acceptodds
Under review as a conference paper at ICLR 2027

Frustratingly Easy Majority Voting Using Length Conditioning

Abstract

Majority Voting (MV) is a simple and widely used strategy which improves the reasoning accuracy of Large Language Models (LLMs) by allocating additional inference-time computation. Specifically, in MV multiple Chain-of-Thoughts (CoTs) are independently sampled and the most frequent final answer is selected. In this work, we propose Length-Conditioned Majority Voting (LCMV), a Test- Time Training approach which improves MV leveraging an empirical observation that has recently been established in several studies: longer CoTs tend to be associated with a lower probability of correctness. We formulate this inductive bias as a length-conditioned prior probability, which is modeled by a simple logistic function. Moreover, we adapt this prior at test time using only unlabeled testing data and without fine-tuning the LLM. Finally, the adapted prior is used to weight the vote of each sampled answer. Extensive experiments across multiple LLMs and reasoning benchmarks show that LCMV consistently improves over both standard MV and existing length-based and confidence-based aggregation strategies, while retaining the simplicity and model-agnostic nature of MV. Our code is available in the Supplementary Material and will be published after this article is accepted.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Majority Voting, Reasoning LLMs, Test-Time Training, Parallel Test-Time Scaling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7W89OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes length-conditioned majority voting, which assigns each sampled answer a logistic weight based on its reasoning length. Two parameters start with a preference for shorter responses and are adapted across incoming questions using ordinary majority agreement as pseudo-labels. The method does not require future questions or correct-answer labels, but collects the current question's responses before updating and voting. The language model remains frozen. Experiments span four existing protocols, comparing unweighted voting, length-based selection, and confidence-based aggregation. Reported averages favor LCMV, and an ablation suggests an improvement over fixed length weighting.

Critical strengths
  1. The online adaptation setting is worthwhile. Updating a weighting rule as questions arrive, without access to future questions or correctness labels, is a meaningful constraint compared with fitting on a complete offline collection. Adapting only two scalars adds little computation once responses exist and requires no language-model fine-tuning. Relative to Shortest Majority Vote, the contribution is this online adaptation, rather than length weighting itself. A reliable improvement under these constraints would be practically useful.

  2. The evaluation includes relevant breadth and controls. The four protocols cover reasoning and instruction-tuned models, different sample counts, mathematics, and GPQA. Table 5 directly compares fixed, online, and offline weighting, rather than reporting only comparisons to ordinary majority voting. Repeating generation seeds and question orders is a sensible attempt to examine variability in the sequential setting.

Critical weaknesses
  1. The incremental value of online adaptation is not established convincingly enough. Table 5 is the most direct evidence: online weighting improves the three protocol averages over fixed weighting by 0.7, 1.0, and 1.5 points. These gains are promising, but their paired uncertainty is unclear. The reported 25 runs combine five generation seeds with five question orders per seed; separating those sources of variation matters when assessing reliability. The broader comparisons do not resolve this: Section 5.2 imports MV, FFS, and Shortest MV in Table 1, and all baselines in Table 3, from earlier publications, whereas LCMV uses newly generated responses. Matching nominal protocols does not ensure identical samples or parsing. Some baselines are reproduced, but the principal comparisons still need evaluation on the same responses. The experimental argument should center on whether adaptation reliably improves over fixed weighting and established inexpensive aggregators, with benchmark breadth supporting that conclusion.

  2. The learning signal needs a clearer interpretation and mechanism analysis. Equation 5 labels agreement with the sampled majority, which can itself be wrong. Learning that target does not automatically yield the correctness probability asserted in Equation 4. This is not a reason to dismiss the heuristic: reweighting can still change the winner and improve accuracy. However, the paper should show whether adaptation repairs incorrect majority answers more often than fixed weighting, and how often it damages correct ones. That analysis would explain when the online update contributes useful information rather than reinforcing the majority's existing preferences.

Questions for authors
  1. Contribution clarification: What is the paired improvement and uncertainty for Online versus Fixed LCMV in Table 5? How do MV and Shortest MV compare on exactly the same responses and grader?
  2. Mechanism clarification: When ordinary MV is incorrect, how often does adaptation recover the correct answer compared with fixed weighting, and how often does it damage a correct MV answer? Does the learned weight track actual correctness or only majority agreement?
  3. Online robustness clarification: How much variation comes from response generation versus question ordering, particularly the first question used for normalization? Does the benefit over fixed weighting persist across those choices?
3 comments
AuthorsOpenReview ↗

Answers to Reviewer 7W89

We thank the Reviewer for their work and feedback, and for appreciating our method, its unsupervised and online strategy and its usefulness, as well as for appreciating the diversity and breadth of our experimental evaluation. We provide responses to all questions below.

Q1: The incremental value of online adaptation is not established convincingly enough. Table 5 is the most direct evidence: online weighting improves the three protocol averages over fixed weighting by 0.7, 1.0, and 1.5 points. These gains are promising, but their paired uncertainty is unclear. The reported 25 runs combine five generation seeds with five question orders per seed; separating those sources of variation matters when assessing reliability. (...) What is the paired improvement and uncertainty for Online versus Fixed LCMV in Table 5?

A1: Following your suggestion, we report below Tab. 5 in which the results are computed with different combinations of random seeds. Specifically, the first column “5 seeds x 5 permutations” corresponds to the 25 runs used for each model and dataset in Tab. 1-3 and summarized in Tab. 5. On the other hand, “5 generative seeds” corresponds to using only 5 random seeds for the autoregressive CoT generation with temperature, without any question order permutation. Finally, “1 generative seed, 5 permutation seeds” indicates the use of only one random seed for generation and 5 random seeds for permutation.

5 generation seeds × 5 permutations (Online LCMV)5 generation seeds × 5 permutations (Fixed LCMV)5 generation seeds (Online LCMV)5 generation seeds (Fixed LCMV)1 generation seed × 5 permutation seeds (Online LCMV)1 generation seed × 5 permutation seeds (Fixed LCMV)
Ave. Tab. 158.9±0.658.2±0.758.6±0.958.4±0.859.6±0.358.9 ±0.7
Ave. Tab. 273.7± 1.372.7±1.273.8±1.773.1±2.073.5±0.572.3±0.9
Ave. Tab. 362.4± 0.460.9±0.562.4±0.561.0±1.262.9±0.261.2±0.6
Average65.0±0.563.9±0.564.9±0.964.2±1.265.3±0.364.1±0.6

The results in the table above show that there is no substantial difference in the use of multiple seeds for generation and/or permutation.

Q2: Section 5.2 imports MV, FFS, and Shortest MV in Table 1, and all baselines in Table 3, from earlier publications, whereas LCMV uses newly generated responses. Matching nominal protocols does not ensure identical samples or parsing. Some baselines are reproduced, but the principal comparisons still need evaluation on the same responses. (...) How do MV and Shortest MV compare on exactly the same responses and grader?

A2: Unfortunately, it is not very clear to us what you mean by "same responses and grader". If you mean that the comparison must be conducted using the same ground-truth answers, then, of course we use the same answers as the other approaches (as these are unique and no others exist). Furthermore, we have not only used the same answers, but—where available—we have also adopted the same string parsers, verifiers, prompts and generation hyperparameters proposed in the papers from which we drew our experimental protocols and against which we benchmarked our work (Sections 5.1, 5.2, and App. B). This, in fact, implicitly favors our competitors.

If, on the other hand, by "responses" you are referring to the CoTs generated by other methods, please note that these are not publicly available, and therefore we do not have access to them. This is, in fact, a quite common situation: authors which generate CoTs using non-deterministic sampling (i.e., with a temperature other than 0) very rarely make the CoTs generated by their models public. What is typically disclosed is the temperature value used, along with other details—such as the verifier and/or the prompt—which allow for a fair comparison of the resulting accuracy. For further details, please refer to App. B.

AuthorsOpenReview ↗

Answers to Reviewer 7W89 (2nd part)

Q3: Equation 5 labels agreement with the sampled majority, which can itself be wrong. Learning that target does not automatically yield the correctness probability asserted in Equation 4. This is not a reason to dismiss the heuristic: reweighting can still change the winner and improve accuracy. However, the paper should show whether adaptation repairs incorrect majority answers more often than fixed weighting, and how often it damages correct ones. That analysis would explain when the online update contributes useful information rather than reinforcing the majority's existing preferences. (...) When ordinary MV is incorrect, how often does adaptation recover the correct answer compared with fixed weighting, and how often does it damage a correct MV answer? Does the learned weight track actual correctness or only majority agreement?

A3: The simplest way to quantitatively measure how frequently our method repairs incorrect majority answers (and, conversely, how often it damages correct ones), is by using accuracy. In fact, by definition, accuracy is the number of correct responses relative to the total number of responses. Therefore, a higher accuracy for Online LCMV compared to MV shows precisely how often our approach improves the pseudo-labels obtained using Eq. (5). For the sake of brevity, we show below only the mean values, taken from Tab. 1, 2 and 3, calculated for each protocol by averaging across all the datasets and models used:

MVOnline LCMVFixed LCMV
Ave. Tab. 157.158.9±0.658.2±0.7
Ave. Tab. 271.873.7±1.372.7±1.2
Ave. Tab. 356.762.4±0.460.9±0.5
Ave. Tab. 665.867.4-
Average62.965.663.9

In the above table we also added the average accuracy of the first protocol (Fig. 2), obtained across all the values of kk and all the models, whose detailed results are reported in Tab. 6. This comparison shows that we consistently improve over MV, independently of the setting, demonstrating that our approach is able to improve the pseudo-labes computed with standard MV (Eq. (5)). Note that other methods do not always achieve an improvement over MV. For instance, when using protocol 1, proposed by Short-3@k, the latter underperforms compared to MV when using DeepSeek-R1-Distill-Qwen-32B (Fig. 2a and Tab. 6). In the third protocol (Tab. 2), proposed by FFS, FFS is, on average, inferior to MV. In the above table, we also added the average results of Fixed LCMV, which again show that our unsupervised adaptation strategy improves over the fixed-prior variant.

Furthermore, following your suggestion, we analyzed how often Online LCMV changes the aggregated MV decision (Eq. (2)) from correct to incorrect, and vice versa. To reduce the number of possible combinations, we used an example with k=3k = 3. Specifically, we used DeepSeek-R1-Distill-Qwen-32B tested on HMMT February 2025 from Tab. 6. Using k=3k = 3, we get 1,140 different subsets of responses for a single question (please see the description of Protocol 1 on lines 311–313). Given that HMMT February 2025 consists of 30 questions, the experiment was repeated for a total of 1,140 x 30 = 34,200 runs. In the table below, column “Response configuration” indicates the configuration of the votes used in Eq. (2). For instance, “3C” means 3 correct answers, “2C–1E” means 2 correct and 1 incorrect answer, while “1C–1E–1E” means 1 correct and 2 different, incorrect answers. Column “Number of runs” indicates the number of runs for which MV produced that configuration. Finally, “Incorrect → Correct” and “Correct → Incorrect” denote the number of times Online LCMV turned an incorrect MV decision into a correct one, and vice versa.

Response configurationNumber of runsIncorrect → CorrectCorrect → Incorrect
3C6,93800
2C–1E4,02604
1C–1E–1E5,0741,889693
2E–1C1,303100
3E1,50000
2E–1E4,99700
1E–1E–1E10,36200
Total34,2001,899697

In the example shown in the table above, most of the changes are focused on the configuration “1C–1E–1E”. In this configuration, since there is no winner, MV chooses at random, thus only one-third of the 5,074 runs, on average, will be correct. For the remaining two-thirds of incorrect decisions, Online LCMV corrects 1,889 runs while damaging 693 correct ones. Overall, Online LCMV corrects 1,899 decisions and damages 697. The difference, 1,202, when compared to the total (34,200), corresponds to an increase in accuracy of approximately 3.5%, as shown in Table 6 in the entry corresponding to this experiment.

AuthorsOpenReview ↗

Answers to Reviewer 7W89 (3rd part)

Q4: How much variation comes from response generation versus question ordering, particularly the first question used for normalization? Does the benefit over fixed weighting persist across those choices?

A4: As mentioned in Sec. 5.1 and 5.2, we randomly permute the question order to show that our method (both the Fixed and the Online versions) is not sensitive to this order, including the normalization based on the first question. This robustness is demonstrated by the low standard deviation obtained in Tab. 1, 2, and 3. For instance, in Tab. 1, the standard deviation for Online LCMV is lower than that of the other methods for which it was possible to repeat the experiments using different random seeds. The same result was obtained with the first experimental protocol (Fig. 2 and Table 6). Also in that case, the average standard deviation (Tab. 6) is lower for Online LCMV with respect to MV and Short-3@k in 12 and 13 out of 16 comparisons, depending on the adopted base LLM (lines 840-847). Note that in the first protocol, the question order is fixed but the experiments are based on randomly extracting subsets of kk CoTs drawn from a set of 20 generations, thus in every subset the first question is different (see App. C.1 for more details). Finally, please note that also Fixed LCMV was tested using the same random order, thus the benefit shown in Tab. 5 takes into account the first-question normalization (Sec. 5.3).

Reviewer_d7ypOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper weights majority voting by CoT length using a simple logistic function that is adapted online at test time on unlabeled questions.

Critical strengths
  1. Overall, I think the proposed method is reasonable and can be effective in some cases (though see Weaknesses 1 and 3). These cases are: (i) the model is strong enough that the majority-vote signal is offten acceptable; (ii) the difficulty of the questions is relatively consistent. Otherwise there is a mismatch—harder questions should naturally require longer reasoning.

  2. The experimental coverage is relatively broad, involving multiple models and reasoning tasks.

Critical weaknesses
  1. The method only works under limited conditions, as noted above. In particular, when questions in the test set have different difficulties, the method will fail.

  2. Importantly, the table formatting and result presentation are poor. Several main result tables are too crowded and lack professionalism.

  3. Related to Weakness 1, a reasoning chain is neither better when longer nor better when shorter. Using only logistic regression to model the length variable is not sufficiently reasonable. The method and abstract also only mention one side and ignore the other—that longer reasoning is not always better.

Questions for authors

Please refer to the above comments.

3 comments
AuthorsOpenReview ↗

Answers to Reviewer d7yp

We thank the Reviewer for their work and feedback, and for appreciating the proposed method and the broad scope of the experimental coverage. We provide responses to all questions below.

Q1: The method only works under limited conditions: (i) the model is strong enough that the majority-vote signal is often acceptable.

A1: Please, note that this condition is common to the entire Majority Voting paradigm, and, to some extent, to all parallel test-time scaling methods (Sec. 1). In fact, all the MV approaches are implicitly based on the assumption that each individual response has a probability of being correct that is higher than that of random chance because it is an ensemble method (Sec. 1). If the base model is not strong enough, all MV methods fail. However, this is a mild condition, as LLMs are often able to provide correct answers, as evidenced by the fact that the MV paradigm has been widely and successfully adopted for several years, both in scientific papers and in industrial work. Note also that using (standard) MV for collecting pseudo-labes (lines 10-11 of Alg. 1) is inspired by (Zuo et al. (2025)), a popular TTT method using MV-based pseudo-labes in conjunction with Reinforcement Learning.

Q2: Limited conditions: (ii) when questions in the test set have different difficulties, the method will fail. (...) harder questions should naturally require longer reasoning.

A2: We agree with the Reviewer that, if the relative differences in question difficulty (and the corresponding lengths of the CoTs) are extremely variable, our method—much like all other approaches proposed in the literature that rely on the same length-based inductive bias—most likely will struggle. However, the testing sets used in our experiments exhibit high internal variance in terms of difficulty and the corresponding length of the CoTs. For instance, Fig. 1a shows the distribution of CoT lengths for GPQA-Diamond, obtained using R1-Distill-Qwen-32B. These lengths range from 359 to 16,000 tokens, showing a very high variability of the response length. A similar variance is present in all other datasets used in this paper, and this same high variance has also been observed by other authors who empirically examined the same phenomenon before us (Sec. 1 and 2). Despite this high variance in CoT lengths, Online LCMV almost always succeeds in improving the accuracy of MV and outperforming the other state-of-the-art methods against which we compared it—including those that do not employ an inductive bias based on length (e.g., CER in Tab. 1 and 3).

Furthermore, following your suggestion, we have added a new section in the appendix (App. C.4) to evaluate the robustness of Online LCMV against a degree of variability in question difficulty that exceeds that of the benchmarks used in the main paper. More precisely, for each of the Tab. 1–3, we merged the respective datasets and randomly shuffled them, using 25 permutations. Unlike the permutations used in the main paper (which were applied intra-dataset), in this case, we permuted the entries of the entire merged dataset, resulting in sequences of questions randomly drawn from different datasets. For example, for Tab. 2, the merged and shuffled dataset randomly alternates questions drawn from AIME2024, AIME2025-I, AIME2025-II, and GPQA-Diamond (more details in App, C.4). In this way, during the execution of Alg. 1, the model is forced to adapt to an incoming stream of questions covering different topics (mathematics, biology, etc.), with a high variance of the question difficulty and the length of the corresponding CoTs. We report below the results.

ModelDatasetMVFFSShortest MVShort-3@kCEROnline LCMVOnline LCMV (merged dataset)
DeepSeek-R1-Distill-Qwen-32BAIMO72.861.973.770.4±2.674.4±2.475.3±0.875.1±0.8
DeepSeek-R1-Distill-Qwen-32BGPQA-Diamond63.361.263.564.1±0.561.7±0.765.0±1.064.9±1.0
DeepSeek-R1-Distill-Qwen-14BAIMO71.762.071.567.8±2.171.8±2.373.4±1.373.0±1.4
DeepSeek-R1-Distill-Qwen-14BGPQA-Diamond56.156.656.457.9±1.857.3±1.558.5±1.957.6±1.7
DeepSeek-R1-Distill-Qwen-1.5BAIMO40.026.242.238.9±2.840.7±3.542.7±1.741.7±2.2
DeepSeek-R1-Distill-Qwen-1.5BGPQA-Diamond29.527.730.229.6±1.529.0±1.729.1±2.129.1±2.4
QwQ-32B-PreviewAIMO51.340.850.846.4±2.853.3±1.451.3±1.351.1±1.0
QwQ-32B-PreviewGPQA-Diamond62.256.862.258.9±1.359.3±1.561.7±1.261.4±1.4
LIMOAIMO68.862.270.068.0±2.057.3±3.472.0±1.772.2±1.5
LIMOGPQA-Diamond55.550.155.855.1±1.454.0±1.559.8±1.059.6±1.1
Average57.150.557.655.7±0.755.9±0.958.9±0.658.6±0.7
AuthorsOpenReview ↗

Answers to Reviewer d7yp (2nd part)

ModelDatasetGreedyMVLFSFFSOnline LCMVOnline LCMV (merged dataset)
Phi-4-Reasoning-PlusAIME202471.780.080.076.783.3±4.383.0±4.3
Phi-4-Reasoning-PlusAIME2025-I66.773.380.066.776.0±5.076.3±5.4
Phi-4-Reasoning-PlusAIME2025-II76.793.386.786.779.4±7.978.4±8.2
Phi-4-Reasoning-PlusGPQA-Diamond62.962.770.267.266.0±0.566.1±0.6
DeepSeek-R1-Distill-Qwen-32BAIME202468.377.960.680.080.5±2.679.8±2.7
DeepSeek-R1-Distill-Qwen-32BAIME2025-I51.753.349.759.661.3±2.761.3±2.8
DeepSeek-R1-Distill-Qwen-32BAIME2025-II46.760.060.046.771.7±9.270.7±9.6
DeepSeek-R1-Distill-Qwen-32BGPQA-Diamond60.762.160.162.664.2±2.063.9±1.9
QwQ-32BAIME202479.283.580.181.284.9±3.284.6±3.5
QwQ-32BAIME2025-I60.069.953.559.969.3±5.469.0±5.4
QwQ-32BAIME2025-II69.281.970.478.083.2±3.983.7±4.4
QwQ-32BGPQA-Diamond59.764.757.665.264.5±1.464.2±1.5
Average64.571.867.469.273.7±1.373.4±1.6
ModelDatasetGreedyMVP(True)PENLNELLCEROnline LCMVOnline LCMV (merged dataset)
LLaMA-3.1-8B-InstructGSM8K82.889.687.685.286.286.283.890.090.5±0.890.3±0.9
LLaMA-3.1-8B-InstructMATH53.455.456.852.052.853.650.458.262.5±0.962.6±0.9
LLaMA-3.1-8B-InstructMathQA60.063.265.264.465.261.665.468.268.2±1.367.9±1.5
Mistral-7B-InstructGSM8K44.862.246.655.859.060.055.665.266.4±1.366.6±1.5
Mistral-7B-InstructMATH17.020.413.619.020.220.019.624.022.5±0.822.6±0.9
Mistral-7B-InstructMathQA20.220.812.422.620.019.422.622.624.7±1.424.5±1.3
Olmo-2-7B-InstructGSM8K84.285.082.084.483.878.084.888.888.1±0.288.3±0.4
Olmo-2-7B-InstructMATH37.842.540.041.040.039.242.648.054.2±0.754.0±0.6
Olmo-2-7B-InstructMathQA45.252.051.844.850.048.847.459.457.5±0.957.6±1.0
LLaMA-3.2-3B-InstructGSM8K75.278.473.273.077.078.675.282.685.8±0.885.9±0.9
LLaMA-3.2-3B-InstructMATH46.451.244.244.042.640.040.256.062.3±0.962.0±1.0
LLaMA-3.2-3B-InstructMathQA55.459.652.255.654.258.457.462.865.6±0.565.8±0.7
Average51.956.752.153.554.253.653.860.562.4±0.462.3±0.5

The results presented in the three above tables show that the decrease in accuracy—compared to the use of each dataset individually—is marginal. This further confirms the robustness of Online LCMV with respect to the variability of the difficulty of the questions and the length of the generated CoTs.

In App. C.4 we also added a new figure (Fig. 4) which is the analogue of Fig. 1, where we show the distributions of the correct and incorrect responses computed using a merged dataset. This figure, along with the corresponding analysis provided in App. C.4, helps explain why Online LCMV works even with such heterogeneous datasets. In fact, the two distributions in the example shown remain roughly unimodal in shape, and can therefore be (partially) discriminated using logistic regression.

Q3: Importantly, the table formatting and result presentation are poor. Several main result tables are too crowded and lack professionalism.

A3: Thank you for pointing this out. Unfortunately, the main result tables (Tab. 1, 2, and 3) were small in the submitted version of the paper due to space constraints. In the new version submitted alongside this rebuttal, we have enlarged them. Furthermore, please note that we have followed the official ICLR table style. Any additional suggestions regarding formatting are welcome.

AuthorsOpenReview ↗

Answers to Reviewer d7yp (3rd part)

Q4: Related to Weakness 1, a reasoning chain is neither better when longer nor better when shorter. Using only logistic regression to model the length variable is not sufficiently reasonable. The method and abstract also only mention one side and ignore the other—that longer reasoning is not always better.

A4: This question is not very clear to us. The fact that longer reasoning chains are sometimes correct is stated clearly, for instance, in Sec. 1 and 2, and it is illustrated in Fig. 1a. Importantly: this is our main motivation. In fact, the main difference between our proposal and most of the previous work is that we want to use all the k answers, without hard filtering the longer ones. This concept is expressed in lines 106–112, which we include here for the sake of completeness:

“However, we argue that these filtering methods are suboptimal because they completely discard the k − m longest sequences (m ≥ 1), for which the probability of being correct—despite being lower than that of the others—is nonetheless non-zero. This is illustrated in Fig. 1a: the tail of the correct response distribution extends below that of the incorrect answers. Consequently, it is possible for long responses to be correct, albeit with a lower prior probability. Hence, instead of filtering out long sequences, we propose to weight the vote of each response during the MV aggregation phase using a (learned) length-conditioned correctness probability.”

Moreover, in Sec. 5.2 we compare methods which discard long sequences with vote-weighting methods (like ours, which includes all sequences in a weighted vote), and we show that the latter empirically outperform the former, concluding that:

“These results support our hypothesis in Sec. 1: longer responses can still provide useful evidence, and using their weighted votes is more effective than discarding them.” (lines 392-393).

If, on the other hand, the Reviewer means that logistic regression alone is insufficient to determine the correctness of a response, we would like to point out that this is not what we do. In fact, our method is a weighted modification of MV, which, in turn, is based on consensus. Our decision is taken using Eq. (6), which is based on the logistic function and the aggregated consensus.

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

This paper proposes using sequence length as a prior to improve majority voting in reasoning LLMs. Specifically, the authors introduce an online, unsupervised, probabilistic method that exploits this prior.

Critical strengths
  1. Good writing.
  2. This paper shows that, for certain LLMs, trajectory length correlates with correctness in mathematical reasoning.
Critical weaknesses
  1. The length–correctness relationship is weak and only partially characterized. The relationship is in fact non-monotonic — short sequences are more likely to be correct and medium-length ones more likely to be incorrect — so length carries signal only in part of the range. Moreover, this observation is established only for mathematical reasoning; whether it extends to other tasks is unknown, which limits the generality of the claimed prior.

  2. The baseline in Eq. (3) is fixed rather than adaptive. Eq. (3) appears to derive its baseline solely from the first question's length. Why not update the baseline online as samples are collected? An online estimate would seem both more natural under the paper's streaming assumption and more robust to the heavy variance of early samples.

  3. The empirical gains appear marginal. Fig. 2 shows an improvement of roughly one percentage point over majority voting, and in the ablation, the fixed-LCMV and offline-LCMV variants differ by less than one percentage point. Differences of this size are difficult to distinguish from run-to-run variance.

  4. The limit of the length prior is not established. I appreciate that the authors operate under a deliberately limited unsupervised setting. It would nonetheless be informative to probe the ceiling of the length inductive bias in a supervised setting: pre-collect labeled data and fit the logistic model on it, to see how much of the remaining gap is attributable to the online/unsupervised constraint rather than to the length prior itself. Such a comparison would clarify whether the method's limitations stem from the prior or from the setting.

Questions for authors

See Critical Weaknesses.

4 comments
AuthorsOpenReview ↗

Answers to Reviewer xbzZ

We thank the Reviewer for their work and feedback, and we provide responses to all questions below.

Q1: The length–correctness relationship is weak and only partially characterized. The relationship is in fact non-monotonic — short sequences are more likely to be correct and medium-length ones more likely to be incorrect — so length carries signal only in part of the range.

A1: We completely agree with the Reviewer on this point and, in fact, this is our main motivation, which differentiates our proposal from most of the previous work in this field. In more detail, the fact that length–correctness relationship is weak is implicit in the concept of (negative) correlation, and it is largely emphasized, e.g., in Sec. 1 and Sec. 2, and illustrated in Fig. 1a, which shows that correct and incorrect distributions largely overlap. Please note that, of course, our decision criterion is not based on length alone. In fact, our decision criterion is a (weighted modification) of the Majority Voting (MV) consensus. In other words, the decision is taken aggregating the consensus of multiple, independent answers. Our proposal differs from the widely adopted MV criterion by exploiting the length–correctness (negative) correlation to include an inductive bias in this aggregation mechanism (Eq. (6)). Importantly, our main motivation in proposing a weighted voting mechanism is that long sequences should not be discarded. In fact, most of the previous works which exploit our same inductive bias simply discard the k−mk-m longest answers (Sec. 1 and 2). Conversely, we use all the kk answers, exactly because, as you rightly point out, the relationship between length and correctness is weak. Consequently, it is possible for some correct answers to be longer than incorrect ones. This concept is expressed, e.g., in lines 107–112, which we include here for the sake of completeness:

“However, we argue that these filtering methods are suboptimal because they completely discard the k−mk − m longest sequences (m≥1m ≥ 1), for which the probability of being correct—despite being lower than that of the others—is nonetheless non-zero. This is illustrated in Fig. 1a: the tail of the correct response distribution extends below that of the incorrect answers. Consequently, it is possible for long responses to be correct, albeit with a lower prior probability. Hence, instead of filtering out long sequences, we propose to weight the vote of each response during the MV aggregation phase using a (learned) length-conditioned correctness probability.”

Moreover, in Sec. 5.2 we compare length-based filtering methods (which discard long sequences) with vote-weighting methods (like ours, which includes all sequences in a weighted vote), and we show that the latter empirically outperform the former, concluding that:

“These results support our hypothesis in Sec. 1: longer responses can still provide useful evidence, and using their weighted votes is more effective than discarding them.” (lines 392-393).

Q2: Moreover, this observation is established only for mathematical reasoning; whether it extends to other tasks is unknown, which limits the generality of the claimed prior.

A2: We respectfully disagree with the Reviewer on this point. For instance, Fig. 1a was computed using the GPQA-Diamond dataset, which contains questions on biology, physics, and chemistry, but not math. Please, note that we used GPQA-Diamond in our experiments (e.g., Tab. 1, 2, 4 and 5). Moreover, the generality of our inductive bias has been amply demonstrated by a long series of previous works, some of which are mentioned in lines 52–100 and in Sec. 2. Please, note that many of these studies use heterogeneous domains and datasets in their experiments—not just mathematical ones. To cite a few (non-exhaustive) examples of works using also non-mathematical datasets:

  • Fatemi et al. (2025) use also MMLU-STEM
  • Liu et al. (2025) use also Olympiad Bench
  • Michael et al. (2025) use also GPQA-Diamond
  • Aradhye et al. (2025) use also GPQA-Diamond
  • Yi et al. (2025) use also Olympiad Bench, MMLU and LiveCodeBench
  • Devic et al. (2025) use also NonAmbigQA, MMLU, MediQ, MedMCQA, SciQ, MedQA, MMLU-Pro-NoMath, SuperGPQA and others
  • Parascandolo et al. (2026) use also MMLU-STEM
  • Vanhoyweghen et al. (2026) use also HLE.
AuthorsOpenReview ↗

Answers to Reviewer xbzZ (2nd part)

Q3: The baseline in Eq. (3) is fixed rather than adaptive. Eq. (3) appears to derive its baseline solely from the first question's length. Why not update the baseline online as samples are collected? An online estimate would seem both more natural under the paper's streaming assumption and more robust to the heavy variance of early samples.

A3: The reason why we adopted a normalization depending only on the first question (Eq. (3)) is explained in lines 238-242. Specifically, given that standardization is commonly used to stabilize learning by fixing the scale of input values, recalculating its parameters for each question would lead to a variation of the standardized values (l’l’) rather than a normalization. Conversely, using only the first question allows us to keep the standardization parameters fixed. To empirically test the robustness of our method with respect to this choice, in the experiments in Tab. 1 through Tab. 5 we use 25 random permutations of the questions’ order (lines 347-348). For instance, in lines 368-372, we show that the average standard deviation of our results in Tab. 1 is lower than the other methods for which we could repeat the experiments with random seeds. The same results are confirmed in App. C.1 (lines 840-847) for the experiments of Fig. 2, where again our method gets a standard deviation lower than the competitors. This demonstrates that our approach is not sensitive to the order in which the questions arrive, including the first question used for normalization.

Nevertheless, we agree with the Reviewer that this is an interesting ablation experiment. Following your suggestion, we modified Alg. 1 removing the condition “if j=1j = 1” in line 6 and, hence, recomputing the standardization parameters μ\mu and ss for each specific question qjq_j. Using this online normalization, we repeated all the main experiments (Fig. 2 and Tab. 1-3). The details are shown in the new Appendix section C.3 (please, see the new version of our manuscript). We report below the results of this online normalization compared to the values of Tab.1-3 and we refer to App. C.3 for the graph corresponding to Fig. 2.

ModelDatasetMVFFSShortest MVShort-3@kCEROnline LCMVOnline Normalization LCMV
DeepSeek-R1-Distill-Qwen-32BAIMO72.861.973.770.4±2.674.4±2.475.3±0.874.8±1.0
DeepSeek-R1-Distill-Qwen-32BGPQA-Diamond63.361.263.564.1±0.561.7±0.765.0±1.064.8±1.1
DeepSeek-R1-Distill-Qwen-14BAIMO71.762.071.567.8±2.171.8±2.373.4±1.373.3±1.3
DeepSeek-R1-Distill-Qwen-14BGPQA-Diamond56.156.656.457.9±1.857.3±1.558.5±1.958.3±1.6
DeepSeek-R1-Distill-Qwen-1.5BAIMO40.026.242.238.9±2.840.7±3.542.7±1.741.6±2.1
DeepSeek-R1-Distill-Qwen-1.5BGPQA-Diamond29.527.730.229.6±1.529.0±1.729.1±2.129.1±2.0
QwQ-32B-PreviewAIMO51.340.850.846.4±2.853.3±1.451.3±1.351.4±1.2
QwQ-32B-PreviewGPQA-Diamond62.256.862.258.9±1.359.3±1.561.7±1.261.6±1.3
LIMOAIMO68.862.270.068.0±2.057.3±3.472.0±1.771.7±1.8
LIMOGPQA-Diamond55.550.155.855.1±1.454.0±1.559.8±1.059.4±1.0
Average57.150.557.655.7±0.755.9±0.958.9±0.658.6±0.5
ModelDatasetGreedyMVLFSFFSOnline LCMVOnline Normalization LCMV
Phi-4-Reasoning-PlusAIME202471.780.080.076.783.3±4.383.3±4.3
Phi-4-Reasoning-PlusAIME2025-I66.773.380.066.776.0±5.075.7±5.4
Phi-4-Reasoning-PlusAIME2025-II76.793.386.786.779.4±7.978.7±7.9
Phi-4-Reasoning-PlusGPQA-Diamond62.962.770.267.266.0±0.565.9±0.5
DeepSeek-R1-Distill-Qwen-32BAIME202468.377.960.680.080.5±2.680.5±2.8
DeepSeek-R1-Distill-Qwen-32BAIME2025-I51.753.349.759.661.3±2.761.3±2.7
DeepSeek-R1-Distill-Qwen-32BAIME2025-II46.760.060.046.771.7±9.270.9±9.4
DeepSeek-R1-Distill-Qwen-32BGPQA-Diamond60.762.160.162.664.2±2.063.8±1.8
QwQ-32BAIME202479.283.580.181.284.9±3.284.3±3.7
QwQ-32BAIME2025-I60.069.953.559.969.3±5.469.3±5.4
QwQ-32BAIME2025-II69.281.970.478.083.2±3.984.0±3.3
QwQ-32BGPQA-Diamond59.764.757.665.264.5±1.464.3±1.4
Average64.571.867.469.273.7±1.373.5±1.5
AuthorsOpenReview ↗

Answers to Reviewer xbzZ (3rd part)

ModelDatasetGreedyMVP(True)PENLNELLCEROnline LCMVOnline Normalization LCMV
LLaMA-3.1-8B-InstructGSM8K82.889.687.685.286.286.283.890.090.5±0.890.6±0.7
LLaMA-3.1-8B-InstructMATH53.455.456.852.052.853.650.458.262.5±0.963.3±1.0
LLaMA-3.1-8B-InstructMathQA60.063.265.264.465.261.665.468.268.2±1.368.5±1.3
Mistral-7B-InstructGSM8K44.862.246.655.859.060.055.665.266.4±1.366.7±1.4
Mistral-7B-InstructMATH17.020.413.619.020.220.019.624.022.5±0.822.6±0.9
Mistral-7B-InstructMathQA20.220.812.422.620.019.422.622.624.7±1.425.0±1.5
Olmo-2-7B-InstructGSM8K84.285.082.084.483.878.084.888.888.1±0.288.2±0.3
Olmo-2-7B-InstructMATH37.842.540.041.040.039.242.648.054.2±0.754.6±0.6
Olmo-2-7B-InstructMathQA45.252.051.844.850.048.847.459.457.5±0.958.1±1.1
LLaMA-3.2-3B-InstructGSM8K75.278.473.273.077.078.675.282.685.8±0.886.1±0.8
LLaMA-3.2-3B-InstructMATH46.451.244.244.042.640.040.256.062.3±0.962.6±0.8
LLaMA-3.2-3B-InstructMathQA55.459.652.255.654.258.457.462.865.6±0.566.2±0.7
Average51.956.752.153.554.253.653.860.562.4±0.462.7±0.4

As expected, the results of the online normalization are slightly lower than those obtained by computing the standardization parameters using only the first question. However, the difference is small and this variant still manages to outperform the other methods.

Q4: The empirical gains appear marginal. Fig. 2 shows an improvement of roughly one percentage point over majority voting, and in the ablation, the fixed-LCMV and offline-LCMV variants differ by less than one percentage point. Differences of this size are difficult to distinguish from run-to-run variance.

A4: Please, note that the gains are actually a bit higher, and that the number of runs—for each of the 8 datasets and 11 different LLMs—ranges from 25 experiments (in Tables 1–5) to thousands or even hundreds of thousands for the experiments in Figure 2, thus the computed averages are quite robust. Specifically, we report the exact numerical values corresponding to the curves plotted in Fig. 2 in App. C.1, Tab. 6. Depending on kk, the improvement of our method with respect to MV ranges from 0.7 to 2.7 in Fig 2a and from 1.6 to 2.4 in Fig 2b, with an average of 1.3 and 1.9 points, respectively. Moreover, these values ​​represent the average obtained by repeating the experiments using, for each kk, all subsets of kk responses contained within the set of 20 generated responses (lines 311-313). For instance, with k=3k=3, we repeated the experiments using 1,140 runs, and, with k=10k=10, we repeated the experiments 184,756 times. For each value of kk and for each model, our method consistently outperforms both MV and Short-3@k. Therefore, the advantage of Online LCMV is sharp and consistent, and is not due to statistical fluctuations.

Similarly, the ablation experiments shown in Tab. 5 are the average of all the experiments reported in Tab. 1, 2 and 3, including all the datasets and the LLMs. The full results are reported in Tab. 11, 12 and 13. Please, note that each of the accuracy values ​​for Online LCMV reported in Tab. 1, 2, and 3 represents, in turn, the average of 25 runs with different random seeds (lines 346-349). Thus, for instance, the average gain of 1.1 points reported in Tab. 5 of Online LCMV (65.0) with respect to Fixed LCMV (63.9), which is higher than the standard deviation (0.5), is obtained considering all the experiments of Protocols 2-4, all the random seeds and all the datasets and the tested LLMs. We have emphasised this in the new version of our manuscript (lines in blue, 470-473).

In general, we agree with the Reviewer that, in some cases, the improvements over the competing methods examined are small. However, we adopted experimental protocols taken from 4 different state-of-the-art approaches (Sec. 5.1), each protocol being based on the specific testing scenario of the protocol’s proposer (e.g., number of generated answers kk, prompts, answer extraction and verifiers, sampling hyperparameters, token budgets, etc.). In each scenario, the second-best result (after ours) was achieved by a different approach (see the averages reported in Tab. 6, 1, 2 and 3). This suggests that, unlike our method, the other approaches fail to guarantee consistent gains across varying scenarios. We believe that, overall, these results—obtained using 8 datasets and 11 different LLMs—demonstrate the robustness and generality of our method, which consistently outperforms the state-of-the-art solutions, regardless of the specific experimental scenario.

AuthorsOpenReview ↗

Answers to Reviewer xbzZ (4th part)

Q5: It would be informative to probe the ceiling of the length inductive bias in a supervised setting: pre-collect labeled data and fit the logistic model on it, to see how much of the remaining gap is attributable to the online/unsupervised constraint rather than to the length prior itself. Such a comparison would clarify whether the method's limitations stem from the prior or from the setting.

A5: We thank the Reviewer for this suggestion. We added a new Appendix section in the revised version of our manuscript (App. C.5) dedicated to these experiments. Specifically, we used the fourth experimental protocol (Tab. 3) because it is the only one whose datasets contain a training split. For each dataset, we used the training split to fit the parameters of our logistic regression (w0w_0 and w1w_1), this time using supervised data. Since we need positive and negative answers, and the latter are not provided with the training data (training data only contains questions associated with correct CoTs), we generate responses which are then labeled using the same verifier used for testing. In more detail, for each training question qq, at training time we use the LLM to generate k=10k = 10 CoTs. The verifier associates a label (positive or negative) to each generated answer. Then, we use these labels to train w0w_0 and w1w_1. Finally, the parameters are frozen and the logistic function is tested on the testing split using Eq. (6) to aggregate consensus. We report below the results.

ModelDatasetGreedyMVP(True)PENLNELLCEROnline LCMVSupervised LCMV
LLaMA-3.1-8B-InstructGSM8K82.889.687.685.286.286.283.890.090.5±0.890.6±0.9
LLaMA-3.1-8B-InstructMATH53.455.456.852.052.853.650.458.262.5±0.962.9±1.4
LLaMA-3.1-8B-InstructMathQA60.063.265.264.465.261.665.468.268.2±1.368.4±1.3
Mistral-7B-InstructGSM8K44.862.246.655.859.060.055.665.266.4±1.366.5±1.7
Mistral-7B-InstructMATH17.020.413.619.020.220.019.624.022.5±0.822.4±0.8
Mistral-7B-InstructMathQA20.220.812.422.620.019.422.622.624.7±1.425.2±1.3
Olmo-2-7B-InstructGSM8K84.285.082.084.483.878.084.888.888.1±0.288.1±0.4
Olmo-2-7B-InstructMATH37.842.540.041.040.039.242.648.054.2±0.754.4±0.9
Olmo-2-7B-InstructMathQA45.252.051.844.850.048.847.459.457.5±0.958.2±1.0
LLaMA-3.2-3B-InstructGSM8K75.278.473.273.077.078.675.282.685.8±0.886.0±0.9
LLaMA-3.2-3B-InstructMATH46.451.244.244.042.640.040.256.062.3±0.962.5±0.9
LLaMA-3.2-3B-InstructMathQA55.459.652.255.654.258.457.462.865.6±0.566.0±0.6
Average51.956.752.153.554.253.653.860.562.4±0.462.6±0.5

The table above shows that Online LCMV and Supervised LCMV perform nearly on par, indicating that—at least in this setting—the proposed unsupervised training approach is able to model the length-based probabilistic prior quite effectively. Note that, in the same table, the average accuracy of Online LCMV (62.4) is 5.7 points higher than that of MV (56.7).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Answers common to all Reviewers

We thank all the Reviewers for their work and their comments. We submitted a new version of our paper where the main changes are:

  • Tables 1, 2, and 3 have been enlarged to improve their readability.

  • We have added three appendix sections (C.3, C.4, and C.5), along with corresponding tables and figures, to present the results of new experiments suggested by the Reviewers.

All changes are underlined in blue.

Below we answer to all the questions of the Reviewers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.