acceptodds
Under review as a conference paper at ICLR 2027

When Should Self-Consistency Trust Trajectory Signals? Competition-Aware Reselection for LLM Reasoning

Abstract

Self-consistency (SC) is a strong and widely used aggregation method for LLM reasoning, improving accuracy by majority voting over multiple sampled trajectories. We find that SC selection becomes substantially less reliable under strong competition between the two leading answer clusters, where additional evidence is needed to distinguish the correct candidate. By examining probability-based trajectory signals, we further find that correct answer clusters tend to have higher mean trajectory signals than competing incorrect clusters. Based on these findings, we propose CARS (Competition-Aware Reselection for Self-Consistency), which uses trajectory-level signals to revise SC’s selection between the Top-2 candidates only under strong competition. Across the evaluated models and mathematical and scientific reasoning benchmarks, CARS achieves improvements over Vanilla SC and Trajectory-Signal-Weighted Self-Consistency on the vast majority of tasks. Further cross-model, cross-benchmark, and sensitivity analyses show that the competition-dependent gain pattern remains robust across settings and is not tied to a particular trajectory signal or gating threshold.

Primary area:
generative models
Keywords:
LLM Reasoning, Self Consistency, Test-Time Scaling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Le6EOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CARS combines self-consistency with trajectory scores through a simple rule: retain the majority answer unless the two leading candidates have similar support and the runner-up has a higher mean score. Using the same sampled responses for all methods, it improves pooled accuracy by about one percentage point across three model families. The analysis shows why trajectory scores can help in close contests but cause harm when applied indiscriminately.

Critical strengths
  • S1. The main insight is clear and useful: auxiliary evidence need not be trusted equally on every problem. The resulting method is easy to understand and adds no new generation passes.

  • S2. The unconditional-reselection baseline makes a convincing case that selective intervention matters. Reporting corrected and newly broken answers helps explain the accuracy changes.

  • S3. All methods use the same responses, and the paper includes multiple model families and threshold sensitivity analyses. These are meaningful controls for a lightweight answer-selection method.

Critical weaknesses
  • W1. The separation between developing the selection rule and validating its transfer is not sufficiently clear. Tail-30% certainty is chosen following a preliminary comparison, while the competition threshold is introduced as an empirical operating point. The shared threshold, positive gains across three families, and sensitivity analyses are useful evidence, but the paper does not clearly identify which data informed these choices and which were reserved for evaluating the complete fixed rule. A held-out evaluation and paired uncertainty estimates would make the transfer claim easier to assess. The concern is the validation protocol, rather than the absolute number of additional correct answers.

  • W2. The incremental value of the specific hard-gating rule over stronger lightweight aggregation alternatives is not yet sufficiently isolated. Prior work has already explored alternative ways of combining answer frequency with trajectory-level confidence or certainty. For example, CISC uses a softmax-weighted combination of confidence scores and answer frequency, while self-certainty Borda voting combines frequency and certainty through rank-based weighting. Separately, RISC uses a candidate-to-leader support ratio as one feature in a learned candidate-ranking framework; for the runner-up candidate, this quantity corresponds to q2/q1q_2/q_1. These precedents do not establish that CARS’s exact Top-2 hard-gating rule already exists, and CARS’s simple competition-conditioned formulation remains distinct. However, the current comparisons against SC, UTRS, and raw-score TSWS show an advantage over the tested alternatives without fully isolating whether the gain specifically requires the proposed hard gate, rather than a stronger lightweight combination of candidate support and trajectory evidence. One representative comparison under a common development protocol would better substantiate the incremental contribution of the aggregation rule.

Questions for authors
  • Q1. If the signal and threshold are fixed using separate development data, does CARS still improve on self-consistency on held-out problems or benchmarks? This would directly test whether the proposed operating rule transfers.

  • Q2. How does CARS's fixed Top-2 gate differ from RISC's use of relative support in learned candidate ranking? Could the authors clarify which contribution lies in the competition-conditioned empirical finding and which lies in the particular decision rule?

  • Q3. Could the authors compare against one representative lightweight alternative, such as CISC-style softmax weighting or Borda rank weighting, using the same responses and trajectory signal? A common development protocol would help isolate the aggregation rule from differences in signal quality or tuning. A same-signal adaptation should be distinguished from a full reproduction of the original method.

  • Q4. How sensitive is CARS to the string-based answer clustering strategy? The appendix states that trajectories are clustered solely by parsed final-answer strings, while semantic or mathematical equivalence is used only for correctness evaluation. As a result, equivalent answers expressed in different forms may remain separate voting clusters. Because both Vanilla SC and CARS’s competition statistic R=q2/q1R=q_2/q_1 depend directly on these cluster counts, could the authors report how often equivalent answers are split across clusters, or provide a sensitivity analysis using normalized or equivalence-aware clustering?

  • Q5. What typically distinguishes the answers CARS rescues from those it incorrectly overturns? A few contrasting examples could clarify when the trajectory signal adds useful evidence and when it reinforces a plausible but wrong solution.

  • Q6. Does the benefit remain at smaller sampling budgets? Reusing subsets of the existing responses would help show whether the competition rule is useful when fewer than 128 trajectories are available.

Reviewer_5EY2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CARS modifies self-consistency by consulting trajectory scores only when the two leading answer clusters have similar vote counts. It selects the runner-up when that cluster has a higher mean score and otherwise retains the majority answer. The default score measures certainty over the final 30% of each trajectory. With 128 trajectories per question, CARS improves pooled accuracy by 0.86–1.03 percentage points across three 7B–8B models.

Critical strengths
  • Reusing the same trajectories across methods isolates the effect of the aggregation rule.
  • By restricting the diagnostic analysis to cases where the correct answer is among the two leading candidates, the paper separates selection errors from coverage failures. The finding that trajectory signals become more useful as the vote gap narrows provides a specific empirical basis for competition-aware gating.
  • The unconditional reselection control and rescue/broken analysis provide a clear motivation for selective intervention.
Critical weaknesses
  1. The baseline comparison does not sufficiently establish the contribution relative to prior confidence-based aggregation. UTRS and TSWS are useful controls, but both use the same tail-certainty signal as CARS. Relevant alternatives include Perplexity Consistency from RPC [1] and CER’s confidence aggregation [2]. Their omission leaves open whether competition gating improves on stronger uses of model probabilities. These methods should be compared with their prompting and compute differences accounted for.

  2. The reported gains are small relative to the unresolved statistical uncertainty. Evaluation uses a single generation seed and provides no paired confidence intervals. An improvement of roughly one percentage point is encouraging, but insufficient to establish reliability across questions and sampled trajectories. Repeated sampling and paired analysis are particularly important for a method that changes only a small subset of decisions.

  3. The parameter-selection protocol is not sufficiently specified. The threshold of 0.70 is described as an empirical operating point, but the paper does not clearly identify held-out data for choosing it or the tail window. The sensitivity analysis is useful, yet does not resolve whether the reported gains persist when both choices are fixed before evaluation.

  4. The rule does not account for uncertainty in the cluster means. The same vote ratio can arise from a few trajectories or many, and any positive score difference can trigger reselection once the gate opens. Exact ties are handled, but the stability of near-tie decisions is not established. Results stratified by cluster size and Top-2 mass, together with a variance-aware control, would clarify this limitation.

  5. Evaluation is limited to a single budget of 128 trajectories, despite the method’s dependence on vote counts and estimated cluster means. Results at 16, 32, and 64 samples are needed to establish practical usefulness. The main benchmarks are predominantly mathematical, and the stronger Qwen3-8B model appears only in a conditional diagnostic rather than a full CARS evaluation.

[1] Zhi Zhou, Yuhao Tan, Zenan Li, Yuan Yao, Lan-Zhe Guo, Yu-Feng Li, and Xiaoxing Ma. A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning. NeurIPS, 2025.

[2] Ali Razghandi, Seyed Mohammad Hadi Hosseini, and Mahdieh Soleymani Baghshah. CER: Confidence Enhanced Reasoning in LLMs. ACL, 2025, pp. 7918–7938.

Questions for authors
  1. Does CARS outperform established probability-based aggregation methods under matched generation and validation budgets?
  2. Do the gains remain significant across independent trajectory samples, using paired comparisons with self-consistency?
  3. How do accuracy and intervention reliability vary with sampling budget, cluster size, and Top-2 mass? Does a minimum score margin improve near-tie decisions?
  4. What are the full CARS results on Qwen3-8B, and what additional time and memory are required to record the trajectory scores?
Reviewer_83CKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks when self-consistency (SC) majority voting should be overridden by probability-based trajectory signals. Each problem's answer distribution is summarized by the Top-2 mass M = q1 + q2 and the competition ratio R = q2/q1 (Eq. 1), where clusters are formed by exact final-answer strings. The diagnostic analysis in Sec. 2 makes three observations. First, on the "Top-2-controlled" subset, where the correct answer is one of the two largest clusters, SC accuracy depends mainly on R and drops as R approaches 1 (Fig. 1b). Second, the correct cluster's mean tail-30% certainty (Eqs. 3-5) tends to be higher than that of competing wrong clusters (Fig. 2). At the trajectory level, the within-question AUROC of tail-30% certainty for separating correct from incorrect trajectories is about 0.61-0.62 (Table 4, App. B.1). Third, always choosing between C1 and C2 by this signal (UTRS) hurts accuracy overall but helps in the high-R region (Fig. 3).

The proposed method, CARS (Eq. 6), keeps the SC answer unless R >= rho = 0.70. In that case it returns C2 if C2's mean signal is strictly higher. The evaluation uses Llama3.1-8B-Instruct, Qwen2.5-7B-Instruct and InternLM3-8B-Instruct with thinking disabled. Each model samples N = 128 trajectories with one seed (App. A.1). The benchmarks are AIME24/25 (60 problems), AMC23 (40), MATH-500, Minerva (272), OlympiadBench (674) and GPQA-Diamond (198), 1,744 problems in total (Table 3). CARS improves pooled accuracy over SC by +0.86, +0.86 and +1.03 pp (Table 1). Signal-weighted voting (TSWS) gains +0.57, +0.00 and +0.23 pp, and UTRS loses 3.0-5.1 pp. Further experiments cover four signal variants (Table 2), a threshold sweep over rho in [0.50, 0.90] (Fig. 6, Table 5), per-benchmark gain maps (Fig. 5) and a descriptive analysis of Qwen3-8B without running CARS end-to-end (Fig. 7, Sec. 4.5).

Critical strengths
  • A clear diagnosis that yields a simple, cost-free rule.
    • Restricting to the Top-2-controlled subset separates coverage failures from selection failures (Fig. 1a vs. 1b).
    • The signed rescue/broken map (Fig. 3, App. C.2) shows that the harm of unconditional reselection is concentrated in the low-R, high-M corner, where SC is already reliable. The benefit appears at high R.
    • The intervention accounting makes the trade-off concrete. Unconditional reselection gives 271 rescues against 491 broken cases, while CARS gives 116 against 68 (Sec. 4.2, Table 1).
    • The resulting rule (Eq. 6) needs no training and no extra generations, only post-hoc aggregation of log-probabilities already recorded during sampling (Sec. 3.2).
    • The gating effect is also reasonably stable. In Table 5, every rho in {0.60, 0.70, 0.80, 0.90} gives a positive net rescue on all three models. Only Llama at rho = 0.50 is negative (58 rescues vs. 62 broken, -0.23 pp). Fig. 6 likewise stays above zero for all three models from rho ≈ 0.52 to 0.90.
Critical weaknesses
  • Size and significance. The net gains are 15, 15 and 18 of 1,744 problems per model (Table 5 at rho = 0.70), from one sampling seed (seed 42, App. A.1), with no CIs or significance tests. My exact McNemar tests on the rescue/broken counts give p ≈ 0.09 for Llama (42/27), p ≈ 0.05 for Qwen (33/18) and p ≈ 0.03 for InternLM (41/23). Pooling gives p < 0.001 (116/68), but only when model-problem pairs are treated as independent, although all three models answer the same problems. These are also the models used to design the method. Per-benchmark changes are one to a few problems.
    • Choices made on the test data. The tail window was picked by AUROC on the same three models and data (App. B.1, Table 4). Even there, Tail-30% is best only for Llama and on average: Qwen is best with Tail-40% (0.6192) and InternLM with Tail-10% (0.6243). The gate idea was read off Fig. 3, which pools the same instances. rho = 0.70 is "an empirical operating point" (Sec. 4.1) inside a shaded "higher-gain region" (about 0.63-0.75) that Fig. 6 draws on the same data. That band contains only Qwen's peak (about 1.09 pp at rho ≈ 0.65); InternLM peaks at about 1.38 pp (rho ≈ 0.78) and Llama at about 1.09 pp (rho ≈ 0.83). The sign of the gain is fairly robust (see Strengths), but its size ranges from -0.23 to about +1.38 pp in Fig. 6 (-0.23 to +1.26 pp in Table 5). The abstract's "not tied to a particular ... gating threshold" therefore overstates the evidence.
  • Weak baselines and a narrow scope: it is not shown that hard Top-2 gating beats a tuned frequency-confidence combination, or that it transfers.
    • Baselines. Voting clusters are raw answer strings: "different strings remain separate voting clusters even when both are correct" (App. A.4). This splits equivalent correct answers, lowers q1 and inflates R, so part of the roughly 1 pp gain may only recover fragmentation errors that canonicalized SC (standard practice) avoids. TSWS sums raw tail-30% certainty "without exponentiation or z-score normalization" (App. A.4). Its score is therefore |Ci| times the mean signal, which (my inference) nearly always reproduces SC; this fits its 30-98 overrides and +0.00 to +0.57 pp gains (Table 1). The comparison also leaves out CISC with a tuned temperature (Taubenfeld et al., 2025), DeepConf filtering plus voting (Fu et al., 2026), self-certainty Borda voting (Kang et al., 2025), a soft log q_i + beta * s̄_i combination, and a random or length-based C1/C2 tie-break inside the gate. The tie-break control is needed given a signal AUROC of about 0.61.
    • Scope. All end-to-end results use 7-8B non-thinking models at N = 128. The stronger Qwen3-8B is analysed only descriptively (Fig. 7), and that figure gives no evidence of a gain at the default gate. Reading the per-bin rates and counts for R >= 0.70, SC selects correctly 49 of about 81 times and certainty-based Top-2 selection (the CARS rule inside the gate) 45 times. That is a net of about -4, which is within noise, as the authors' own caveat about sparse high-R bins implies (p. 9, l. 448-449). The difference is only about +1 of 49 for R >= 0.80 and about +6 of 31 for R >= 0.85. Fig. 5, which the paper uses to show the "recurrence of this pattern across individual settings" (Sec. 4.3, p. 7), omits GPQA-Diamond, AIME and AMC. On GPQA, unconditional reselection beats both CARS and SC for Llama (UTRS 34.34 vs. CARS 33.84 vs. SC 32.83) and Qwen (UTRS 36.87 vs. CARS 35.35 vs. SC 32.83) (Table 1). The UTRS-CARS gaps are 1 and 3 of 198 problems, which is small, but this offers no support for the paper's principle to "retain frequency-based SC under weak competition" (Sec. 2.4).
Questions for authors

see above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.