same source pools of 32 responses. ThinkMerge aggregates up to K=32K=32 usable traces with answer temperature 0.6. Results are averaged over five seeds. MS outperforms ThinkMerge on both benchmarks with both (K,P)(K,P) configurations.

MethodHumanEval+MBPP+
ThinkMerge (K=32)70.2470.95
MS (K=8, P=4)73.9074.39
MS (K=16, P=2)73.0573.44

3) Weight concentration and self-certainty.

Indeed, the normalized particle weights are strongly concentrated. For (K,P)=(8,4)(K,P)=(8,4) and (16,2)(16,2) mean effective sample size ranges from 1.003 to 1.166, and the largest weight exceeds 90% in 78.3--99.7% of cases.

However, self-certainty did not outperform MS. We evaluated self-certainty on 32 saved base-model responses per MBPP+ problem, following the authors’ released scoring implementation. Self-certainty is computed under the base model, and the highest-scoring response is selected. Both methods use a nominal budget of 32 paths (N=32N=32 for self-certainty, KP=32KP=32 for MS) and a maximum generation length of 4096 tokens.

ModelSelf-certainty (N=32)MS (8,4)MS (16,2)
Qwen2.5-7B51.0666.3566.72
Qwen2.5-14B65.0874.3973.44

4) Coding baselines and public tests.

We compare execution-filtered voting on MBPP+ using the same 32 candidates as ordinary voting. For a fair comparison with access to the same public tests, we also evaluate two MS variants: filtering final particles by setting failing particles' weights to zero before renormalizing and selecting, and filtering the initial 32 responses before using their reasoning traces in MS. The latter uses up to KK usable passing traces per particle and ordinary weighted final selection. If no checks can be extracted or no usable candidate passes, each method falls back to its unfiltered procedure. As you can see from the table, MS with filter outperforms filtered voting. Moreover, original MS is still very competitive with filtered majority voting, despite not executing the public tests.

Model(K,P)Majority voteFiltered voteMSMS: final filterMS: initial filter
Qwen2.5-14B(8,4)50.2677.5175.4078.3178.04

Not executing public tests during decoding is a deliberate choice. Our goal is to compare the models’ underlying reasoning abilities, rather than augment them with tools or execution feedback. This allows us to assess whether our method improves performance solely by changing the inference target. None of the original baselines executes public tests during decoding either, ensuring a fair comparison.

5) Qwen checkpoints.

Both are base checkpoints: Qwen/Qwen2.5-7B and Qwen/Qwen2.5-14B. We will include these exact identifiers in the experimental setup.


We thank the reviewer for the constructive feedback, which helped us clarify our contribution and strengthen the experimental comparisons. We hope these clarifications and additional results address your concerns, and kindly ask you to reconsider the score in light of them.

Thanks for running extra experiments and being so transparent with the numbers. However, the new results reinforce my main concerns: 1. ESS of ~1.0 confirms that at P=2 or 4, the method isn't really sampling from the sharpened marginal distribution as the theory suggests, but simply doing argmax selection over a few ThinkMerge-style proposals. 2, on math it still trails simple Majority Vote, and on coding, once basic public-test filtering is applied, the large gap over Majority Vote mostly disappears. Given this mismatch between the theoretical framing and the practical results, I will keep my score

We thank the reviewer for this follow-up comment and respectfully disagree with this interpretation of our results.

1. Low ESS does not invalidate the target or the importance correction. The exact theoretical result concerns the multi-trace decomposition; the finite-particle sampler is explicitly an approximation. Concentrated weights suggest that this approximation may be coarse when using few particles. We agree that selection is then close to an argmax, but an argmax of the importance weights, which encode the target-to-proposal correction, rather than of the model probability. The correction is therefore decisive in determining which particle is selected. Crucially, Figure 2 shows that increasing P at fixed K improves the results, whereas uniform selection among particles would leave the output distribution unchanged. MS also outperforms both ThinkMerge and self-certainty in the requested comparisons. Together, these findings support the value of the proposed target and its importance correction beyond logit aggregation or alternative reranking, while showing that improving the particle approximation improves performance.

2. Execution feedback changes the setting being evaluated. Access to execution feedback can naturally level performance, but our goal is to improve inference without that feedback, the setting shared by all original baselines. On coding tasks, MS achieves the strongest results among the compared methods in this setting, including Power Sampling and temperature sampling. Adding execution feedback provides an additional capability; the resulting convergence in performance does not negate the improvement achieved without it. We already acknowledge majority voting’s strength on mathematics, and our theory makes no claim of universal superiority over voting, since both pursue closely related self-consistency objectives.

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

The paper proposed a test-time scaling method, named marginal sharpening. Instead of doing power sampling, they do powering with the marginal distribution on the answers. The authors carried out experiments to show the performance of their method comparing with some baselines: Majority Vote, Low Temperature and Power Sampling, across several models and math reasoning and code datasets.

Critical strengths

The empirical results: I think the idea is simple and effective in some datasets like MBPP+ and SxH. In long context reasoning. Its performance is quite incremental for Maths reasoning datasets. The theoretical part: I think it is clear and easy to follow.

Critical weaknesses

There are a lot of places in the Table 1 and 2 for maths datasets, the marginal sharpening obtains a weaker results than the majority vote. Meanwhile, the performance on code datasets is significantly improved. Do the authors have an explanation for it?

Most of the models in the experiments are quite outdated and small size. The maths testing datasets also quite easy. The authors could try their methods with recent and larger reasoning models and harder maths datasets, such as AIME24-25, OlympiadBench.

The experiment results are carried out for 32 paths. In practice, 32 is quite large number. I think the gap will be smaller if the authors try their methods with 8 or 16 paths.

Questions for authors

I like the simplicity of the idea, and it works for some datasets. Maybe the authors need to investigate more to show its effectiveness in that direction.

1 comment

We thank the reviewer for recognizing the method's simplicity, theoretical clarity, and gains on code and text-to-SQL. Below we address the main questions:

1) Math versus code.

We discuss this distinction in the third paragraph of the introduction and Section 5.2. Both MS and majority voting favor answers supported by multiple reasoning traces. Mathematical answers can often be normalized and compared, making voting an effective way to aggregate that support. In Python code generation, different programs can implement the same solution, splitting votes across distinct strings. MS generates a shared answer supported by multiple traces without requiring independently generated complete programs to match. This provides a natural explanation for the different behavior across the two domains.

2) Model and benchmark coverage.

Our choice of models follows prior power-sampling studies: Qwen2.5-7B appears in Power Sampling, Scalable Power Sampling, and Power-SMC; Phi-3.5-Mini-Instruct appears in Power Sampling and Power-SMC. We extend this setting in two directions: 1. We add Qwen2.5-14B, a larger model with twice the parameters 2. We evaluate three reasoning families with T=16384, four times the T=4096 limit used for the standard-model experiments. Thus, our six-model evaluation tests both increased model size and substantially longer reasoning.

The mathematics benchmarks remain challenging for these models. MATH-500 contains competition problems, while Minerva includes 272 university-level STEM problems, including differential equations and special relativity. Standard-model base accuracy on Minerva is only 12.1–19.6%, and even the strongest reasoning-model result reaches 52.6%. These benchmarks leave substantial headroom and provide a meaningful test of inference-time improvements.

We additionally evaluate Qwen2.5-14B on all 60 AIME24–25 problems with a 4096-token generation limit. The table reports mean accuracy over five seeds. MS improves over MCMC Power Sampling and that matches our results on the other math benchmarks.

MethodAIME24–25 accuracy (%)
Base sampling (T=1)5.67
Low temperature (T=0.25)11.00
Majority voting (32 samples)16.67
MS (K=8, P=4)11.67
MS (K=16, P=2)15.00‾\underline{15.00}
Power Sampling11.00

3) Smaller path budgets and computational cost.

We already present results for smaller total path budgets in Figure 2. Even with a smaller number of paths, MS improvement over the baseline remains substantial. For example, at K=8K=8, P=1P=1 or K=4K=4, P=2P=2, MS improves over the baseline by 15–20 percentage points on the coding benchmarks.

We now directly compare majority voting and MS at budgets of eight and sixteen paths on Python coding and SQL datasets using Qwen2.5-14B.

MethodHumanEval+MBPP+SxH
8 paths
Majority voting52.4457.1411.45
MS (K=4, P=2)69.7672.1734.94
16 paths
Majority voting53.0555.2916.27
MS (K=4, P=4)72.8073.2838.43

Even at eight and sixteen paths, MS retains substantial advantages over majority voting on code and SxH, with gains of roughly 15--20 percentage points on the coding benchmarks.


We thank the reviewer for the constructive questions, which helped us clarify the paper's contribution and strengthen its empirical support. We hope these clarifications and additional results address your concerns. We would appreciate your reconsideration of the score in light of them.

Reviewer_nBpUOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes power sampling similar to Karan and Du (2025) but only for the answer-part of a trace. The proposed algorithm relies on sequential importance sampling and often outperforms Karan and Du (2025) in performance and speed.

Critical strengths
  • The major contribution of the paper is to show that sharpening over the answer part of a generated trace improves upon sharpening the full trace (i.e., including the reasoning).
Critical weaknesses
  • The proposed algorithm is an SIS and feels rather similar to Azizi et al. (2026). As main contribution of the paper it feels not significant enough.
  • The paper only compares the method to the MCMC baseline (Karan and Du, 2025) but does not compare performance and speed to more recent work such as Ji et al. (2026) and Azizi et al. (2026). I expect Azizi et al. (2026) to be rather close in speed as the proposed algorithm.
  • The models in the main results are by today's standards rather small in their parameter counts (4B, 7B, 14B), old (Qwen2.5, Phi3.5), and few (3 models).
  • The models in the second part of the evaluation are small and the paper does not compare to power-sampling baselines.
  • Often, improvements seem to be insignificant.
Questions for authors
  • Zenn and Geiping (2026) show that often larger probability does not imply larger correctness, can you make any comment on how this behaves when considering only sharpening over the marginal? I.e., is the correlation here better?
  • Would it be interesting to compare your results to power self-consistency (i.e., self-consistency for samples from the power distribution) in Zenn and Geiping (2026)?
  • How does the performance and runtime compare to Power-SMC of Azizi et al., (2026) and Scalable Power Sampling of Ji et al, (2026)? Are additional improvements justifying (potentially increased?) runtimes?
  • Do the observed performance improvements translate to other (i.e., non-mathematical and non-code) benchmarks?
4 comments

We thank the reviewer for recognizing the value of sharpening the answer marginal and for suggesting additional baselines that strengthen our experimental evaluation. Below we address the main questions:

1) Inference target and correlation with correctness

Our main contribution is to sharpen the answer marginal distribution instead of a sequence distribution. The four-trace table in Section 2 (Preliminaries), with probabilities 0.20, 0.20, 0.20, and 0.40, shows that this target can favor a different answer from sequence-level sharpening. For integer KK, an exact multi-trace identity yields particles containing KK traces and one shared answer, with a tractable importance correction. SIS itself is established; the novelty lies in the target and its multi-trace construction.

To demonstrate this target's practical advantage, we evaluate the correlation between estimated marginal probability and correctness. We use the first 100 MBPP+ problems with Qwen2.5-14B. For each problem, we sampled 32 candidate solutions and an independent bank of 32 reasoning traces, estimating each answer's marginal probability by averaging its conditional probability over the bank. In the table below, we report the mean Pearson correlation between estimated marginal probability and correctness, along with importance-reweighted accuracy for K=2K=2 and K=4K=4. As you can see, the estimated marginal probability correlates more strongly with correctness than the joint sequence probability and yields higher reweighted accuracy. The mean paired correlation improvement is 0.084 (95% problem-bootstrap interval: [0.043, 0.129]).

ScoreMean Pearson rAccuracy, K=2Accuracy, K=4
Joint sequence probability0.04367.0566.35
Estimated answer marginal0.12876.5676.93

Accurately approximating marginal probabilities is indeed difficult. Table 1 from the main paper shows that MS improves accuracy on MBPP+ by 20 percentage points over the MCMC baseline, demonstrating that sharpening the answer marginal instead of the whole sequence can yield practical gains even with a small number of traces.

2) Comparisons with newer power samplers.

Our original comparison used one representative sequence-level power sampler, Karan and Du's MCMC method. MCMC Power Sampling, Scalable Power Sampling, and Power-SMC all aim to sample from the same full-sequence power distribution.

We now add Power-SMC, choosing it as the most latency-efficient of these alternatives. We use the official implementation on both coding datasets with Qwen2.5-14B, the largest model in our main comparison. Power-SMC uses 32 particles, exponent 4, and a 4096-token limit. The Power-SMC results are averaged over five seeds.

MethodHumanEval+MBPP+
MCMC Power Sampling67.4045.00
Power-SMC63.5453.54
MS (K=8, P=4)73.9074.39
MS (K=16, P=2)73.0573.44

Power-SMC does not close the gap to marginal sharpening: both MS configurations remain substantially more accurate on both datasets. The consistent advantage of MS supports the distinction between sharpening the joint sequence and sharpening the answer marginal.

With N=32 particles for Power-SMC and KP=32 trace-conditioned streams for MS, both methods support 32-way batch-parallel decoding. Thus, at this matched stream budget, their asymptotic forward-pass complexity is identical: O(NT).

3) Model coverage and long-generation evaluation.

Our models choice overlaps with prior power-sampling studies: Qwen2.5-7B appears in all three, and Phi-3.5-Mini-Instruct in both Karan and Du and Power-SMC. We extend this setting in two directions: 1. We add Qwen2.5-14B, a larger model with twice the parameters 2. We evaluate three reasoning families with T=16384, four times the T=4096 limit used for the standard-model experiments. Thus, our six-model evaluation tests both increased model size and substantially longer reasoning.

Figure 4 compares MS with the authors' official Power Sampling implementation on the reasoning model Qwen3-4B. MS is more accurate at T=2048T=2048, 40964096, and 81928192. At T=8192T=8192, it achieves 84% versus 69% accuracy, with a 75th-percentile runtime of 7m47s versus 1h39m per problem. This runtime growth made a full Power Sampling comparison at T=16384T=16384 computationally infeasible.

4) Size and reliability of the gains.

The results in Table 1 are averaged over five independent runs. Compared with MCMC Power Sampling, the best MS configuration improves accuracy by 29.4 percentage points on MBPP+ and 6.5 points on HumanEval+ for Qwen2.5-14B. For Qwen2.5-7B, the gains reach 41.2 points on MBPP+ and 7.9 points on SxH. Even the smaller gains often exceed the standard errors estimated across the five seeds.

5) Comparison with power self-consistency.

We thank the reviewer for this suggestion. We evaluated power self-consistency by voting over five independent MCMC Power Sampling outputs per problem. For Qwen2.5-14B, accuracies (Power Sampling -> Power self-consistency) are: MATH-500, 77.56 -> 80.60; Minerva, 28.82 -> 30.51; SxH, 36.14 -> 37.35; HumanEval+, 67.44 -> 70.73; and MBPP+, 45.03 -> 44.18. MS is more accurate on every benchmark except MATH-500, despite comparing one MS run per problem with five full Power Sampling runs. These results strengthen the case for sharpening the answer marginal, even when Power Sampling is augmented with majority voting.

6) Domains beyond mathematics and Python code.

We selected verifiable reasoning tasks to make them reproducible and avoid dependence on an LLM judge. The coding benchmarks are a natural choice: correctness is measured by executing the generated program against public test cases. SxH extends this to text-to-SQL, where correctness is measured by executing the generated SQL query against a public database. These results show that the gains extend beyond Python code generation.


We thank the reviewer for the constructive feedback. We hope that we have addressed all of the concerns and that our answers made the contributions of the paper clearer. If you don't have further questions, we would appreciate your reconsideration of the score in light of these clarifications and additional results.

Reviewer_nBpUOpenReview ↗

Response to the author's rebuttal

I would like to thank the authors for writing a detailed and thorough rebuttal addressing my points. At the same time I want to sincerely apologize for following up this late. I will take this for my scoring into account.

  1. Inference target and correlation with correctness

Interesting preliminary results; for the main paper, it might be interesting to include more datasets into this ablation as it seems to further strengthen the method.

Our original comparison used one representative sequence-level power sampler, Karan and Du's MCMC method. MCMC Power Sampling, Scalable Power Sampling, and Power-SMC all aim to sample from the same full-sequence power distribution.

I see your point but, in this case, one should not focus large parts of the paper on runtime improvements (which are obviously very strong when comparing to an MCMC method) but rather on the distribution that one is sampling from. Therefore, I feel like I have to insist on these results being computed for all the considered baselines in the paper to make a fair comparison. From the preliminary results on two of the datasets, it seems like this will turn out to be in the favor of the proposed method.

  1. Model coverage and long-generation evaluation.

I do see that these papers all test relatively small models and datasets (and I would have today the same critique about them as well). Still, for the method to be adopted broadly and usefully it might make sense to run (only the proposed method) for larger datasets and models. Since if the proposed method is indeed effective, one would see improvements on models and datasets where running the other methods might simply turn out to be infeasible.

  1. Comparison with power self-consistency.

I think for this evaluation, it would have been interesting to evaluate runtimes and a comparison to self-consistency more closely. I.e., does it make sense to run power self-consistency instead of self-consistency given the increased runtime.

  1. Domains beyond mathematics and Python code.

I asked for this baseline as other work (Zenn and Geiping, 2026) shows that the gains in mathematics and code typically do not transfer to other benchmark datasets. I would have been interested whether this is also the case if only the marginal was sharpened.

I would like to thank the authors again for their efforts and rebuttal. In light of the new results I will raise my score for the paper and weakly recommend acceptance. This is based on my assumption that the additional results (and missing baselines) will be included (alongside additional results asked for in this response) prominently in an updated version of the paper.

We sincerely thank the reviewer for the thoughtful follow-up, constructive suggestions, and positive reassessment of our paper. Most of the additional results presented in the rebuttal are already included in the appendix, and we confirm that the revised paper will be restructured to present all new results prominently in the main text.

We agree that including runtime comparisons is valuable. A detailed runtime comparison between Marginal Sharpening and MCMC Power Sampling is already provided in Figure 6. Majority voting has a similar runtime, as it uses the same budget, while Power-SMC takes slightly longer due to particle resampling, despite still underperforming Marginal Sharpening. We will detail these comparisons, together with runtimes for all other methods, in the revised version.

We also agree that testing our method on newer models is important. This is exactly what we aimed to capture in Table 2, where we evaluated three models (Qwen3, Ministral3, and SmolLM3) with long reasoning traces (T = 16K). These results already show the benefit of the proposed method in this setting.

We thank you again for the time and effort you have put into reading the paper, and we are glad that the discussion addressed your concerns and has led to a more positive assessment of our work.

Comments on the paper

1 comment

We thank the reviewers and the Area Chair for their time and constructive feedback. The discussion has helped us clarify our contribution and strengthen the experimental evaluation. We summarize these points below for ease of reference.

Our main contribution is to motivate sharpening the answer marginal for reasoning: an answer can receive substantial support across many traces even when no individual trajectory dominates. Starting from this objective, we derive a practical sampler through an exact multi-trace decomposition, with finite-particle convergence guarantees (Proposition 3.1) and a locally variance-optimal proposal (Proposition A.1). This connects a principled inference target to a tractable sampling algorithm.

1. Model coverage and computational budgets. Our evaluation spans six models of sizes 3B, 4B, 7B, and 14B, five benchmarks, and generation lengths up to 16K tokens (Tables 1–2). Figure 4 shows improved accuracy and order-of-magnitude speedups over MCMC Power Sampling in long-generation settings. During the rebuttal, we additionally evaluated AIME24–25 and compared methods at budgets of eight and sixteen paths, where substantial gains persist.

2. Target versus sampler. The new Power-SMC comparison (Table 3) shows substantial gains at a matched budget of 32 decoding streams: 73.90% versus 63.54% on HumanEval+ and 74.39% versus 53.54% on MBPP+. These results support the advantage of changing the inference target beyond the choice of sequence-level sampler.

3. Comparisons with alternative heuristics. Following the reviewers’ suggestions, the revision adds power self-consistency (Table 6), ThinkMerge (Table 7), and self-certainty (Table 8). MS outperforms ThinkMerge on both Python coding benchmarks and self-certainty on MBPP+ for both model sizes. A single MS run also outperforms voting over five independent Power Sampling outputs on four of five benchmarks. These results support benefits beyond logit aggregation, confidence-based selection, and voting over power samples.

4. Contribution of the importance correction. Figure 2 shows gains of 4–9 percentage points when increasing PP at fixed KK, supporting the correction’s contribution beyond the proposal alone. We reported the requested weight-concentration statistics and acknowledged that the approximation can be coarse with few particles. Even when selection approaches an argmax, it is an argmax of the target-to-proposal importance correction, which determines the selected answer.


We acknowledge majority voting’s strength on mathematics. Both majority voting and Marginal Sharpening follow the principle of self-consistency, and our goal is to demonstrate its advantages over sequence-level Power Sampling. Our contribution is to formalize this principle through answer-marginal sharpening and derive a practical sampler with demonstrated advantages, particularly on coding.

We hope the additional experiments and clarifications provide helpful context for the final assessment. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.