acceptodds
Under review as a conference paper at ICLR 2027

When Can Scores Improve a Vote?

Abstract

A language model can answer a problem by generating several solutions and choosing the most common final answer. We characterize exactly when scores make voting errors decrease faster as more solutions are sampled. The key information is how scores distinguish the correct answer from its leading wrong rivals. We study independent samples for one fixed problem with finitely many possible answers, where errors occur but the correct answer is more likely than any individual wrong answer. Some fixed rule with positive, bounded score weights makes errors decrease faster than ordinary voting at an exponential rate if and only if the correct answer’s score distribution cannot be reproduced by mixing those of the most frequent wrong answers. Two perfectly calibrated examples (whose scores match correctness frequencies) have identical score–correctness statistics and answer frequencies, yet differ in this capacity for improvement. Complementing this large-sample characterization, we identify the weights that preserve the long-run winner under calibration and give conditions for improving count ties safely at finite budgets. On a shared cohort of 252 competition-mathematics problems, score weighting improves eight-candidate accuracy by 2.78 percentage points for Math-7B (95% interval [1.06, 4.63]) and 2.91 for Phi-4 ([1.19, 4.76]). Count ties account for most of both gains. Further experiments evaluate scoring cost, transfer, and score-based diagnostics.

Primary area:
foundation or frontier models, including LLMs
Keywords:
process reward models, voting, plurality voting, scoring votes, LLM math solving

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_2hfpOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies when score-based weighting improves plurality voting over sampled language-model responses. Its main theoretical result characterizes when a bounded, strictly positive score-only weighting rule can improve the asymptotic error probability bound, under the assumptions of i.i.d. sampling and a uniquely correct population mode. The criterion depends on whether the correct answer’s conditional score distribution lies outside the convex hull of those of the most frequent incorrect answers. The paper also characterizes weighting rules that universally preserve the correct population mode or prevent accuracy degradation under exact calibration, and provides guarantees for score-based tie-breaking and harmful winner reversals. Validation combines constructed distributions, numerical checks, and AIME experiments across three language models. At an eight-candidate budget, raw-score weighting improves accuracy by 2.78 percentage points for Math-7B and 2.91 percentage points for Phi-4, with most gains arising from resolving count ties.

Critical strengths
  1. Comprehensive analysis. The paper examines score-weighted voting from several perspectives, and the theoretical results appear reasonable and provide useful insights into when weighting helps or hurts.

  2. Attention to voting ties. The paper explicitly studies tie-breaking techniques in LLM test-time inference, which I think is new and interesting to the literature.

Critical weaknesses
  1. Limited treatment of adaptive inference. The paper mainly studies voting under fixed sampling budgets or in the asymptotic limit. However, it remains unclear how the proposed weighting rules could support adaptive sampling, stopping, and aggregation. This limits their practical guidance for settings where the inference budget is chosen dynamically.

  2. Practical value of tie-breaking needs further validation. The main empirical gains are demonstrated with eight candidates, where count ties may be relatively common. It would be useful to examine whether these gains persist at larger budgets, such as 40 candidates, where ties may become less frequent. Another natural baseline is to resolve a tie by sampling additional candidates until a unique count winner emerges. Comparing accuracy and sampling cost against this baseline would help establish the practical value of score-based tie-breaking.

  3. The limitation of a correct population mode assumption. The main theoretical results assume that the correct answer is already the unique population mode. However, an important motivation for using scores is to address cases where this basic assumption of self-consistency fails. In particular, the paper could better explain when retaining one base vote per candidate through hη(v)=1+ηb(v)h_\eta(v)=1+\eta b(v) is preferable to directly using scores or selecting the highest-scoring candidate. A clearer comparison with Best-of-N, especially when the population mode is incorrect, would strengthen the motivation of introducing hη(v)h_\eta(v).

  4. The presentation is difficult to follow. The paper contains repeated qualifications and defensive explanations that interrupt the main argument. Consolidating these caveats and presenting a clearer distinction between the central results and supplementary details would improve readability.

Questions for authors

Please see my critical weakness.

3 comments

Thank you for highlighting adaptive inference and the case where the most common answer is wrong. We added the sampling comparison you suggested, clarified the stopping guarantee, and tested how scores correct a wrong vote leader.

1. Should we score a tie or generate more solutions?

We compared these two choices directly. In a separate adaptive-inference test on 120 AIME problems, we replayed the same saved Math-7B solutions with Qwen PRM:

  • Score the tie: Keep an answer that has the most votes. If several answers tie, score only their solutions and choose the tied answer with the largest total score.
  • Sample more: Generate one more solution at a time until one answer leads, up to 128 solutions.

Each policy uses only the solutions available so far. Starting with 40 solutions:

PolicyAccuracyMean solutions generatedMean solutions scored
Score the tie25.83%40.000.98
Sample until one answer leads24.17%40.980

Scoring ties with 40 solutions also matched plain voting's observed accuracy with 128 solutions: 25.83%. That uses 88 fewer generated solutions, with fewer than one scored solution per problem on average. Our cost model estimates 24 GPU-seconds saved per problem compared with sampling until one answer leads (Appendix A.4; Table 16).

The stopping guarantee covers decisions made during sampling. The half-floor rule gives each solution weight max(score, 0.5). Under independent sampling from the same model and scorer, with scores calibrated for that problem, the probability that this rule ever replaces a correct count leader ahead by at least m votes with a wrong answer is at most exp(−2m). This holds across all sample sizes, including stopping based on the answers seen so far (Section 5.1; Proposition 7).

2. What happens when ties become less common?

Scores also improve accuracy by choosing an answer with fewer votes. Across all 362 AIME problems unused in our submission, ties occur in 29% of groups at eight solutions and 6% at 128.

With Math-7B and Qwen PRM, weighting beats count-first at 16, 32, 40, and 64 solutions. At 40 solutions, the gain is 2.21 percentage points [0.83, 3.68]. Count-first uses scores only to break ties, so this comparison measures the benefit beyond tie-breaking.

With Phi-4 on 120 AIME problems, the gain over count-first appears at all six tested budgets, reaching 3.33 points [0.83, 6.67] at 128 solutions (Section 6).

3. Can scores correct a wrong vote leader? Why keep a base vote?

Yes. Equation (14) gives an example where a wrong answer appears 40% of the time and the correct answer appears 35%. Scores let voting recover the correct answer.

Solutions per voteScore-weighted votingHighest-score selection
4094.4%80.0%
12899.9%80.0%

The weighting results use 200,000 simulated groups per budget; highest-score accuracy is calculated exactly. As more solutions are generated, weighted voting's accuracy approaches 100%, while highest-score selection approaches 80% (Appendix A.6; Table 21).

A base vote controls how much agreement matters. With weight c + score, each solution contributes c votes plus its score. Setting c = 0 uses scores directly; c = 1 keeps one vote per solution and adds the score.

In this example, the correct answer wins in the long run for 0 ≤ c < 2.2, including c = 1. The half-floor rule also corrects this example. Under calibration, it additionally preserves every already-correct long-run winner (Theorem 2).

This distinguishes two goals: correcting a wrong count leader and preserving a correct one. In practice, we keep one base vote and choose the score-based bonus using separate problems with known answers.

The model experiments also show the value of combining votes. With Math-Shepherd on all 362 AIME problems, score-weighted voting beats highest-score selection by 13.26 points [9.12, 17.13] at 128 solutions.

4. Presentation

We followed your suggestion to consolidate the qualifications. The main text now presents the central results with plain English language explanations, one scope table, and one limits paragraph. Section 6 summarizes the experiments, and the appendix contains the detailed procedures.

I genuinely like the core idea of combining majority voting with score weighting. Thank you for the detailed response, which addresses several of my initial concerns.

However, a few important questions and issues remain:

  1. Adaptive sampling vs. fixed-budget sampling: Regarding my first point, my main question is whether your framework can genuinely support adaptive sampling rather than relying on fixed-budget sampling (or merely breaking ties at the end of a fixed budget)—for example, how does it perform in an early-stopping setting compared with existing methods such as Adaptive Self-Consistency?

  2. Scope of tie-breaking gains: Regarding my second point, my concern is not addressed simply by scaling from 8 to 40 samples. If you continue increasing the sampling budget (e.g., to 80) or switch to a stronger LLM, ties will naturally occur even less frequently. Perhaps the authors should explicitly scope their claim and highlight that their method is particularly helpful for less consistent LLMs (or harder regimes where output variance is high).

  3. Theoretical characterization of cc and comparison between c=1c=1 and c=0c=0: Regarding my third point, it is unclear where the threshold 2.2 in c<2.2c < 2.2 comes from. A promising improvement would be to add formal theoretical results (e.g., theorems or propositions) that rigorously characterize the required conditions and this exact threshold. Furthermore, the current experimental results do not seem to convincingly support that c=1c =1 consistently outperforms c=0c = 0.

  4. Revision quality and presentation: Thank you for the significant effort you put into revising the manuscript. However, as Reviewer axPD noted, "such a substantially revised manuscript would be better evaluated through a new review cycle." Moreover, I am still not fully satisfied with the revised version. For example, (i) The newly added table on Page 2 does not work well and is likely to confuse readers rather than clarify the main message; (ii) Experimental Findings: Organizing the experimental section into 9 Findings dilutes the focus. Highlighting the 3 most important findings would be much more effective. (iii) Inconsistent Notation: Notation inconsistencies significantly detract from the paper's readability. For example, at least three different symbols are used for probability throughout the manuscript (PP, Pr⁡\Pr, and P\mathbb{P}). Please carefully check these instances and streamline the notation to improve overall readability.

I hope these comments can help the authors further improve their paper.

Thank you for your continued engagement and support for the core idea. Your follow-up identifies four points we should make clearer.

1. Adaptive stopping

Choosing when to stop and choosing how to combine answers are separate decisions. Our weighting rules can be applied when sampling stops based on the answers seen so far.

Proposition 7 remains valid at such stopping times under independent sampling and per-problem calibration. It bounds how often the half-floor rule overturns a correct count leader; it does not establish that the leader is correct.

Our experiment compares scoring a tie with sampling until one answer leads. We have not run a direct comparison with Adaptive Self-Consistency.

2. Where tie-breaking helps

Tie-breaking is useful when generated answers disagree enough to tie. This includes small budgets and difficult problems, including difficult problems for strong models. Answer disagreement is the relevant factor, rather than model size alone.

Full weighting also changes answers outside ties. Across 362 AIME problems, it beats count-first by 2.21 percentage points [0.83, 3.68] at 40 solutions. We distinguish this benefit from tie-breaking (Section 6).

3. Where the threshold 2.2 comes from

The threshold follows exactly from Equation (14). For weight c + score, each answer receives a base contribution from its frequency plus its score contribution. In that example, the expected totals per sample are:

  • Correct answer: 0.35c + 0.19
  • Leading wrong answer: 0.40c + 0.08
  • Other wrong answer: 0.25c + 0.08

The correct answer beats both exactly when 0 ≤ c < 2.2, because:

0.35c + 0.19 > 0.40c + 0.08 exactly when c < 2.2.

Under independent sampling, these expected totals determine the unique long-run winner. Thus, 2.2 is an exact threshold for this example, rather than a fitted parameter or a universal recommendation.

The base controls how much agreement counts relative to scores. Its benefit depends on the problem. The separate half-floor rule has a precise preservation guarantee under calibration (Theorem 2), rather than a guarantee of always beating raw scores.

4. Presentation and review scope

We agree with your specific suggestions: move the page-2 table to the appendix, focus the experiments on three main findings, and use consistent probability notation.

The three main findings are gains from breaking ties, gains beyond ties, and comparisons with highest-score selection. This provides a simpler structure for the same central results.

We appreciate the limited review time and would be grateful for consideration of the submission’s theoretical contribution and the new experiments we ran during the rebuttal phase that answer your questions.

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

This paper investigates when score-weighted voting can outperform standard plurality voting in LLM reasoning, providing a theoretical characterization of the conditions under which weighting improves the asymptotic error decay rate. The authors further analyze the role of score calibration and finite-sample voting, and demonstrate empirical improvements on mathematical reasoning benchmarks across multiple language models.

Critical strengths
  1. The paper provides a rigorous necessary-and-sufficient characterization of when score weighting improves plurality voting’s asymptotic error exponent.
  2. The experiments demonstrate consistent gains across multiple models, with careful statistical analysis and practical insights into tie-breaking.
Critical weaknesses

The main theorem relies on unknown problem-specific score distributions and asymptotic analysis, limiting its direct applicability to practical finite-budget inference.

Questions for authors

Given that Theorem 1 relies on unknown problem-specific score distributions, how can its characterization be used to guide the design of practical voting rules under limited sampling budgets?

3 comments

Thank you for your positive assessment of the theory and experiments. Your question led us to add the following practical voting procedure below and test it on new problems.

1. How do we use the theorem to choose a voting rule?

We use problems with known answers to choose the voting rule, then test that rule on separate problems (not used to choose the rule). The known-answer problems used to choose the rule are called development problems.

Which problems did we use? We used 120 AIME development problems from our submission, split into 60 for training and 60 for validation. Training proposes voting rules; validation chooses among them. These are separate from the new AIME test batches we mentioned in our replies to the other reviews.

For the expanded development experiment, Math-7B generated 384 solutions for each of the 120 development problems. Each solution has a final answer and scores for its reasoning steps. We combine those step scores into one score for the solution. The known correct answer lets us check whether a voting rule selects the right answer.

Every solution receives one vote plus a bonus based on its score:

Vote weight = 1 + bonus multiplier × score-range bonus.

We use the 60 training problems to divide scores into four ranges. We try bonuses of 0, 0.5, or 1 for each range, with multipliers from 1/16 to 8. This creates different ways to give more influence to solutions in particular score ranges.

We then compare the proposed rules on the other 60 development problems, e.g. the validation set. For each rule, we combine groups of 40 generated solutions, add the vote weights for each final answer, and select the answer with the largest total. We choose the rule with the highest validation accuracy.

With Qwen PRM and multiplied step scores, the experiment selected the following rule:

A solution scoring at least 0.580 counts as nine votes. A solution below 0.580 counts as one vote.

For example, a solution scoring 0.60 contributes nine votes to its final answer, while one scoring 0.40 contributes one. All solutions giving the same final answer contribute to that answer's total.

More details, including experimental results, can be found in Appendices A.2 and A.8–A.9.

2. What does the theorem contribute?

It tells us which bonuses make voting errors shrink faster. Correct solutions must receive a higher average bonus than solutions giving each of the most common wrong answers. We check these wrong answers separately, rather than combining them into one group.

For independent solutions where the correct answer is most likely, a small enough bonus meeting this condition improves the long-run error rate.

This guides our voting rule: one vote per solution, plus a score-based bonus. We choose the bonus and its size using problems with known answers, comparing accuracy with 40 solutions per vote.

3. Does the chosen rule work on new problems and another model?

Yes. Across all 362 AIME problems, including ones not included in our submission, our chosen rule beats its count-first version.

With 40 solutions per vote, the gain is 2.39 percentage points [0.74, 4.14] when we multiply each solution’s step scores. Using their geometric mean, the gain over count-first is 1.75 points [0.18, 3.41]. These gains come from changing the vote winner, rather than only breaking ties.

We also reused the geometric-mean rule unchanged on Phi-4’s solutions. On its 120-problem test, it beat using scores directly by 4.17 points [0.83, 8.33], with 128 solutions per vote (Appendix A.9; Tables 32 and 38).

4. Do simple rules help with small sampling budgets?

Yes. Benefits already appear with eight solutions. Across all 362 AIME problems unused in our submission, Math-7B with Qwen PRM improves over plain voting by 1.78 percentage points [0.93, 2.61] at eight solutions and 2.39 points [0.92, 3.96] at 40.

At 40 solutions, weighting also beats count-first by 2.21 points [0.83, 3.68] on these same 362 problems. This shows that scores help beyond breaking ties.

With the weaker Math-Shepherd scorer, weighted voting beats choosing the single highest-scoring solution by 13.26 points [9.12, 17.13] at 128 solutions on the same problems.

These results support starting with count-first and testing full weighting on separate problems with known answers (Section 6).

5. Does the theorem predict the long-run outcome?

We test this separately in six constructed cases. Each answer appears equally often across the cases, and solutions with a given score are correct equally often. We change only which wrong answers receive high scores.

Testing the same set of voting rules in all six cases matches Theorem 1's prediction: errors shrink faster as more solutions are added exactly where its condition holds (Table 19).

Section 8 now gives practical extensions based on these findings.

Thank you for the substantial revision and additional experiments. The proposed practical weight-selection procedure is a useful improvement. However, it still relies primarily on empirical search guided by validation accuracy, instead of directly operationalizing the distributional separation criterion in Theorem 1. Therefore, I consider my original concern partially addressed.

Additionally, the manuscript has undergone extensive rewriting and differs substantially from the original submission. Given the limited time available during the rebuttal period, it is difficult to thoroughly assess all these changes. I would respectfully suggest that such a substantially revised manuscript would be better evaluated through a new review cycle.

Thank you for reading the revision and recognizing the practical procedure as an improvement. We appreciate the time you have already spent on the paper. Two points may help clarify the remaining concern.

1. The paper also applies the separation criterion directly.

You are right that the validation-based procedure chooses weights by their accuracy on development problems. Our previous response emphasized that procedure. The direct application of Theorem 1 is a separate analysis in Appendix G.7.

That analysis groups scores by answer and solves a linear program to find a bonus that favors the most frequent answer over each plausible leading rival. Confidence bounds account for sampling uncertainty. The correctness of the reference answer is checked separately.

Page 100 gives a concrete example using 256 Math-7B solutions to AIME 2006-II-3. For the bonus “add extra weight when the score is at least 0.5”:

  • 171 of 188 solutions giving the correct answer receive the bonus.
  • 7 of 46 solutions giving the leading wrong answer receive it.

After accounting for sampling uncertainty, the correct answer’s average bonus still exceeds the rival’s by at least 0.406. The frequency checks also establish the theorem’s required ordering of answers. Theorem 1 then guarantees that a sufficiently small bonus makes voting errors shrink faster as more independent solutions are added.

This directly checks the theorem’s condition on model outputs. The audit certifies two of the 120 examined problems. It establishes the long-run improvement condition; the separate validation study selects rules for accuracy at practical sampling budgets.

2. The revision strengthens the same central contribution.

The central contribution remains the exact characterization you assessed positively in your original review. The distinction between calibration and voting quality, and the guarantees for tie-breaking, also remain central.

We rewrote the explanations in response to the presentation concerns and added experiments on the questions raised by the reviewers. The additional experiments support the submission’s theoretical questions; they do not replace those questions with a different contribution.

For your specific concern, the most relevant passage is Appendix G.7, pages 99–101, especially the worked example on page 100. It shows the direct connection between measured score distributions, a concrete bonus, and Theorem 1’s guarantee.

We understand the limited review time and appreciate the time you've spent evaluating our submission.

Reviewer_4aRzOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2(+1)
Read review
Paper summary

The paper studies when per-candidate scores, such as process reward model (PRM) scores, can improve plurality voting over sampled LLM solutions to a single problem. In an iid model where the correct answer is the most frequent but not necessarily a majority, Theorem 1 characterizes when some bounded, strictly positive weighting improves plurality’s error exponent. This happens if and only if the correct answer’s score distribution lies outside the convex hull of the score distributions of the most frequent wrong answers. A toy example in table 1 shows two calibrated populations with identical score: correctness statistics where only one admits improvement, which argues that calibration alone is not the right diagnostic. The authors also characterize which weights preserve the correct winner on every calibrated distribution, and give finite-budget guarantees for using scores only to break count ties. On 252 AIME problems with eight candidates each, PRM-score weighting improves accuracy by about 2.8 points for Qwen2.5-Math-7B and 2.9 for Phi-4, and most of the gain comes from breaking count ties. Overall, the theory is rigorous and the calibration insight is useful, but the experiments mostly test tie-breaking at small budgets that the main theorem does not address, and the empirical scope is narrow.

Critical strengths
  • Theorem 1 is an exact iff characterization, and its geometric condition, that the correct answer’s score distribution lies outside the rivals’ mixture hull, is intuitive.
  • Table 1 shows clearly that calibration, and pooled reliability plots in particular, miss the information that matters for voting: how high scores are spread across wrong answers.
  • Overall the theoretical treatment and results are detailed and rigorous.
Critical weaknesses
  • Theorem 1 concerns error exponents as N goes to infinity, yet the headline results are limited to N = 8. Most of the gains, e.g., 18 of 21 for Math-7B, come from count ties. Count-first refinement, which uses scores only at ties, retains most of this gain (2.38 of 2.78 points) while leaving the exponent unchanged. So the main theorem does not explain the main empirical result.
  • Raw weighting performs as well as highest-score selection (0.00 [−1.19, 1.19]). 13 of count-first’s 18 net gains occur on pools where every answer is distinct, where count-first is just highest-score selection. The results therefore mostly show that a PRM beats an arbitrary first-occurrence tie-break when vote counts are tied, which is expected. Against random-tie plurality the gain drops to 2.20 points.
  • The experiments are limited with only one scorer, one benchmark, and mainly one budget. The Math-1.5B result is inconclusive. Results across budgets (16–128) are exploratory and appear only in the appendix. I recommend the authors validate the theory on more models and benchmarks to test its generalizability.
  • The writing seems to be AI-generated with little editing from the authors to ensure clarity, and so it is very difficult to read.
Questions for authors
  • Do any gains come from overriding a count leader rather than from tie-breaking? How does raw weighting compare with count-first as N grows from 8 to 128?
  • Do the results hold for other scorers and benchmarks, or is the dominance of tie-breaking specific to AIME?
3 comments

Thank you for recognizing the theory as “detailed and rigorous.” We added experiments that directly answer your two questions and rewrote the main text for clarity.

1. Do scores help beyond breaking ties?

Yes. Scores improve accuracy by choosing an answer with fewer votes.

Count-first picks the answer with the most votes and uses scores only to break ties. Weighted voting adds the scores for each answer and chooses the largest total. Beating count-first therefore shows a gain beyond tie-breaking.

On 242 AIME problems unused in our earlier submitted version of the paper, weighting beats count-first by 2.62 percentage points at 40 solutions. The 97.5% confidence interval is [0.55, 4.82].

Across both test batches, 120 initial problems in our submission plus 242 additional problems, totaling 362, Math-7B with Qwen PRM beats count-first at 16, 32, 40, and 64 solutions.

With Phi-4, weighting beats count-first at all six tested budgets, from +1.46 [0.68, 2.34] at eight solutions to +3.33 [0.83, 6.67] at 128. Multiplying the step scores instead of taking their minimum raises the gain at 128 to +7.50 [3.33, 12.50].

On 60 problems from our submission, using new solutions, the gain over count-first grows from 3.33 points [0.56, 6.67] at 40 solutions to 10.00 [3.33, 18.33] at 128.

We test the theorem’s long-run prediction separately. Six constructed cases have the same answer frequencies and the same accuracy at each score. We change only which wrong answers receive high scores. Testing the same set of weights in every case matches Theorem 1’s prediction: errors shrink faster as more solutions are added when the conditions of the theorem holds (Section 6; Table 19).

2. Do the gains extend to other scorers, benchmarks, and baselines?

We now test two generators, two scorers, four benchmarks, three ways to combine step scores, and budgets from 8 to 128.

Weighted voting also beats choosing the single highest-scoring solution. Both methods use the same generated solutions and scores, matching generation and scoring costs. Selected gains at 128 solutions, using minimum step scores:

Test setScorerGain over highest-score selection
362 unused AIME, Math-7BMath-Shepherd+13.26 [9.12, 17.13]
120 AIME, Phi-4Math-Shepherd+5.83 [0.83, 11.67]
146 earlier MATH-500 problemsMath-Shepherd+6.85 [2.05, 11.64]
Same MATH-500 problemsQwen PRM+4.11 [0.68, 8.22]
120 GSM-HardMath-Shepherd+3.33 [0.83, 6.67]

Gains beyond tie-breaking also appear on MATH-500. With Qwen PRM, weighting beats count-first at 16 and 32 solutions. It also improves on plain voting at eight solutions by 1.24 points [0.34, 2.27], even though plain voting is already 89.1% accurate.

The benefit holds across different ways of scoring a solution. On the first 120 unused AIME problems at 40 solutions, weighting beats plain voting using the minimum, product, and geometric mean of step scores (Appendix A.3).

Section 6 now presents these findings in the main text, with the full budget curves in Table 2 and Figure 1. Appendix A.9 gives the latest experiments.

3. How have we improved the writing?

We rewrote the abstract, introduction, and main text around two questions: when does overruling help, and when is tie-breaking safe?

Each formal result is followed by a plain English explanation. Section 6 presents the experiments in a clearer way.

4. Soundness

We would appreciate your reassessment of the soundness rating based on the revised paper and new experiments. Please identify any specific statement, proof, or experimental claim that needs further clarification so we can address it directly.

Thank you for the new experiments. The gains over count-first at 16–64 solutions answer my main question regarding whether scores help beyond tie-breaking at moderate budgets. However, the gains over best-of-N are shown only with Math-Shepherd. With Qwen PRM on AIME the comparison is inconclusive or favors best-of-N, as you note in your reply to Reviewer MwJr. The link to Theorem 1 also rests only on constructed cases. I have raised my soundness score to 2. Regarding the writing, the authors made drastic changes that I think would be difficult to assess given the remaining time in the discussion phase, so I will keep my original presentation score.

Thank you for confirming that the new experiments answer your main question and for raising the soundness score. We would like to clarify the two remaining scientific points.

1. Weighting also beats best-of-N with Qwen PRM.

On 146 MATH-500 problems, weighting with Qwen PRM beats best-of-N by 4.11 percentage points [0.68, 8.22] at 128 solutions, using the same generated solutions and scores (Table 37).

The advantage therefore extends beyond Math-Shepherd. The results distinguish when voting helps across different scorers and tasks.

2. Theorem 1 also has a direct application to real model outputs.

Appendix G.7 checks the theorem’s condition on Math-7B outputs and certifies it on two of 120 problems.

The worked example on AIME 2006-II-3 makes this concrete. A score-based bonus goes to 171 of 188 correct solutions, compared with 7 of 46 solutions giving the leading wrong answer. Statistical checks confirm both the required answer-frequency ordering and a positive bonus gap.

Under independent sampling, Theorem 1 then guarantees that a sufficiently small bonus makes errors shrink faster as more solutions are added. This directly connects real model outputs to the theorem’s long-run guarantee (page 100; Table 70).

3. The rewrite addresses your presentation concern.

The central theoretical results remain the same. We simplified their explanations and added experiments answering the reviews.

For a focused check of the writing, Section 3.2 (page 4) explains the bonus rule, and Section 6 (pages 7–9) presents the experimental findings.

We appreciate your continued engagement. With your main experimental question answered, we would be grateful for your reconsideration of the overall recommendation in light of these clarifications.

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

This paper studies when score-weighted plurality voting improves on ordinary plurality for aggregating language-model answers. Under an iid, finite-answer model where the correct answer is the unique population mode, it characterizes the existence of bounded, strictly positive score-only weights that improve the asymptotic error exponent. The criterion depends on whether the correct answer’s conditional score distribution lies outside the convex hull of the leading incorrect answers’ score distributions. The paper also characterizes calibration-based preservation guarantees and provides a finite-budget guarantee for score-based tie-breaking under conditional stochastic dominance. Experiments on AIME show modest accuracy improvements, primarily attributable to resolving count ties.

Critical strengths
  1. A precise theoretical characterization. The main result provides a necessary and sufficient condition for improving the error exponent, rather than only a sufficient condition for a particular weighting rule.

  2. A useful distinction between calibration and aggregation quality. The constructed populations share identical score–correctness statistics and answer frequencies but differ in whether weighting can improve the exponent. This clearly demonstrates why answer-specific score distributions matter.

  3. Careful separation of different guarantees. The paper distinguishes preservation of the population winner, improvement in asymptotic error rates, and finite-budget accuracy. Its counterexamples help prevent conflating these properties.

  4. Transparent empirical reporting. The experiments use paired candidate pools and problem-level uncertainty estimates. The paper also discloses retrospective analyses, inconclusive findings, and limitations of the learned-weight experiments.

Critical weaknesses
  1. The connection between the central theorem and the main empirical gains remains limited. The theorem concerns asymptotic error exponents, whereas the main experiments use eight candidates and attribute most gains to count ties. These results establish useful finite-budget effects but provide limited direct evidence for the central asymptotic mechanism.

  2. The characterization does not yet yield a broadly actionable weight-selection procedure. Applying the result requires information about answer-specific score distributions and leading rivals, which are unknown at deployment. The valid certification analysis succeeds on only 2 of 120 Math-7B problems. This does not imply that improvement is impossible elsewhere, but it illustrates the current difficulty of operationalizing the theorem.

  3. Evidence for learning transferable weights is preliminary. The primary selection screen retains plurality. The positive learned-weight result comes from a secondary screen on retrospectively assigned problem roles, with bin preprocessing shared across those roles. Consequently, it provides limited evidence for prospective generalization to unseen problems.

  4. The empirical advantage over simple score-based alternatives is not established broadly. In the main Math-7B study, raw-score weighting has the same observed accuracy as highest-score selection. Evaluation is also concentrated on one benchmark family, one scorer, and a small candidate budget, limiting conclusions about the practical value of weighting beyond simpler alternatives.

  5. The incremental theoretical contribution needs sharper positioning. The proof draws on classical large-deviation, convex-separation, and minimax arguments. The paper would benefit from a more explicit account of which results are unavailable from prior weighted-voting analyses and why the new characterization materially changes our understanding. The use of established mathematical tools is not itself a weakness; the concern is the clarity of the novelty and significance argument.

Questions for authors
  1. Can you evaluate several candidate budgets on fresh model samples and demonstrate whether the proposed separation criterion predicts improvements beyond tie-breaking gains?

  2. Can you provide a practical procedure for choosing both the bonus function and its intensity, with all preprocessing and selection fixed before evaluation on new problems?

  3. The criterion depends on exactly leading rivals. How should a practitioner handle nearly tied rival frequencies, where small estimation errors may change the identified rival set and the useful intensity range?

  4. Which results most clearly distinguish this work from prior error-exponent analyses of weighted voting? An explicit comparison of assumptions and conclusions would help establish the theoretical novelty.

  5. Do the conclusions persist across additional scorers and task families, and under matched-compute comparisons with highest-score selection and count-first tie-breaking?

2 comments

Thank you for the questions. We added fresh samples, larger budgets, a weight-selection procedure, and tests with more models, scorers, and benchmarks.

1. Do scores help beyond breaking ties?

Yes. Scores improve accuracy by choosing an answer with fewer votes. Count-first picks the answer with the most votes. If several answers tie for the lead, it picks the tied answer with the highest total score. Count-first uses scores only to break ties. Weighted voting adds the scores for each answer and chooses the largest total. Beating count-first therefore shows a benefit beyond tie-breaking.

On 242 AIME problems unused in our submission, the test at 40 solutions shows a 2.62-percentage-point gain over count-first. Its 97.5% confidence interval is [0.55, 4.82].

Across all 362 unused problems, weighting beats count-first at 16, 32, 40, and 64 solutions with Math-7B and Qwen PRM. With Phi-4, it wins at all six budgets, reaching +3.33 [0.83, 6.67] at 128 solutions.

We test the theorem's long-run prediction separately. Six constructed cases have the same answer frequencies and the same accuracy at each score. We change only which wrong answers receive high scores. Testing the same set of weights in every case matches Theorem 1's prediction: errors shrink faster as more solutions are added exactly where the conditions of the theorem hold (Section 6; Table 19).

2. How do we choose the weights before testing?

We choose the entire voting rule on separate development problems, then keep it fixed.

  1. Split 120 development problems into 60 for training and 60 for validation.
  2. Use training scores to define four score ranges. Try bonuses of 0, 0.5, or 1 in each range, multiplied by a factor from 1/16 to 8. Every solution gets one base vote plus its bonus.
  3. Training proposes rules; validation chooses one by accuracy with 40 solutions.
  4. Freeze the score ranges, bonuses, and multiplier before testing on new problems.

Our chosen voting rule beat its count-first version on both AIME test batches: +2.78 [0.28, 5.83] and +2.20 [0.14, 4.41] percentage points. Each vote used 40 solutions, with each solution scored by multiplying its step scores.

Using geometric-mean scores, a rule chosen on Math-7B also improves Phi-4's accuracy when reused unchanged: +4.17 [0.83, 8.33] over using scores directly at 128 solutions (Appendices A.2, A.8–A.9).

3. Nearly tied wrong answers

When wrong answers have similar vote counts, we check whether the proposed weights favor the correct answer over each of them.

With 512 samples per run, this cut false predictions of improvement from 48.5% to 0.4% of runs. The rule recommended changing the weights in 2.7% of runs (Table 20).

4. What does the theory add to prior work?

Theorem 1 gives an exact test for whether any bounded, positive score-based weighting can improve plain voting's long-run error rate. It applies to independent samples where the correct answer is the most likely answer, and allows each wrong answer to have its own score distribution.

Earlier work studies voters with known reliability or finds optimal weights under a specified model. Kuang et al. assume interchangeable wrong answers and conditional independence. Our result handles arbitrary dependence between answers and scores and also gives the exact initial effect of adding a score bonus.

The key new insight is that which wrong answers receive high scores matters. Table 1 gives two cases with identical answer frequencies and identical relationships between scores and correctness, yet opposite answers to whether weighting can improve the long-run rate. Section 7 now makes this distinction explicit.

5. Do the results extend to other tasks and matched-cost comparisons?

We test two generators, two scorers, four benchmarks, three ways to combine step scores, and budgets from 8 to 128.

Weighted voting and best-of-N use the same solutions and scores, matching generation and scoring costs. Selected gains over best-of-N at 128 solutions, using minimum step scores:

Test setScorerAccuracy gain
362 unused AIME, Math-7BMath-Shepherd+13.26 [9.12, 17.13]
120 AIME, Phi-4Math-Shepherd+5.83 [0.83, 11.67]
146 earlier MATH-500 problemsMath-Shepherd+6.85 [2.05, 11.64]
Same MATH-500 problemsQwen PRM+4.11 [0.68, 8.22]
120 GSM-HardMath-Shepherd+3.33 [0.83, 6.67]

Section 6 and Appendix A.9 give the results in more detail.

Reviewer_upZaOpenReview ↗

Response to the Authors

Thank you for addressing my concern. I will raise my score.

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

The paper asks when scoring individual LLM solutions (here, min-step PRM scores) adds value over plain answer counting in self-consistency voting. For a fixed problem with iid samples and a correct population mode, Theorem 1 gives an exact criterion: some bounded, strictly positive score-only weight improves plurality’s error exponent iff the correct answer’s conditional score distribution lies outside the convex hull of the leading wrong rivals’ conditional score distributions. Two perfectly calibrated populations with identical score–correctness statistics but different wrong-answer allocation (Table 1) show that calibration alone is uninformative. Theorems 2–3 characterize which weights preserve a correct long-run winner under calibration versus protect accuracy at every finite budget, with exact counterexamples separating the two. Proposition 7 bounds harmful half-floor reversals of a count lead ≥ m by e^{−2m}; Proposition 1 shows count-first score tie-breaking never hurts under conditional stochastic dominance. Empirically, raw-score weighting gains 2.78 pp (Math-7B) and 2.91 pp (Phi-4) over plurality on 252 AIME problems at N = 8, with most of the gain from resolving count ties.

Critical strengths

The hull criterion is clean, correct and genuinely clarifying; the Table 1 construction makes the calibration-versus-allocation point sharply and is the kind of example that changes how one evaluates verifiers. The population/finite-budget distinction is handled with unusual rigor, including exact tie-independent counterexamples. The empirical protocol is honest: pre-declared replication, explicit post hoc labelling, compute-matched plurality, and a decomposition of gains into tie-breaks versus count overrides.

Critical weaknesses

The headline theory and the headline empirics do not explain each other. The gains come overwhelmingly from count ties (18/21 and 17/22), and the paper itself proves that tie-only refinement leaves the error exponent unchanged, so Theorem 1 is not the mechanism behind the observed improvement. Worse, 13 of 18 count-first gains occur on all-distinct pools where the rule reduces to highest-score selection, and raw weighting versus highest-score selection is exactly 0.00 pp; the practical takeaway is close to “when votes tie, pick best-of-N,” which is known (Kokkinos & Margaritis 2014) and would be the natural strong baseline rather than a post hoc control.

The exponent criterion is asymptotic and hard to use: the certified crossover in one law is N ≥ 396, a positive direction can be certified on only 2/120 problems, and the learned-weight primary screen retains plurality.

Empirical breadth is narrow (one scorer, one PRM scalarization chosen ad hoc, AIME only, ~37% base accuracy, N = 8 only in the prospective study, Math-1.5B inconclusive). Related-work positioning against Kuang et al. (2026) and CISC should be sharper. Finally, the manuscript is 92 pages and the main text is dense with scope caveats to the point of obscuring the argument.

Questions for authors

As shown in weaknesses

4 comments
Reviewer_MwJrOpenReview ↗

Lower my score

The authors did not try to answer my questions at all. After considering the other reviewers’ assessments and revisiting my initial evaluation, I have lowered my score.

AuthorsOpenReview ↗

New experiments and answers to the review

Thank you for the review. We apologize for the late reply. Completing the requested experiments took a lot of time with the computing resources available to us. We should have posted an interim update instead of complete silence.

1. Scores help beyond breaking ties.

You are right that the original eight-solution gains came mainly from ties. The new experiments show gains from choosing an answer with fewer votes.

Count-first uses scores only to break ties. Weighted voting adds the scores for each answer and chooses the largest total. Its gain over count-first therefore measures the benefit beyond tie-breaking.

We tested another 242 AIME problems unused in our submission.

With Math-7B and Qwen PRM, weighting beat count-first by 2.62 percentage points at 40 solutions per vote. The 97.5% confidence interval [0.55, 4.82] accounts for both tests and stays above zero. The test at 128 solutions showed +2.48 points, 95% interval [−0.41, 5.37].

Across all 362 unused problems, weighting beats count-first at 16, 32, 40, and 64 solutions. With Phi-4, it beats count-first at all six budgets; the gain at 128 is 3.33 points [0.83, 6.67], or 7.50 [3.33, 12.50] when step scores are multiplied.

These are gains at fixed budgets. We test the long-run claim on six constructed cases that differ only in which wrong answers get high scores. The tested voting rules reduce errors faster as more solutions are added exactly when Theorem 1 predicts (Section 6).

2. Weighted voting is more than selecting the highest-scoring solution.

Best-of-N chooses one solution. Weighted voting combines support for the same answer. At 128 solutions, their accuracy differs as follows:

Test setScorerWeighted voting's gain over best-of-N
362 unused AIME, Math-7BMath-Shepherd+13.26 [9.12, 17.13]
120 AIME, Phi-4Math-Shepherd+5.83 [0.83, 11.67]
146 earlier MATH-500 problemsMath-Shepherd+6.85 [2.05, 11.64]
Same MATH-500 problemsQwen PRM+4.11 [0.68, 8.22]
120 GSM-Hard problemsMath-Shepherd+3.33 [0.83, 6.67]

Gains are percentage points with 95% intervals, using the minimum step score. On AIME with Qwen PRM, the comparison is inconclusive for Math-7B; best-of-N wins for Phi-4 at 64 and 128. The advantage over best-of-N depends on the scorer and task.

We agree that score-based tie-breaking is known and cite Kokkinos and Margaritis. Proposition 1 adds a condition under which it cannot hurt accuracy.

3. The rule can be chosen on separate problems and then reused.

Weighting already helps with four solutions in this example: 62.94% accuracy versus 53.38% for plain voting. Our bound guarantees an advantage at every vote size of 396 or more; 396 is not a minimum requirement for improvement. In that same example, the half-floor rule already wins with four votes: 62.94% versus 53.38%.

The 2/120 result counts successful certificates, not all problems that benefit from weighting.

We now choose how much each score counts in the vote using separate development problems, then keep that rule fixed during testing.

When step scores are multiplied, using the chosen rule for all votes beats using it only to break ties on both unused AIME batches at 40 solutions: +2.78 [0.28, 5.83] and +2.20 [0.14, 4.41].

With geometric-mean scores, a rule chosen using Math-7B also works on Phi-4 without changes. It beats using scores directly by 4.17 points [0.83, 8.33] at 128 solutions.

4. The evaluation is broader, and the related-work comparison is clearer.

We now test two generators, two scorers, four benchmarks, and six budgets from 8 to 128. We combine step scores in three ways: minimum, product, and geometric mean. On MATH-500, plain voting is already 89.1% accurate at eight solutions, yet weighting adds 1.24 points [0.34, 2.27].

CISC: We added this confidence-weighted voting baseline. Using Qwen PRM scores beats it by 0.99 points [0.05, 2.03] at eight solutions on the first batch. Table 1 also shows that scores can identify correct solutions equally well, yet allow faster long-run error reduction in only one of two cases.

Kuang et al.: Their optimal-weight result assumes interchangeable wrong answers and conditional independence, and allows negative weights. Our theorem gives an exact test for improvement using bounded positive weights while allowing each answer to have its own score distribution (Section 7).

5. We rewrote the paper to improve the presentation.

The ten-page main text centers on two questions: when does overruling help, and when is tie-breaking safe? Assumptions and limits are collected in one table and a limits paragraph. Section 6 presents the new findings; Appendix A.9 gives the latest experiments.

Thank you for the detailed response and the new experiments; I have read the rebuttal and the revised paper. The new results show that weighting can beat count-first at N = 16–64, which answers part of my first concern. But three issues remain.

(1) The link to Theorem 1 still rests only on constructed laws. In that same family (Table 19), rules gain 4.75–8.87 points over count-first at N = 40 where D = 0, so the criterion does not predict finite-budget gains. On AIME the correct answer is the empirical mode on fewer than half of the problems (23/60 and 51/120 in the appendix), which is outside the theorem’s premise.

(2) With the Qwen PRM, raw weighting is not separated from best-of-N at any budget on the 362 problems.

(3) The criterion remains hard to use. The N = 4 example is a tie-breaking gain, since count-first attains the same accuracy there. The fitted rule with the primary minimum score is worse than raw weighting at N = 32 and 40 (Table 32).

Thank you for confirming that the new results establish gains beyond tie-breaking. We would like to answer each remaining point directly.

1. The theorem is applied to real model outputs.

The connection to Theorem 1 goes beyond constructed examples. Appendix G.7 implements its separation test on Math-7B outputs and certifies two AIME problems.

For AIME 2006-II-3, the analysis checks that the correct answer is the most likely answer and finds a score-based bonus that favors it over its leading rival. After accounting for sampling uncertainty, the bonus gap remains at least 0.406. Under independent sampling, Theorem 1 then guarantees faster long-run error decay for a sufficiently small bonus (page 100).

The 23/60 and 51/120 figures describe the most frequent answers in finite samples. Our direct application checks the population-mode assumption separately; it does not assume that every benchmark problem satisfies it.

Regarding D = 0, a rule can win at 40 solutions without making errors shrink faster in the long run. Theorem 1 tells us exactly when that stronger improvement is possible. The constructed examples demonstrate why these two outcomes must be distinguished.

2. The best-of-N comparison establishes benefits with both scorers.

The Qwen PRM comparison on AIME is one part of the evaluation. At 128 solutions, weighting beats best-of-N by:

Scorer and taskAccuracy gain
Qwen PRM, 146 MATH-500 problems4.11 points [0.68, 8.22]
Math-Shepherd, 362 AIME problems13.26 points [9.12, 17.13]

Both comparisons use the same generated solutions and scores. These results directly establish that combining votes can outperform choosing one top-scored solution.

On Qwen PRM with AIME, the established benefit is over plain voting and count-first. Across the broader evaluation, we also establish benefits over best-of-N. The conclusions identify where each benefit occurs.

3. The practical evidence comes from rules tested on new problems.

The four-solution example answers the concern about the 396-solution bound: improvement can begin much earlier. Our evidence for practical weight selection comes from separate experiments.

Table 32 also contains positive results for the minimum-score fitted rule. Across all 362 AIME problems, at 64 solutions it beats plain voting by 2.35 points [0.97, 3.87] and count-first by 1.66 points [0.55, 2.90].

Raw weighting performs better at the two budgets you identify. That comparison determines which rule performs better there; it does not remove the fitted rule’s demonstrated gains over counting.

Learning also improves on raw weighting itself. With geometric-mean scores, a rule chosen using Math-7B and reused unchanged on Phi-4 gains 4.17 points [0.83, 8.33] at 128 solutions (Table 38).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overview of edits and new experiments

Overview of the new experiments

Thank you for the feedback. We apologize for posting late; we waited for the requested main experiments to finish, which took a while due to compute limits.

The new experiments show that scores can improve voting by overruling the simple count, not just by breaking ties.

What we tested. We generated 128 solutions per problem on 362 AIME problems unused in our earlier experiments already included in our submission, 60 earlier AIME problems, and GSM8K, GSM-Hard, and MATH-500. We tested Math-7B and Phi-4, two scorers (Qwen PRM and Math-Shepherd), and votes using 8–128 solutions. Two independent implementations checked the latest results.

How to read the results. Plain voting picks the most common answer. Weighted voting adds the scores behind each answer and picks the largest total. Count-first uses scores only to break ties. The gain over count-first therefore measures the benefit of overruling the count. Gains are accuracy percentage points; brackets give 95% confidence intervals.

1. Scores improve voting, and the benefit goes beyond ties.

On all 362 unused AIME problems, with Math-7B and Qwen PRM (minimum step score):

Solutions per voteGain over plain votingGain over count-first
8+1.78 [0.93, 2.61]+0.21 [−0.35, 0.73]
16+2.21 [1.04, 3.45]+1.14 [0.21, 2.07]
32+2.62 [1.10, 4.21]+2.07 [0.83, 3.45]
40+2.39 [0.92, 3.96]+2.21 [0.83, 3.68]
64+2.76 [0.97, 4.56]+1.93 [0.41, 3.45]
128+1.93 [−0.55, 4.42]+1.93 [−0.28, 4.42]

At eight solutions, most of the benefit comes from breaking ties. At 16–64, choosing an answer with fewer votes adds a clear gain.

2. Larger numbers of solutions confirm the benefit of overruling.

With 40 solutions per vote, weighting beat count-first by +2.62 points [0.83, 4.55]. The gain remains positive after accounting for both tests: 97.5% interval [0.55, 4.82]. The test at 128 solutions showed a similar gain: +2.48 [−0.41, 5.37].

3. The gains also appear with a different generator.

Using Phi-4 on the first 120 AIME problems, weighted voting beat both plain voting and count-first at all six tested budgets. At 128 solutions, it gained +3.33 points [0.83, 6.67] over each.

Using the product of step scores instead of the minimum raised the gain to +7.50 points [3.33, 12.50].

4. With a weaker scorer, combining votes beats trusting one solution.

Best-of-N selects the single highest-scoring solution. With Math-Shepherd's minimum step score and 128 solutions, weighted voting beat it by:

Test set and generatorGain over best-of-N
362 unused AIME, Math-7B+13.26 [9.12, 17.13]
120 AIME, Phi-4+5.83 [0.83, 11.67]
146 earlier MATH-500 problems, Math-7B+6.85 [2.05, 11.64]
120 GSM-Hard problems, Math-7B+3.33 [0.83, 6.67]

5. Learned weights work on new problems and another generator.

We used separate development problems to choose how much each solution counts in the vote based on its score. We then tested the same voting rule on new problems, without changing it. Using products of step scores, the fitted rule beat its own count-first version at 40 solutions on both unused AIME batches: +2.78 [0.28, 5.83] and +2.20 [0.14, 4.41].

We chose a voting rule using separate Math-7B development problems. Compared with using each solution’s score directly as its vote weight, this rule improved accuracy by 1.93 percentage points [0.14, 3.86] on 242 new problems, using 40 solutions per vote.

We then applied the same rule unchanged to Phi-4’s solutions. It again beat using the scores directly, by 4.17 percentage points [0.83, 8.33], using 128 solutions per vote. Both tests combine each solution’s step scores using their geometric mean.

6. Controlled examples show why the theorem's condition matters.

We created six test cases. Across the cases, each answer appeared at the same rate, and solutions with a given score were correct equally often. Choosing the highest-scoring solution also gave the same accuracy.

We changed only which wrong answers received high scores. We then tested the same set of voting rules in every case. Theorem 1 correctly predicted when weighted voting would reduce errors faster than plain voting as more solutions were added. Thus, Knowing how scores relate to correctness is not enough; which wrong answers receive those scores also matters.

Revision. The ten-page main text now centers on two questions: when does overruling the vote help, and when is tie-breaking safe? Section 6 summarizes the findings; Appendix A.9 gives the latest experiments. Changes are blue; the formal theoretical results are unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.