acceptodds
Under review as a conference paper at ICLR 2027

When to Vote, When to Rewrite: Disagreement-Guided Strategy Routing for Test-Time Scaling

Abstract

Large Reasoning Models (LRMs) achieve strong performance on mathematical reasoning tasks but remain unreliable on challenging instances. Existing test-time scaling methods, such as repeated sampling, self-correction, and tree search, improve performance at the cost of increased computation, yet often exhibit diminishing returns on hard problems. We observe that output disagreement is strongly correlated with instance difficulty and prediction correctness, providing a useful signal for guiding instance-level strategy selection at test time. Based on this insight, we propose a training-free framework that formulates test-time scaling as an instance-level routing problem, rather than allocating more computation within a single strategy, dynamically selecting among different scaling strategies based on output disagreement. The framework applies lightweight resolution for consistent cases, majority voting for moderate disagreement, and rewriting-based reformulation for highly ambiguous instances. Experiments on seven mathematical benchmarks and three models show that our method improves accuracy by 3%–7% while reducing sampling cost compared to existing approaches.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Mathematical Reasoning, Test-Time Scaling, Adaptive Routing

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_5mkdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a simple workflow that treats test-time scaling as an instance-level strategy routing problem. It uses answer consistency across stages to choose early stopping, voting, or rewriting for each problem. The method needs no extra training and is easy to integrate. Experiments across multiple models and math benchmarks report average accuracy gains of about 3–7 percentage points over strong baselines, with lower inference costs.

Critical strengths
  1. The workflow is simple, needs no extra training, and is easy to integrate. This gives the method practical value.
  2. Experiments across multiple models and math benchmarks report average accuracy gains over other strong baselines, with lower inference costs. The results suggest practical benefits in both accuracy and cost.
Critical weaknesses
  1. I find the novelty limited. The method combines consistency checks, voting, and rewriting in a staged routing workflow. Its added contribution over existing methods needs clarification. Table 4 tests budgets of 4, 6, and 8 generations. It remains unclear how further sampling of the original problems compares with rewriting at larger budgets on a fixed high-disagreement subset.

  2. The benefits of rewriting and the reliability of the comparisons need clarification. In Table 3, OnlyRewrite scores 90.4, 59.0, and 56.7 on Math500, Olympiad, and AIME24, respectively. These scores exceed SCoP's 88.2, 52.6, and 40.0 in Table 1. Table 1 reports the standard deviation of the average score across three random seeds, but the evaluation protocol remains unclear. It is unclear whether these experiments use consistent evaluation settings.

Also, some text in Figures 2–4 is too small to read clearly.

Questions for authors
  1. Do the main experiments and ablation studies adopt consistent evaluation settings (particularly for AIME24/25)? Does the gain stem from fluctuations in the evaluation? Will a higher number of sampling iterations n (e.g., 16, 32, 64) yield different results?

  2. Appendix C requires “remove unnecessary descriptions” and "simplify", yet the author attributes some of the errors to over-compression. How does the method perform with a paraphrasing prompt that does not require compress text?

Reviewer_UGNaOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a training-free framework for test-time scaling that routes problem instances to different inference strategies based on disagreement among sampled answers. The framework consists of three stages: Disagreement Filter, Vote Resolve, and Rewrite & Rethink. Instances with consistent predictions are directly accepted, those showing limited disagreement are resolved through voting, and those with persistent disagreement undergo problem rewriting followed by re-reasoning. Experiments across mathematical reasoning and code generation benchmarks report improvements in accuracy and computational efficiency over several existing approaches.

Critical strengths
  1. The proposed framework is simple and training-free, making it easy to integrate into existing reasoning models without additional training.

  2. The paper reports encouraging empirical results across multiple benchmarks and models, together with analyses of inference cost and the relationship between output disagreement and prediction accuracy.

Critical weaknesses
  1. This paper does not provide particularly novel or insightful perspectives, as the key ideas presented have already been explored in existing studies on test-time scaling.

  2. The proposed strategy routing mechanism is overly simplistic, and the claimed voting component is algorithmically redundant. The main methodological contribution appears limited to a heuristic combination of existing techniques, rather than a principled strategy-selection mechanism. More importantly, Algorithm 1 reveals that the Vote Resolve stage does not actually perform meaningful aggregation across the four sampled answers. Specifically, this stage is activated only when the first two answers disagree (a1≠a2a_1 \neq a_2) and the second pair agrees (a3=a4a_3 = a_4). Under these conditions, the plurality vote over all four answers necessarily returns a3a_3, regardless of the values of a1a_1 and a2a_2. Thus, the proposed voting operation is functionally equivalent to directly accepting the second agreeing pair.

  3. There appear to be substantial discrepancies between the partition-level statistics and the overall experimental results. For example, on Math500 with Qwen3-8B, Table 6 reports NDS/MDS/SDS proportions of 66.4%, 10.0%, and 23.6%, respectively. Combining these proportions with the corresponding strategy accuracies in Table 9 yields an expected overall accuracy of approximately 85.5%, whereas Table 1 reports 92.8%. Similar discrepancies arise on Gaokao En and Olympiad. Moreover, combining the partition proportions in Table 6 with the conditional baseline accuracies in Table 8 does not reproduce the baseline accuracies reported in Table 1. Although these discrepancies could reflect different experimental runs or evaluation protocols, the paper does not provide sufficient information to reconcile them.

In addition, the experimental comparisons do not adequately establish the specific contribution of disagreement-guided routing. Table 3 shows that rewriting alone already accounts for a substantial portion of the reported improvement. Although Table 9 compares strategies across disagreement-based partitions, the paper lacks controlled comparisons against alternative routing rules under matched computational budgets, such as random routing, entropy-based routing, or confidence-based routing. Consequently, it remains difficult to determine whether the observed improvements arise from the proposed routing mechanism itself or primarily from the existing rewriting strategy and differences in computational allocation.

Questions for authors
  1. In Algorithm 1, when a1≠a2a_1 \neq a_2 and a3=a4a_3 = a_4, majority/plurality voting over the four answers necessarily selects a3a_3. What additional benefit does the voting operation provide compared with directly accepting the second agreeing pair? How does this support the claimed contribution of the Vote Resolve stage?

  2. Could the authors clarify whether Tables 6, 8, and 9 use the same instance partitions, model generations, and evaluation settings as Table 1? In particular, why do the weighted conditional accuracies derived from the partition-level statistics differ substantially from the reported overall accuracies? Please provide the exact evaluation protocol and corresponding instance-level statistics.

1 comment

Dear Reviewer UGNa,

Thank you for your careful review. We clarified that direct selection and plurality voting are equivalent in the default two-round case, while the second pair still provides the consistency check that determines whether an instance is resolved or routed to rewriting. We also explained that Tables 6–9 are independent diagnostic runs, reported exact route-level statistics from an end-to-end run, and added a matched random-routing baseline controlling candidate outputs, route proportions, strategies, and computation. Disagreement-guided routing improves average accuracy by 6.7 points under this control. These clarifications and analyses have been incorporated into the revised PDF.

We would appreciate your assessment of whether these results address the concerns and would be grateful if you could reconsider the rating. We remain happy to clarify any outstanding issues.

Thank you for your time.

Reviewer_W8SqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a simple but effective test-time scaling method consisting of three stages: (1) sample two answers and check whether they agree; (2) if the first two answers disagree, sample two additional answers and check their agreement, applying majority voting if they agree; (3) if both stages fail to reach agreement, classify the problem as challenging and apply problem rewriting before generating a final answer. The method improves performance on various math benchmarks with Qwen3-4B, Qwen3-8B, and DeepSeek-R1-Distill-Llama-8B, and also demonstrates some generalization to coding benchmarks.

Critical strengths
  1. The method is effective on math benchmarks, achieving substantial average accuracy gains while reducing inference cost.
  2. The ablation studies of the different stages and the analysis of rewriting applicability are helpful.
Critical weaknesses
  1. The novelty of the proposed method is relatively marginal as it mainly combines self-consistency with problem rewriting, both techniques well studied by prior work.
  2. The baseline comparisons are relatively limited. I would like to see broader comparisons with prior work on adaptive inference-budget allocation for test-time scaling, as the proposed method shares a similar underlying motivation of applying more advanced test-time scaling strategies only when needed.
  3. From the analysis, rewriting appears to primarily address hidden-information errors, a very specific type of failure. This raises concerns about how well the method generalizes to other domains, especially given that the improvements on coding benchmarks appear smaller than those on math benchmarks.
  4. Some design choices, such as using two rounds of voting and two samples per round, appear somewhat ad-hoc and may not scale effectively. It would be valuable to explore a more general formulation that scales effectively with larger inference budgets.
  5. It would be valuable to see how the method generalizes to larger models, such as GPT or Claude, on more challenging benchmarks.
Questions for authors

n/a

2 comments

Dear Reviewer W8Sq,

Thank you for your careful review. We have clarified that the contribution lies in instance-level routing across TTS strategies rather than combining voting and rewriting. The Dynamic Voting comparison shows that our method improves both accuracy and efficiency over adaptive allocation within a single strategy, while the Pass@6 analysis separates the benefit of routing from simply increasing the sampling budget. The coding experiments examine cross-domain generalization, the 4/6/8-call results demonstrate extension to different budgets, and the Qwen3-32B and GPQA-Diamond evaluations provide evidence on larger models and more challenging tasks. The corresponding clarifications have also been incorporated into the revised PDF.

We would appreciate your assessment of whether these results address the concerns and would be grateful if you could reconsider the rating. We remain happy to clarify any outstanding issues.

Thank you for your time.

Thank you for the response. While the larger-scale experiments and increased call budgets are helpful, I remain concerned about the limited novelty, gains being focused on hidden-information settings, and lack of discussion of broader applicability. I therefore maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.