acceptodds
Withdrawn from ICLR 2027

BEYOND THE FINAL ANSWER: CHARACTERIZING AND SELECTING REASONING TRAJECTORIES

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sampling multiple reasoning trajectories creates two decisions: which unfinished trajectories deserve more computation, and which terminal answer should be submitted. These stages can benefit from different evidence. A trajectory can merit continuation before its current answer is reliable, while aggregate hidden-state geometry does not determine when substantial updates occur. We show that early answer-support histories predict later correctness even when both the current candidate and an isolated trial answer are wrong. Update timing also helps terminal comparison, including against aggregate-geometric controls. We introduce the Support History Score (SHS), combining candidate dominance frequency, confidence when leading, and historical competitive margin, and the Trajectory Outcome Score (TOS), based primarily on the temporal centroid of representational updates. SHS guides Continuation Selection (CS); two-stage Answer Selection (AS) retains two trajectories with SHS and compares their final submissions with TOS. Across three reasoning models and five mathematical and scientific benchmarks, support history improves continuation selection by 1.4–3.1 percentage points over matched endpoint-only scoring. Within the same retained pairs, TOS improves accuracy by 2.0–4.8 points over uniform choice. Both procedures reduce generated tokens while matching or exceeding majority-vote accuracy in most settings. The reasoning models remain unchanged; supervised two-input TOS extensions use problem-disjoint cross-fitting.

Primary area:
interpretability and explainable AI
Keywords:
inference-time scaling, reasoning trajectory selection, test-time compute allocation, answer-support dynamics, hidden-state dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_DKuwOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how to choose among multiple reasoning trajectories at test time. It separates two decisions: which unfinished trajectory should receive more computation, and which completed trajectory should provide the final answer.

For continuation selection, the paper proposes the Support History Score (SHS), which tracks how candidate-answer support changes over time. For final answer selection, it proposes the Trajectory Outcome Score (TOS), based mainly on when hidden-state changes occur during the reasoning trajectory. The methods are evaluated on three reasoning models across AIME 2024 to 2026, Omni-64, and GPQA-Diamond.

Critical strengths
  1. The paper makes a useful distinction between continuation selection and final answer selection.

The separation between deciding which trajectory deserves more computation and deciding which completed trajectory should provide the final answer is clear and useful. SHS and TOS are designed around these two decisions, giving the framework a clear structure.

  1. The experimental design is solid.

The paper uses the same sampled trajectories across methods and includes several controlled comparisons. The history vs. endpoint experiment fixes the observation endpoint, score weights, trajectory pool, and continuation protocol, so the effect of using trajectory history is relatively well isolated. The additional controls for SHS screening and TOS ranking also strengthen the empirical analysis.

Critical weaknesses
  1. The benefit of support history is mainly demonstrated on AIME.

The matched history vs. endpoint comparison shows consistent gains across the three models, but the paper notes that these gains come from the AIME settings because the selected Omni-64 and GPQA configurations use only one measurement. The additional GPQA sensitivity analysis is mixed across observation points. This makes it difficult to tell how broadly the history component itself generalizes beyond AIME.

  1. The advantage of TOS over simpler terminal-selection signals is modest and model-dependent.

TOS improves over uniform selection, but its margin over the fixed geometric control is relatively small. On Phi, the shorter-trajectory baseline slightly outperforms TOS. Since Table 2 aggregates all datasets for each model, it is difficult to see when the temporal-centroid signal is actually useful.

Questions for authors
  1. Could the authors provide a broader matched comparison of support history and current-point scoring outside AIME?

The GPQA sensitivity result is useful, but the direction of the effect changes with the observation point. A small additional breakdown on Omni-64 or GPQA, across more than one model or observation point, would help clarify how robust the history effect is.

  1. Could the authors provide a per-dataset breakdown of TOS against the simpler terminal-selection controls?

In particular, results against the fixed geometric rule and the shorter-trajectory baseline would help show where the temporal-centroid signal provides a consistent benefit. If the gains are strongly model- or dataset-dependent, clarifying or narrowing the corresponding claim would also address this concern.

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

This paper proposes Support History Score (SHS) to predict whether the unfinished trajectories can arrive a correct answer, and Trajectory Outcome Score (TOS) to prefer the finished trajectories whose presentations change substantially toward the end of reasoning. In Continuation Selection (CS), SHS selects one of four unfinished trajectories to continue. In Answer Selection (AS), SHS retains two trajectories, and TOS selects the final answer after they finish. And the paper has serious readability issues.

Critical strengths

The research question is important. Deciding which trajectory to continue and which final answer to submit are distinct and two important directions. Studying their predictive signals separately is valuable and important to the academic community.

Critical weaknesses

W1. The paper is difficult to read. Here are some examples:

  • In Figure 4, the caption says, “Within-problem AUC on Qwen3-4B, AIME 2024–2026.” But it does not explain how this AUC is calculated or what it measures. For panel (a), it says, “A fixed early low-entropy score evaluated at current and later answer positions.” The details appear only in Appendix A.10. The setup is also easy to misread: the entropy score is calculated once at 4K tokens, while correctness is evaluated again after continuation to 8K, 16K, and 32K tokens. Yet the x-axis is labeled “Number of tokens,” which may lead readers to think that entropy is recalculated at each position.
  • In lines 14–15 of the abstract, the paper states that “aggregate hidden-state geometry does not determine when substantial updates occur.” The term “aggregate hidden-state geometry” is introduced without explanation or a clear connection to the preceding point about current-answer reliability.
  • The introduction is hard to follow. For example, line 42 asks whether “support history over complete candidate identities predicts later success, including when current-answer checks fail.” The preceding discussion contrasts current-answer confidence with continuation ranking, but it does not explain why support history over complete candidate identities is the proposed next step.

W2. The main gains need stronger robustness checks and closer baselines. Table 1 and Figure 5 use one pool of four sampled trajectories per problem, without paired uncertainty for the main CS/AS comparisons. Figure 5’s history gain comes entirely from AIME, because the Omni-64 and GPQA settings use only one observation. The largest CS gain in Table 1—10 points on Qwen3-4B AIME 2025—amounts to three additional correct answers among 30 questions. I would put more weight on this result with per-dataset paired intervals and a second, independently sampled trajectory pool. The comparison is also missing ST-BoN, which the paper cites and which also selects trajectories early, and Entropy Centroids as a terminal-selection comparison for the proposed temporal centroid.

Questions for authors

TOS averages hidden states within 500-token blocks before measuring changes between blocks. How sensitive are its rankings and selection accuracy to block size? A few trajectories annotated with their largest measured changes would help show what those changes correspond to in the reasoning text.

Reviewer_hwtoOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies how to allocate test-time compute among KK parallel reasoning trajectories. At an early position of each trajectory, the authors teacher-force all candidate answers against an appended answer prompt to read out a candidate distribution, and aggregate these readouts over time into a Support History Score (SHS). Continuation Selection keeps only the top-SHS trajectory, and Answer Selection keeps the top two and picks between their final answers with a Trajectory Outcome Score computed from hidden-state update dynamics. On AIME 2024-2026, a 64-problem Omni-MATH subset and GPQA-Diamond with Qwen3-4B, Phi-4-mini-reasoning and R1-Distill-Llama-8B (K=4K=4), the methods match or exceed majority voting in most settings while generating 35-65% fewer reasoning tokens.

Critical strengths
  1. Allocating compute among parallel trajectories early is a timely problem for reasoning models. Separating "which trajectory to continue" from "whether the current answer is reliable" is a sensible framing, and Figure 4(b) shows SHS stays predictive even when the current top candidate is wrong.
  2. The protocol is careful in places: the learned variant is cross-fitted on problem-disjoint folds, and Table 2 includes length and termination heuristics as controls.
Critical weaknesses
  1. The experiments are far from sufficient to support the claims. Three of the five datasets have 30 problems, so the headline +10.0 points on AIME 2025 is 3 problems and half of the deltas in Table 1 are at most one problem, yet no confidence interval is reported for Table 1. Against the trivial shorter-trajectory heuristic, TOS gains only +1.7/-0.6/+1.4 points (Table 2). Moreover, hyperparameters are chosen per model-dataset pair (Table 4, 5), and the two fixed SHS recipes (Line 320-322) attain exactly the maximum (68/90) of the weight grid evaluated on the AIME problems in Appendix A.9, without an explanation of how they were chosen. I would like to see intervals on larger benchmarks (full Omni-MATH, MATH-500), KK up to 16-64, and one shared configuration.
  2. The closest prior work and the decisive baseline are missing. Reading intermediate answers to allocate compute is the idea of Certaindex [2], and the "isolated trial answer" treated as a strawman (Line 261-263) is DEER [3]. Landscape of Thoughts [4] already tracks the likelihood of every answer candidate along a trajectory. Most importantly, DeepConf [1] is a training-free method with the same goal and headline (match majority voting with far fewer tokens) and must be compared at equal tokens. The paper instead compares with CoP and LT, which fall below MV@4 in most Phi and R1 settings (Table 1).
  3. The paper is difficult to follow. Many terms are used before definition ("matched endpoint-only scoring" in Line 25, "stratum" in Line 262, "interaction interval" in Line 437), Figure 1 cannot be read from its caption, and the 500-token readout spacing and the 0-999 candidate set appear only in the appendix (A.2, A.6). See more in questions.

[1] Fu et al. Deep Think with Confidence. ICLR 2026.

[2] Fu et al. Efficiently Scaling LLM Reasoning Programs with Certaindex. NeurIPS 2025.

[3] Yang et al. Dynamic Early Exit in Reasoning Models. ICLR 2026.

[4] Zhou, Zhu, Li et al. Landscape of Thoughts: Visualizing the Reasoning Process of Large Language Models. ICLR 2026.

Questions for authors
  1. The title does not reflect the contribution. Consider "Allocating test-time compute among reasoning trajectories with answer-support histories".
  2. Line 128-129: Which hidden state is ht(ℓ)h_t^{(\ell)}, and why average μh\mu_h uniformly over all layers?
  3. Line 150-158: di,ℓd_{i,\ell} equals ∥vi∥2\|v_i\|_2 defined in Line 140. Please use one symbol.
  4. Line 164-168: How does μh\mu_h correlate with trajectory length? Otherwise "update timing" is not separable from length.
  5. Line 185: SHS requires enumerating a canonical candidate set. How does the method apply to free-form answers or code?
  6. Line 218: log⁡Qtj(a)\log Q_{t_j}(a) is −∞-\infty for candidates with zero mass. How is this handled?
  7. Line 227: sf(t)s_f(t) is defined but never used. Is it what Figure 3 plots?
  8. Line 320-322: On which data were the AIME and GPQA configurations in Table 4 selected?
  9. Table 1: Please add Single and MV@2 rows, and explain why ST@4 beats both CS and AS on R1 AIME 2025, Phi AIME 2026 and Phi GPQA.
  10. Line 481-483: The reported savings exclude the cost of teacher-forcing 1,000 candidates every 500 tokens. Please report this scoring cost for the math tasks and the end-to-end wall-clock against MV@4.
  11. Typo:
    • Figure 4: panels labelled (A)/(B) but caption says (a)/(b)
    • Omni-64 (Table 1) → Omni (Figure 3, 6)
    • R1-8B (Table 2) → R1-Distill-Llama-8B

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

Related papers

Discussion (0)

Sign in to comment.