acceptodds
Under review as a conference paper at ICLR 2027

Reason Forward, Reflect When Needed: Predictive Anchoring with Suffix-Evidence for Revisable LLM Reasoning

Abstract

Human experts seldom revise uncertain intermediate steps at once. They tentatively accept a step, retain alternatives, inspect downstream effects, and revise only evidence-supported parts of the reasoning state. Existing latent-reasoning methods, however, typically activate continuous representations at high-entropy tokens, conflating uncertainty with the need for latent reasoning. Our analysis reveals three findings: Uncertainty is immediately observable but insufficient for intervention. Intervention impacts emerge later yet locally diagnosable. Revision scope should match influence scope. Building on these observations, we propose PASER, a training-free LLM reasoning framework with three key innovations: 1) PASER treats high-entropy uncertainty as a signal for caching predictive anchors, rather than immediately activating latent reasoning. 2) PASER restores each anchor and evaluates the same realized suffix under discrete CoT and predictive soft histories, determining whether the current latent reasoning interference has an effect. 3) By accumulating impact evidence over a longer horizon, PASER maps its strength to hierarchical reflective actions. Evaluated on mathematical, STEM, coding, and general reasoning benchmarks, PASER achieves aggregate performance over competitive training-free baselines across diverse model families and scales.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM, Latent reasoning, Hierarchical reflective actions

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_FBghOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to improve the efficiency and quality of latent reasoning in large language models through a training-free, evidence-guided revision framework. During autoregressive generation, token-level uncertainty signals such as entropy indicate potential reasoning candidates but provide limited guidance on whether intervention will improve the final answer. However, by comparing the discrete and predictive-soft trajectory distributions over a suffix window after the anchor, the downstream influence of activating latent reasoning becomes quantifiable without costly rollouts. Building on this observation, the authors propose PASER, which caches high-entropy tokens as predictive anchors, diagnoses their influence in a delayed manner via Jensen-Shannon divergence, and applies hierarchical reflective actions (KEEP, GUIDE, or ROLLBACK) calibrated by the strength of accumulated evidence. Evaluated on math, STEM, coding, and commonsense reasoning benchmarks across Qwen3, DeepSeek-R1-Distill, Llama-3.1-NN, and DeepScaleR model families, PASER demonstrates consistent improvements over Soft-Thinking, SeLaR, and SwiR while achieving higher throughput and lower GPU memory usage.

Critical strengths

The proposed method is based on three well-motivated findings. 1. Uncertainty is immediately observable but insufficient for intervention. 2. Intervention impacts emerge later yet are locally diagnosable. 3. Revision scope should match influence scope.

In experiments, the paper verified these findings by correlating entropy-trajectory divergence and comparing short-window vs. partial-rollout prediction.

The method is training-free, introduces no additional GPU memory overhead, and actually improves inference efficiency compared to SeLaR and SwiR, while achieving better accuracy. This makes PASER practically deployable without the cost of retraining or fine-tuning.

Critical weaknesses
  1. The method introduces several hyperparameters—entropy threshold δ, suffix window W, history window K, and the JSD thresholds for KEEP/GUIDE/ROLLBACK actions (0.3 and 0.7)—whose optimal values may vary across model families and task domains. While ablations are provided for δ and K, the sensitivity analysis is limited to a single model (Qwen3-8B) and does not cover cross-model or cross-task stability of the full configuration.

  2. The entropy threshold H_t ≥ 0.5 used to identify candidate anchor positions is introduced without justification or sensitivity analysis. The normalized top-3 entropy ranges from 0 to 1, and the choice of 0.5 as the cutoff determines which tokens are considered "uncertain enough" to warrant caching. The paper does not report: (a) what fraction of generated tokens exceed this threshold across different models and tasks, (b) whether the three key findings (especially Finding 1) hold under alternative thresholds (e.g., 0.3, 0.7), or (c) whether the downstream performance of PASER is sensitive to this initial filtering criterion. If the observations are fragile to the choice of H_t threshold, the generalizability of the motivational analysis is called into question.

  3. The pilot study (Appendix B.1) employs a single, fixed intervention modality—replacing the discrete token with a top-3 probability-weighted soft embedding—to measure whether a high-entropy token "needs" latent reasoning intervention, using the resulting JSD over 256 tokens as the sole criterion. A low JSD under soft-embedding substitution merely indicates that this particular intervention does not substantially alter the subsequent trajectory; it does not rule out that other intervention strategies could meaningfully improve the final answer at the same token position. Consequently, the conclusion that "high entropy is not intervention-sufficient" is only valid relative to the soft-embedding intervention, and the paper provides no evidence that the set of "low-JSD high-entropy tokens" identified by this procedure would not benefit from alternative forms of reasoning intervention.

Questions for authors

It would be better if the paper included cross-model hyperparameter sensitivity experiments to further verify whether the default configuration (δ, W, K, β_L, β_H) generalizes across different model families and task domains.

The paper should conduct a sensitivity analysis on the entropy threshold H_t ≥ 0.5 to further verify that the three key findings and the overall performance are robust to alternative threshold choices (e.g., 0.3, 0.7), and to report what fraction of tokens exceed this threshold in practice.

The pilot study (Section 3.2 / Appendix B.1) employs a single, fixed intervention modality—replacing the discrete token with a top-3 probability-weighted soft embedding—to measure whether a high-entropy token "needs" latent reasoning intervention, using the resulting JSD over 256 tokens as the sole criterion. The paper should conduct verification experiments with alternative intervention modalities (e.g., re-sampling from the top-3 candidates, or injecting a retrieval-augmented context) to further verify that Finding 1 ("high entropy is not intervention-sufficient") generalizes beyond the soft-embedding intervention used in the current analysis.

3 comments
AuthorsOpenReview ↗

Response to Reviewer FBgh [part 1]

We sincerely thank you for recognizing the value of our work and for providing constructive feedback, which has helped us further improve the quality of our manuscript.

W1 and Q1: Cross-model and Cross-task Hyperparameter Sensitivity Analysis

We appreciate the reviewer's insightful comments. We agree that evaluating the robustness of the default configuration across different model families and task domains is important for assessing PASER's generalizability.

Original Hyperparameter Selection. We clarify that the original sensitivity analyses of (δ)(\delta) and KK were conducted exclusively on Qwen3-8B across five benchmarks: GSM8K, MATH-500, AIME2024, AIME2025, and HumanEval. Based on these analyses, we selected $$\delta=0.8=0.8 and K=256K=256 as the default values. These parameters were subsequently fixed and directly applied to other models and tasks without further tuning based on their evaluation results. Therefore, the multi-model and multi-benchmark evaluations in the original manuscript already provide indirect evidence of the generalizability of the shared default configuration across models and tasks.

Additional Cross-model Sensitivity Analysis. To further validate the transferability of the default configuration, we conduct additional sensitivity analyses of the key hyperparameters �3��3�, �4��4�, and �5��5� on DS-R1-Qwen-7B (Appendix Tables 18 and 19). The results exhibit similar performance trends across different models, with the original default configuration maintaining stable performance on both models. These findings further support the cross-model robustness and applicability of PASER's default configuration.

Table R1. Sensitivity Analysis of βH\beta_H and βL\beta_L on DeepSeek-R1-Distill-Qwen-7B

(a) Sensitivity to βH\beta_H

βH\beta_HMATH-500AIME2024CommonsenseQAHumanEvalAverage
1.093.6056.6765.1988.4175.97
1.294.0055.8465.6889.0276.14
1.494.2056.6765.9289.0276.45
1.694.4057.5066.2689.6376.95
1.894.2056.6765.6889.0276.39
2.093.8055.8465.6088.4175.91
2.293.4055.8465.0388.4175.67

(b) Sensitivity to βL\beta_L

βL\beta_LMATH-500AIME2024CommonsenseQAHumanEvalAverage
0.293.6055.8465.6888.4175.88
0.493.8056.6765.9288.4176.20
0.694.0057.5066.2689.0276.70
0.894.4057.5066.2689.6376.95
1.094.2056.6765.9289.6376.61
1.294.0055.8465.9289.0276.20
1.493.8055.8465.1989.0275.96

W2 and Q2: Sensitivity Analysis of the Predictive Anchor Entropy Threshold

We thank the reviewer for the constructive comments regarding the choice and robustness of the entropy threshold.

Robustness of Performance and Key Findings across Entropy Thresholds. To examine whether our findings depend on the default threshold of 0.5, we conduct additional benchmark evaluations on DS-R1-Llama-8B and Qwen3-8B using alternative thresholds of 0.3 and 0.7, while keeping all other PASER configurations unchanged (Appendix Table 15). The results show only minor variations in overall performance across different thresholds, indicating that PASER is robust to reasonable changes in the entropy threshold.

Table R2. Sensitivity analysis on HtH_t under Qwen3-8B and DS-R1-Llama-8B.

ModelHtH_tMATH-500GPQAHumanEvalMBPPAvg.
Qwen3-8B0.397.2062.1296.3494.9487.65
0.597.6062.6396.9595.3388.13
0.797.2061.6295.1294.5587.12
DS-R1-Llama-8B0.391.0048.9988.4182.8877.82
0.591.2049.4989.0283.6678.34
0.790.6048.9988.4182.4977.62

Furthermore, we provide additional pilot studies across multiple models and tasks using different entropy thresholds in Appendix Figures 14 and 15. As shown in these figures, although varying the threshold affects the selection range of candidate anchors, our key findings remain qualitatively consistent across different threshold settings.

Overall, these additional experiments validate the robustness of PASER to entropy threshold selection from both cross-model performance and the consistency of its key findings, demonstrating that its effectiveness does not rely on a specific default threshold.

AuthorsOpenReview ↗

Response to Reviewer FBgh [part 2]

W3 and Q3. Generalizability of the Pilot Intervention Mechanism

We thank you for pointing out the limited diversity of intervention mechanisms in our original pilot study. We agree that using only probability-weighted soft embedding replacement for discrete token embeddings is insufficient to establish the robustness of Finding 1 across different intervention mechanisms.

To address this concern, we introduce an alternative intervention strategy, Top-3 Candidate Resampling, which samples a replacement token from the renormalized probability distribution over the top-3 candidate tokens and applies the intervention at the same high-entropy anchor positions. This allows us to examine whether different intervention mechanisms exhibit consistent downstream influence patterns.

Specifically, we conduct additional pilot experiments across multiple models and tasks (Appendix Figures 12 and 13), using a consistent subsequent-token evaluation window to measure the distributional changes induced by the interventions. As shown in Figures 12 and 13, the results consistently reproduce the observation from our original pilot study across different models and tasks: high-entropy positions do not necessarily exert substantial downstream influence, and entropy alone is therefore insufficient to determine whether intervention is warranted. These findings indicate that Finding 1 is not specific to the original Soft Embedding replacement strategy, further supporting its robustness across different intervention mechanisms.

We once again express our sincere gratitude for your constructive suggestions. We hope that these additional experiments and explanations will fully address the relevant issues and further strengthen the empirical support for our research results.

AuthorsOpenReview ↗

Response to Reviewer FBgh

Dear Reviewer FBgh,

Thank you again for your time and constructive comments. As the discussion period is drawing to a close, we would like to kindly check whether our response has addressed your concerns. Looking forward to your reply.

Thanks again for your time and help in improving the quality of this paper.

Regards from the authors.

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

This paper introduces PASER, a training-free framework for revisable language-model reasoning. The work challenges the common assumption that uncertainty should immediately trigger latent reasoning and instead uses downstream evidence to determine whether and to what extent an earlier reasoning state should be revised. Experiments across mathematical, scientific, coding, and general reasoning benchmarks show consistent improvements over several explicit and latent-reasoning baselines.

Critical strengths
  1. Well-motivated distinction between uncertainty and downstream influence. The paper identifies an important limitation of entropy-triggered latent reasoning: a token can be uncertain without materially affecting the subsequent reasoning trajectory. This distinction is intuitive, empirically motivated, and relevant to a broad class of adaptive reasoning methods.
  2. Practical and conceptually appealing revision framework. PASER is training-free and can be applied to existing reasoning models without parameter updates. Its graded revision policy is also more flexible than a binary decision between standard and latent reasoning, since it attempts to match the scope of revision to the estimated scope of influence.
  3. Broad empirical evaluation and useful ablations. The experiments cover multiple model families, parameter scales, and reasoning domains. The ablations separately examine predictive anchoring, diagnostic-window length, long-horizon auditing, and hierarchical reflection, providing evidence that the main components contribute to performance.
Critical weaknesses
  1. Replay-related inference costs are insufficiently specified. PASER requires suffix replay, long-horizon auditing, and occasional regeneration after rollback. It is unclear whether these additional operations, including discarded tokens, are fully counted in the reported efficiency metrics; otherwise, the computational gains may be overstated.
  2. The evaluation emphasizes aggregate accuracy but provides limited behavioral analysis. The paper does not report anchor frequency, short-window pass rates, or the proportions of KEEP, GUIDE, and ROLLBACK. These statistics, together with per-action accuracy and failure cases, would clarify whether the gains are broadly distributed or driven by a small number of costly interventions.
Questions for authors
  1. How strongly do the short- and long-horizon influence scores correlate with final-answer correctness? Could the authors report the conditional accuracy of examples assigned to KEEP, GUIDE, and ROLLBACK, as well as the false-positive and false-negative rates of the intervention decisions?
  2. What happens when the realized suffix used for diagnosis is itself incorrect? In particular, can an incorrect suffix that is similarly supported under both histories pass through the diagnostic stage without revision?
2 comments
AuthorsOpenReview ↗

Response to Reviewer sKBj [part 1]

We sincerely thank you for recognizing the value of our work and for providing constructive feedback. These comments have helped us further clarify our methodology, strengthen our experimental analysis, and improve the overall quality of the manuscript.

W1: Accounting for Replay-related Inference Costs

We appreciate the reviewer's concern regarding the additional computational overhead introduced by PASER.

End-to-end latency accounts for all additional operations. We clarify that the Time per Request (TPR) reported in Figure 16 measures the end-to-end latency of a complete inference request, rather than only the final CoT generation time. Therefore, all operations performed during PASER inference, including Suffix Replay, Long-horizon Audit, GUIDE, and regeneration after ROLLBACK, are included in the total latency. Importantly, the time spent generating tokens that are subsequently discarded during rollback is also fully accounted for.

TPS is calculated based on effective output tokens. We define TPS as the number of tokens retained in the final output divided by the end-to-end inference time. Replayed and discarded tokens are excluded from the numerator, while their computational costs remain reflected in the denominator. Thus, TPS measures effective output throughput under the complete inference cost, rather than the internal token processing rate, without counting replayed or discarded tokens as useful outputs.

End-to-end efficiency. As shown in Figure 16, PASER achieves lower end-to-end latency and higher effective output TPS than SeLaR and SwiR on both Qwen3-1.7B and DeepSeek-R1-Distill-Qwen-7B. Moreover, Tables 7 and 8 show that PASER also incurs lower peak GPU memory usage than these baselines. These results demonstrate that PASER maintains competitive end-to-end inference efficiency even when the overhead of its additional diagnostic and reflective operations is fully accounted for.

W2: Statistical Analysis of Reflective Actions

We thank the reviewer for the constructive suggestions. To further characterize PASER's actual intervention behavior, we provide additional box plots in Appendix Figure 18, showing the per-example execution counts of KEEP, GUIDE, and ROLLBACK on MBPP (coding) and AIME2024 (mathematical reasoning), using Qwen3-8B and DS-R1-Llama-8B.

As shown in Figure 18, ROLLBACK is not frequently triggered. Even on the challenging AIME2024 benchmark, the median number of ROLLBACK operations is only approximately 9–12 per example. This indicates that PASER does not aim to maximize intervention frequency. Instead, it selectively applies revisions based on downstream influence evidence, concentrating interventions on reasoning decisions with substantial downstream effects and avoiding unnecessary computational overhead.

Q1: Conditional Accuracy of Different Reflective Actions

Since a single test example may contain multiple high-entropy anchors and trigger KEEP, GUIDE, and ROLLBACK at different positions, we adopt an anchor-level analysis rather than partitioning examples into three mutually exclusive groups. Specifically, each anchor is associated with the final-answer correctness of its corresponding example after the complete PASER inference process. We then calculate the conditional accuracy for each action as:

$

Acc(a)=∑i=1Mniayi∑i=1Mnia×100%,a∈{KEEP,GUIDE,ROLLBACK}.\mathrm{Acc}(a)= \frac{\sum_{i=1}^{M} n_i^a y_i} {\sum_{i=1}^{M} n_i^a} \times 100\%, \quad a\in\{\mathrm{KEEP},\mathrm{GUIDE},\mathrm{ROLLBACK}\}.

$

where �1��1� denotes the total number of test examples, �2��2� is the number of anchors assigned action �3��3� in example �4��4�, and �5��5� indicates whether the final answer of example �6��6� is correct.

As shown in Appendix Tables 11 and 12, all three reflective actions exhibit favorable conditional accuracy, supporting the rationale of PASER's hierarchical reflection mechanism. Furthermore, differences in conditional accuracy reflect the varying levels of difficulty associated with the corresponding reasoning states: KEEP primarily preserves reasoning trajectories that do not require further revision, whereas GUIDE and ROLLBACK are more frequently associated with states exhibiting different degrees of downstream influence. These results support PASER's design of selecting appropriate reflective strategies based on downstream influence, rather than applying uniform interventions to all high-entropy positions.

Anchor-level Action-conditioned Accuracy on MBPP

ActionTotal AnchorsCorrect AnchorsAccuracy (%)
KEEP1,6231,42287.62
GUIDE94279384.18
ROLLBACK30825382.14

Anchor-level Action-conditioned Accuracy on AIME2024

ActionTotal AnchorsCorrect AnchorsAccuracy (%)
KEEP74864486.10
GUIDE55246283.70
ROLLBACK29222476.71
AuthorsOpenReview ↗

Response to Reviewer sKBj [part 2]

Q2: Potential Errors in the Diagnosed Suffix

We thank the reviewer for raising this important concern. We acknowledge this limitation and would like to further clarify the diagnostic objective and scope of PASER.

Limitation of the diagnostic mechanism. PASER estimates the downstream influence of an earlier uncertain anchor by comparing the relative probability support of the same realized suffix under the original discrete CoT history and the predictive soft history. This support difference measures relative dependence on the preceding history, rather than the absolute correctness of the suffix itself. Consequently, if an incorrect suffix receives similar probability support under both histories, the resulting support difference may remain small, allowing it to pass the short-window diagnostic without triggering further intervention.

Design objective and distinction from existing methods. Importantly, PASER is not designed to guarantee the correctness of every intermediate reasoning suffix. Instead, its objective is to determine whether an uncertain high-entropy decision has a substantial downstream influence and, accordingly, whether intervention is warranted and what scope of revision is appropriate. Existing entropy-triggered latent reasoning methods, such as SwiR and SeLaR, introduce alternative representations or reasoning interventions at uncertain positions; however, such interventions do not guarantee that subsequent generations will be correct either. In contrast, PASER avoids equating uncertainty with the necessity of intervention. Through delayed influence diagnosis, it filters low-impact uncertain positions and concentrates computation and revision on decisions with substantial downstream effects.

Scope and future directions. Therefore, PASER should be viewed as an influence-guided selective reasoning revision framework, rather than a general-purpose intermediate reasoning error detector or an independent answer verifier. Identifying incorrect suffixes that receive similar support under both histories requires complementary external verification signals or task-level correctness checks.

We will explicitly discuss this limitation in the revised Future Work section (Appendix D.2) and highlight the integration of independent verification mechanisms with PASER's influence diagnosis as a promising direction for future research.

We once again express our sincere gratitude to you, the reviewer, for your insightful and constructive comments. We hope that these additional analyses, explanations and discussions will fully address your concerns.

Reviewer_RiaVOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes PASER, a training-free framework for controlling latent reasoning in LLMs. PASER caches recoverable decoding states at high-entropy tokens and continues explicit reasoning. It then compares support for the generated suffix under discrete and soft-embedding histories to estimate downstream influence. Based on this evidence, it selects among preserving the trajectory, steering the next-token embedding, and rolling back to regenerate the suffix. The central contribution is to separate uncertainty detection from intervention decisions and adapt revision scope to estimated influence. Evaluation spans eight models from 1.5B to 32B parameters and mathematical, STEM, coding, and general reasoning benchmarks, supplemented by diagnostic analyses, ablations, sensitivity studies, selected multi-seed evaluations, and computational-cost measurements.

Critical strengths
  1. A well-motivated intervention policy supported by component ablations. The distinction between uncertainty and downstream influence provides a useful basis for deciding when latent computation is warranted. PASER translates this into a concrete mechanism combining recoverable anchors, delayed diagnosis, and graded revision.

  2. Broad empirical evidence of usefulness across models and tasks. PASER achieves stronger reported aggregate results across multiple reasoning domains and model families, with gains extending to larger models.

Critical weaknesses

The paper insufficiently discusses and compares closely related approaches, leaving PASER’s incremental contribution unclear. At a high level, PASER’s rollback branch follows a feedback-driven process of generating subsequent outputs, evaluating them, modifying earlier continuous representations, and regenerating the continuation. This connects closely to several existing methods beyond CopT: LatentSeek; Soft Reasoning; and STIR. STIR is the closest one because it also employs short-horizon lookahead. Related principles also appear in steering methods for explicit reasoning, such as Reliable Control-Point Selection for Steering Reasoning in Large Language Models. The paper should position PASER more precisely within this literature and compare performance. The current evaluation leaves uncertainty about how PASER goes beyond these works, limiting the assessment of its methodological novelty and practical advantages over existing feedback-driven reasoning approaches.

[1] LatentSeek (https://arxiv.org/abs/2505.13308)

[2] Soft Reasoning (https://arxiv.org/abs/2505.24688)

[3] Reliable Control-Point Selection for Steering Reasoning in Large Language Models (https://arxiv.org/abs/2604.02113)

Questions for authors
  1. There are many works about training-free reasoning. As mentioned above, I hope the authors can provide more comparisons in terms of methodological differences, benchmark performance, motivations, and so on.
  2. Candidate Anchor only averages the top-3 entropy. Is there any reason for the choice? Under this criterion, how many fractions are selected to be GUIDE and ROLLBACK?
4 comments
AuthorsOpenReview ↗

Response to Reviewer RiaV [part 1]

We sincerely thank you for recognizing the value of our work and for providing constructive feedback. These comments have helped us further strengthen the clarity, technical positioning, and overall quality of the paper.

Q1: Further Discussion and Comparison with Related Work

We appreciate the reviewer's insightful suggestions. In response, we have expanded the related-work discussion in Appendix D and added empirical comparisons with these methods in Appendix Tables 9 and 10.

Difference from LatentSeek and Soft Reasoning. LatentSeek formulates test-time reasoning as an instance-level latent representation optimization problem, updating latent representations through policy gradients based on reward signals. Soft Reasoning perturbs initial token embeddings and employs Bayesian optimization to search for improved representation configurations. Both methods primarily focus on optimizing reasoning trajectories through representation-space exploration. In contrast, PASER investigates whether an uncertain high-entropy decision actually influences subsequent reasoning and whether intervention is warranted. Rather than directly optimizing latent representations, PASER preserves a recoverable Predictive Anchor and compares the support of the same realized suffix under the original discrete and predictive soft histories, thereby diagnosing the downstream influence of an uncertain decision.

Difference from Reliable Control-Point Selection. Reliable Control-Point Selection identifies stable reasoning boundaries through repeated generations and constructs reliable steering directions using SVD-based projection, aiming to improve the reliability of activation steering. In contrast, PASER diagnoses the actual downstream influence of uncertain decisions through suffix-level distributional evidence. Since behavioral stability does not necessarily imply significant downstream influence, PASER further dynamically selects among KEEP, GUIDE, and ROLLBACK based on the estimated influence, enabling revision scopes to match the extent of downstream effects.

Difference from STIR. Although both STIR and PASER leverage future trajectory evidence for selective latent intervention, they differ fundamentally in their diagnostic objectives, evidence construction, and revision mechanisms. STIR constructs a reusable library of steering vectors from successful and failed generation trajectories, and employs short-horizon lookahead probes to evaluate whether candidate correction directions improve subsequent generation. In contrast, PASER does not rely on a preconstructed correction-vector library. Instead, it restores the cached Anchor prefix state and evaluates the same realized suffix under the original discrete and predictive soft histories to estimate the downstream influence of an earlier uncertain decision. Furthermore, PASER combines short-window diagnosis with long-horizon auditing to distinguish transient effects from persistent trajectory divergence, and selects different revision scopes according to the estimated influence: KEEP preserves the original trajectory, GUIDE performs localized token-level correction, and ROLLBACK restores the Anchor state and regenerates the affected suffix. Thus, while STIR primarily validates whether candidate corrections are beneficial, PASER further diagnoses the influence of uncertain decisions and determines the appropriate scope of revision.

Benchmark Performance Comparison. To further evaluate the effectiveness of PASER, we extend our experimental comparisons in Appendix Tables 9 and 10 to include LatentSeek and STIR. The results show that PASER achieves better performance on the evaluated benchmarks, further supporting the effectiveness of our approach and its advantages over these closely related methods.

Table R1. Comparison with STIR

ModelMethodAIME2024AMC2023MATH-500Avg.
DeepSeek-R1-Distill-Qwen-7BSTIR65.093.388.082.10
SeLaR60.086.793.380.00
PASER70.096.794.887.12
Qwen2.5-7B-InstructSTIR25.066.775.555.73
SeLaR15.060.076.550.50
PASER25.070.078.357.77

Table R2. Comparison with LatentSeek

ModelMethodGSM8KMATH-500AIME2024Avg.
Qwen2-7BLatentSeek84.3857.43.3348.37
SeLaR83.0953.00.0045.36
PASER85.6758.63.3349.20
Qwen2.5-7BLatentSeek90.1475.613.3359.69
SeLaR90.2276.26.6757.70
PASER91.1376.816.6761.53
Llama3.1-8BLatentSeek78.5454.610.0047.71
SeLaR79.2355.26.6747.03
PASER80.1455.816.6750.87
AuthorsOpenReview ↗

Response to Reviewer RiaV [part 2]

Q2: Top-3 Entropy Selection and Reflective Action Statistics

We thank you for the insightful comments. We further clarify our design rationale and provide additional analyses.

Motivation for Top-3 Normalized Entropy and Reflective Action Statistics. PASER does not average the entropy values of the top-three tokens. Instead, it computes the Shannon entropy over the renormalized probability distribution of the top-3 candidate tokens. This design captures competition among the most probable candidates while reducing interference from the low-probability long tail of the full vocabulary. Normalization maps the entropy to [0,1][0,1], providing a consistent and interpretable criterion for candidate anchor selection. Similar Top-3 entropy measures have also been adopted in prior work, such as SeLaR. Importantly, entropy is used only to identify candidate anchors; whether intervention is warranted is subsequently determined by short- and long-horizon influence evidence.

To further characterize PASER's actual intervention behavior, we provide additional box plots in Appendix Figure 18, showing the per-example counts of KEEP, GUIDE, and ROLLBACK on MBPP (coding) and AIME2024 (mathematical reasoning), using Qwen3-8B and DS-R1-Llama-8B. As shown in Figure 18, ROLLBACK is not frequently triggered. Even on the challenging AIME2024 benchmark, the median number of ROLLBACK operations is only approximately 9–12 per example. This indicates that PASER does not aim to maximize intervention frequency; instead, it selectively applies revisions to decisions with substantial downstream influence.

Top-kk Sensitivity Analysis. To further validate the choice of Top-3, we evaluate alternative candidate-set sizes, including Top-5, Top-7, and Top-10, while keeping other PASER configurations unchanged. The results show that Top-3 achieves the best overall performance, whereas alternative candidate-set sizes lead to only limited performance variations. These findings suggest that Top-3 provides a favorable balance between performance and candidate-set size, while demonstrating that PASER is robust to reasonable changes in the Top-kk selection.

Table R3. Performance of Top-kk

Top-kkMATH-500AIME20242WikiMultiHopQAHumanEvalMBPPAverage
397.6080.8382.0096.9595.3390.54
597.4080.0081.0096.3494.9489.94
797.2079.1781.5095.1294.9489.59
1097.2078.3381.0095.1294.1689.16

We sincerely thank you again for the constructive feedback. We hope these additional analyses and clarifications adequately address the concerns.

Thanks for the authors' reply. Still, some minor concerns exist: PARSE introduces several hyperparameters, making it complicated to tune in real scenarios. Considering that they have addressed most of my concerns, I decided to raise the score.

Thank you for your response, for acknowledging our work, and for increasing your score. Your constructive feedback has improved the quality of our paper. Our PASER adopts a minimal-hyperparameter design for training-free reasoning and easy deployment. Its core mechanisms, Predictive Anchoring, Matched-Suffix Replay, Delayed Influence Diagnosis, and hierarchical reflection (KEEP, GUIDE, and ROLLBACK), remain fixed across all experiments. It exposes only a few interpretable runtime parameters, including �0��0�, �1��1�, �2��2�, which control influence filtering, long-horizon auditing, and reflective action selection, respectively. These parameters allow users to balance reasoning quality and computational cost according to task difficulty and resource constraints, without retraining or modifying the core framework.

Reviewer_D4mjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PASER, a training-free framework for revisable LLM reasoning. Instead of switching to soft embeddings as soon as the model is uncertain, PASER caches the decoding state at high-entropy tokens. It later replays the generated suffix under both discrete and soft histories to measure each anchor's influence, and maps the accumulated influence to one of three actions: keep, guide or roll back. Experiments on math, STEM, coding and general reasoning benchmarks across several model families show that PASER achieves the highest average accuracy against six training-free baselines, including Soft Thinking, SwiR and SeLaR. Ablations examine the anchoring strategy, the diagnosis window, the long horizon K and the reflection thresholds.

Critical strengths
  1. The motivation is reasonable. Separating where to cache an anchor, when to judge it and how much to revise is a sensible refinement over methods that intervene at every high-entropy token, and a preliminary analysis supports it.
  2. PASER has the highest average accuracy in every model group in Tables 1 and 2, across math, STEM, coding and general reasoning benchmarks.
  3. The method is training-free. The baselines use their official implementations and released hyperparameters with the same decoding settings, and the paper reports efficiency in tokens per second, time per request and GPU memory.
  4. The paper provides ablations of its main design choices, plus four-seed results on three benchmarks, where PASER beats SwiR and SeLaR on average for both models tested.
Critical weaknesses

Critical Weaknesses:

  1. The hyperparameters appear to have been chosen on the evaluation benchmarks. δ and K are swept on Qwen3-8B over five benchmarks that also appear in the main results, and the defaults score best in both sweeps. On these five benchmarks, PASER leads the best baseline, SwiR, by 2.61 points at the defaults, but falls below SwiR just two grid steps away (δ = 0.4 or 1.2, K = 128 or 384). Could the authors select hyperparameters on held-out data and report how the results change?
  2. The influence score compares how well the discrete and soft histories explain a suffix that was generated from the discrete history. A suffix naturally fits the history it was generated from, so a large score indicates disagreement between the two histories rather than an error in the reasoning. The paper does not appear to report how often each action is taken, or how often an intervention turns a wrong answer into a correct one. Could the authors report these statistics?
  3. Tables 1 and 2 report no standard deviations. Where four-seed means are available (Tables 13 and 14), they are lower than the corresponding table values in 16 of 18 cells, across all methods, which suggests the main tables may slightly overstate typical performance. Several benchmarks are also small: AIME has 30 problems, so one problem is worth 3.3 points. Could the authors report means and standard deviations over multiple seeds for the main tables?
Questions for authors

See weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer D4mj [part 1]

We sincerely thank you for recognizing the value of our work and for providing constructive feedback, which has helped us further improve the quality of our manuscript.

W1: Design Choices and Hyperparameter Settings

We appreciate the reviewer's constructive comments on PASER's design choices and hyperparameter settings. We agree that distinguishing between fixed framework components and tunable runtime parameters is important for understanding the practicality and generalizability of our method. We provide the following clarifications.

Minimal Hyperparameter Design. PASER follows a minimal-hyperparameter design principle to provide a practical, training-free, and easily deployable reasoning framework. Its core mechanisms, including Predictive Anchoring, Matched-Suffix Replay, Delayed Influence Diagnosis, and the three hierarchical reflection strategies (KEEP, GUIDE, and ROLLBACK), remain fixed across all experiments. PASER exposes only a small number of functionally interpretable runtime parameters, such as δ\delta, KK, and (βL\beta_L,βH\beta_H), which control influence filtering, long-horizon auditing, and reflective action selection, respectively. These parameters allow users to balance reasoning quality and computational cost according to task difficulty and resource constraints, without retraining the model or modifying the framework's core mechanisms.

Original Hyperparameter Selection and Cross-dataset Generalization. We clarify that the original sensitivity analyses of $δ\delta $and KK were conducted exclusively on Qwen3-8B across five benchmarks: GSM8K, MATH-500, AIME2024, AIME2025, and HumanEval. Based on these analyses, we selected �4��4� and K=256K=256 as the default values. These parameters were subsequently fixed and directly applied to other benchmarks in the main experiments, including GPQA Diamond, AMC2023, GSM-Hard, Math MMLU, MBPP, MBPP+, LiveCodeBench, LeetCode Contest, HotpotQA, MuSiQue, and StrategyQA, without further tuning based on their evaluation results. Thus, the original multi-model and multi-benchmark evaluations provide additional, albeit indirect, evidence of the generalizability of the shared default configuration across models and tasks.

Additional Robustness Validation with Independent Data Splits. To further mitigate potential hyperparameter selection bias, we conduct additional experiments using disjoint parameter-selection and evaluation subsets, as reported in Appendix Tables 20 and 21. Specifically, we use 41 problems from HumanEval and 100 problems from MATH-500 for selecting �6��6� and KK, while reserving the remaining 123 and 400 problems, respectively, for independent evaluation. We search for the optimal configurations exclusively on the parameter-selection subsets, freeze the selected values, and then evaluate them on the corresponding held-out subsets. As shown in the tables, the configurations selected on the independent subsets maintain stable performance on the held-out evaluation sets, with only marginal differences from our original default configuration. These results provide further evidence of the robustness of PASER's default hyperparameters and indicate that its performance does not rely on tuning over the full evaluation benchmarks.

Table R1. δ\delta selection on independent splits.

DatasetSplitSamplesδ\deltaKKPass@1 (%)
HumanEvalMining410.425692.69
HumanEvalMining410.625695.12
HumanEvalMining410.825695.12
HumanEvalMining411.025692.69
HumanEvalMining411.225692.69
HumanEvalEval1230.825695.93
HumanEvalAll1640.825696.95
MATH-500Mining1000.425697.00
MATH-500Mining1000.625697.00
MATH-500Mining1000.825698.00
MATH-500Mining1001.025697.00
MATH-500Mining1001.225696.00
MATH-500Eval4000.825697.30
MATH-500All5000.825697.60

Table R2. KK selection on independent splits.

DatasetSplitSamplesδ\deltaKKPass@1 (%)
HumanEvalMining410.812895.12
HumanEvalMining410.819295.12
HumanEvalMining410.825697.56
HumanEvalMining410.832097.56
HumanEvalMining410.838495.12
HumanEvalEval1230.825696.75
HumanEvalAll1640.825696.95
MATH-500Mining1000.812896.00
MATH-500Mining1000.819297.00
MATH-500Mining1000.825697.00
MATH-500Mining1000.832096.00
MATH-500Mining1000.838496.00
MATH-500Eval4000.825697.00
MATH-500All5000.825697.60
AuthorsOpenReview ↗

Response to Reviewer D4mj [part 2]

W2: Statistical Analysis of Reflective Actions

We appreciate the reviewer's insightful comments. To further characterize PASER's actual intervention behavior, we provide additional box plots in Appendix Figure 18, showing the per-example execution counts of KEEP, GUIDE, and ROLLBACK on MBPP (coding) and AIME2024 (mathematical reasoning), using Qwen3-8B and DS-R1-Llama-8B. As shown in Figure 18, ROLLBACK is not frequently triggered. Even on the challenging AIME2024 benchmark, the median number of ROLLBACK operations is only approximately 9–12 per example. These results indicate that PASER does not aim to maximize intervention frequency. Instead, it selectively applies revisions based on downstream influence evidence, concentrating interventions on reasoning decisions with substantial downstream effects.

W3. Multi-seed Evaluation and Performance Stability

We thank the reviewer for emphasizing the importance of statistical reliability. We agree that multi-seed evaluation and reporting standard deviations are essential for assessing performance stability.

First, we clarify that our original experiments already include four independent runs on AIME2024, AIME2025, and AIME2026, with the average accuracy reported for all baseline comparisons. The corresponding evaluation protocol is detailed in Appendix A.3. Furthermore, to provide a broader assessment of performance variability, we conduct additional multi-seed experiments on representative benchmarks and report the mean and standard deviation in Appendix Table 24. As shown in the results, despite variations across random seeds, PASER maintains competitive performance, further supporting the reliability and stability of our experimental findings.

Table R3. Comparison of PASER with baseline methods under four random seeds.

ModelMethodHumanEvalMath500AMC2023M-MMLU2WM-HQAGPQAMBPP
Qwen3-8BSwiR94.51 ± 0.5097.30 ± 0.2688.75 ± 1.4494.00 ± 0.8277.88 ± 0.6360.10 ± 0.4194.65 ± 0.59
SeLaR88.56 ± 0.7796.60 ± 0.1693.13 ± 1.2586.00 ± 0.8277.00 ± 0.7160.36 ± 0.6592.51 ± 0.59
PASER96.34 ± 0.5097.40 ± 0.2398.75 ± 0.8395.25 ± 0.5081.63 ± 0.4862.12 ± 0.4195.04 ± 0.37
DS-R1-Llama-8BSwiR82.63 ± 0.7991.00 ± 0.2386.88 ± 2.3972.50 ± 1.2946.25 ± 1.2645.83 ± 0.6383.27 ± 0.55
SeLaR81.56 ± 1.0486.25 ± 0.3081.25 ± 1.4466.75 ± 0.9647.13 ± 0.6343.94 ± 0.4178.21 ± 0.55
PASER88.41 ± 0.5091.05 ± 0.1091.25 ± 1.4477.25 ± 0.9652.25 ± 0.6549.37 ± 0.4883.47 ± 0.23

We sincerely thank you again for the constructive suggestions and hope that these additional analyses and clarifications adequately address the concerns.

AuthorsOpenReview ↗

Response to Reviewer D4mj

Dear Reviewer D4mj,

Thank you again for your time and constructive comments. As the discussion period is drawing to a close, we would like to kindly check whether our response has addressed your concerns. Looking forward to your reply.

Thanks again for your time and help in improving the quality of this paper.

Regards from the authors.

AuthorsOpenReview ↗

Response to Reviewer D4mj

Dear Reviewer D4mj,

Thank you again for your time and constructive comments. As the author-reviewer discussion period is in its final stages, I wanted to gently follow up to see if our responses have satisfactorily addressed all of your concerns.

Thanks again for your time and help in improving the quality of this paper. Thank you very much for your time and consideration.

Regards from the authors.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary and General Response

Dear Senior Area Chair, Area Chair and all Reviewers,

First, we sincerely thank the reviewers \color{red}{\bbox[#ffeeee]{\text{D4mj}}}, \color{red}{\bbox[#ffeeee]{\text{RiaV}}}, \color{red}{\bbox[#ffeeee]{\text{sKBj}}}, \color{red}{\bbox[#ffeeee]{\text{FBgh}}} for their constructive feedback and for recognizing PASER's motivation (\color{red}{\bbox[#ffeeee]{\text{D4mj}}}, \color{red}{\bbox[#ffeeee]{\text{RiaV}}}, \color{red}{\bbox[#ffeeee]{\text{sKBj}}}, \color{red}{\bbox[#ffeeee]{\text{FBgh}}}), practicality (\color{red}{\bbox[#ffeeee]{\text{D4mj}}}, \color{red}{\bbox[#ffeeee]{\text{RiaV}}}, \color{red}{\bbox[#ffeeee]{\text{sKBj}}}, \color{red}{\bbox[#ffeeee]{\text{FBgh}}}), broad empirical evidence (\color{red}{\bbox[#ffeeee]{\text{RiaV}}}, \color{red}{\bbox[#ffeeee]{\text{sKBj}}}, \color{red}{\bbox[#ffeeee]{\text{FBgh}}}) and performance (\color{red}{\bbox[#ffeeee]{\text{D4mj}}}, \color{red}{\bbox[#ffeeee]{\text{RiaV}}}, \color{red}{\bbox[#ffeeee]{\text{sKBj}}}, \color{red}{\bbox[#ffeeee]{\text{FBgh}}}). In response, we have strengthened the methodological comparisons, multi-seed experiments, and behavioral analysis.


  • Related-work discussion and comparisons (Reviewer \color{red}{\bbox[#ffeeee]{\text{RiaV}}})

We expanded Appendix D to discuss LatentSeek, Soft Reasoning, Reliable Control-Point Selection, and STIR. We clarify that PASER's contribution lies not in feedback, lookahead, or rollback individually, but in separating uncertainty identification from intervention decisions, diagnosing downstream influence through matched-suffix evidence, and selecting an appropriate revision scope. Additional comparisons (Table 9 and 10) with LatentSeek and STIR show that PASER achieves superior performance across the evaluated settings.


  • Hyperparameter selection and robustness (Reviewers \color{red}{\bbox[#ffeeee]{\text{D4mj}}} and \color{red}{\bbox[#ffeeee]{\text{FBgh}}})

We clarified that the default hyperparameters were selected on Qwen3-8B and fixed across other models and tasks without further tuning. We added disjoint-subset experiments and cross-model sensitivity analyses (Table 18-21). The results demonstrate that the selected defaults remain competitive across different data partitions and models.


  • Candidate-anchor robustness and observations (Reviewers \color{red}{\bbox[#ffeeee]{\text{RiaV}}} and \color{red}{\bbox[#ffeeee]{\text{FBgh}}})

We added sensitivity analyses of candidate-set size and entropy thresholds, showing stable performance across the evaluated configurations. We also extended the pilot analyses and introduced Top-3 Candidate Resampling (Figure 12, 13) as an alternative intervention. The results consistently support our observation that high entropy does not necessarily imply substantial downstream distributional change, strengthening the empirical motivation for PASER.


  • End-to-end efficiency (Reviewer \color{red}{\bbox[#ffeeee]{\text{sKBj}}})

We clarified that inference time includes suffix replay, long-horizon auditing, GUIDE, regeneration after ROLLBACK, and discarded tokens. Effective TPS counts only retained output tokens over the complete inference time. Under this comprehensive accounting, PASER achieves lower request latency and higher effective throughput than SeLaR and SwiR across both evaluated models. We also report lower peak GPU memory usage.


  • Action analysis (Reviewers \color{red}{\bbox[#ffeeee]{\text{D4mj}}}, \color{red}{\bbox[#ffeeee]{\text{RiaV}}}, and \color{red}{\bbox[#ffeeee]{\text{sKBj}}})

We added per-example distributions (Figure 18) of KEEP, GUIDE, and ROLLBACK operations, as well as anchor-level action-conditioned accuracy (Table 11,12) analyses. These results provide further insight into PASER's intervention behavior across different tasks and models.


  • Expanded multi-seed evaluation (Reviewer \color{red}{\bbox[#ffeeee]{\text{D4mj}}})

We clarified that the original AIME evaluations already average multiple independent runs. We further added multi-seed evaluations (Table 24) across multiple benchmarks and models, reporting means and standard deviations. PASER achieves the highest mean accuracy, providing additional evidence of consistent performance.


Summary

These revisions further establish PASER's central contribution: rather than treating uncertainty as an automatic signal for intervention, PASER uses downstream trajectory evidence to assess influence and determine the appropriate revision scope. The experiments strengthen the empirical evidence, while the clarified behavioral analysis improve the transparency and rigor of the work.

We respectfully ask Area Chair to consider these revisions alongside the manuscript and our detailed reviewer-specific responses. Thank you for your time. Thanks for all Reviewers.

Best regards,
The Authors of Submission

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.