acceptodds
Under review as a conference paper at ICLR 2027

From Traces to Strategies: Training-Free Efficient Reasoning via Execution Feedback

Abstract

Large reasoning models (LRMs) often exhibit overthinking, producing unnecessarily long chains of thought that increase computational cost. Existing approaches shorten reasoning through additional training or inference-time intervention, but remain limited by training cost, model access requirements, or insufficient guidance for specific problem-solving processes. To this end, we propose TraCE, a training-free framework that derives instance-specific reasoning strategies through trajectory analysis and execution feedback. TraCE extracts reusable efficiency rules from native correct reasoning trajectories and iteratively validates them through execution feedback to obtain effective rule sets. The validated rules are then consolidated into a domain-level rule repositories and further induces efficiency guidance shared within each domain. Given a new problem, TraCE first routes it to the most relevant domain, selects relevant and complementary rules, and compiles them into an instance-specific reasoning strategy. The compiled strategy is directly provided to the target model within a single problem-solving call, without updating model parameters. TraCE could be directly deployed on large-scale, production-grade flagship models. We evaluate TraCE on DeepSeek-V4-Pro and GLM-5.3-Flash across both Reasoning and Agent tasks. Experiments show that TraCE consistently reduces reasoning tokens and performs effectively across heterogeneous task settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Efficient Reasoning, Training-Free, Large Reasoning Models

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

This paper proposes TraCE, a training-free framework for improving the reasoning efficiency of large reasoning models. The framework extracts reusable efficiency rules from successful reasoning trajectories, iteratively refines these rules through execution feedback, and consolidates them into domain-level rule repositories. During inference, TraCE employs embedding-based domain routing and MMR-based rule selection to construct instance-specific reasoning strategies without updating model parameters. The authors evaluate TraCE on DeepSeek-V4-Pro and GLM-5.3-Flash across mathematical reasoning, scientific reasoning, and agent planning benchmarks. The results show substantial reductions in output tokens while generally maintaining competitive task accuracy. Additional ablation studies investigate the contributions of instance-level rules, domain-level guidance, and iterative execution feedback.

Critical strengths

S1. Practical Training-Free Framework.

TraCE offers a practical approach to improving reasoning efficiency without requiring additional model training or access to internal model states. This makes the framework potentially applicable to large-scale and black-box reasoning models.

S2. Promising Empirical Results Across Multiple Tasks.

The experiments cover mathematical reasoning, scientific reasoning, and agent planning benchmarks on two large reasoning models. TraCE achieves substantial reductions in output tokens across multiple settings while generally maintaining competitive accuracy. The ablation studies provide preliminary evidence supporting the contributions of its individual components.

**S3. Execution Feedback-Guided Rule Refinement.

TraCE incorporates execution feedback to iteratively refine efficiency rules based on both answer correctness and reasoning length, rather than relying solely on rules extracted from existing trajectories. This provides a useful mechanism for improving the effectiveness of reasoning strategies and reducing unnecessary computation.

Critical weaknesses

W1. Unclear Methodological Novelty Beyond Existing Experience and Skill Learning Frameworks.

TraCE follows a pipeline similar to existing experience-based learning and skill optimization frameworks, such as ExpeL, by extracting reusable knowledge from successful trajectories, refining it through feedback, and retrieving relevant guidance for new tasks. While TraCE focuses on reasoning efficiency, it remains unclear what fundamentally distinguishes its methodology from adapting existing experience-learning approaches to this objective. Moreover, components such as domain-level rule consolidation and MMR-based retrieval largely rely on established techniques. The paper would benefit from a clearer explanation of its unique methodological contributions and direct comparisons with existing frameworks under the same reasoning-efficiency objective.

W2. Insufficient Validation of Rule Effectiveness, Applicability, and Generalization.

TraCE primarily validates rule modifications based on correctness and reduced output length on the same source instance, which does not necessarily guarantee their effectiveness or transferability to unseen problems. Furthermore, embedding-based domain routing and MMR-based rule selection rely on semantic similarity, which may not accurately capture the underlying reasoning structure or determine whether a rule's applicability conditions are satisfied. Although the framework instructs the model to ignore inapplicable rules, it lacks explicit applicability verification. Domain-level consolidation may also introduce overly broad rules or lose important applicability conditions. While the paper demonstrates downstream improvements on held-out benchmarks, it lacks systematic evaluations of individual rule validity, retrieval reliability, and robustness to rule–instance mismatches.

W3. Insufficient Statistical Reliability and Accuracy–Efficiency Trade-off Analysis.

The paper reports substantial reductions in output tokens, but the statistical reliability of the results remains insufficiently examined. Several accuracy differences are relatively small, particularly on benchmarks with limited sample sizes, and the paper primarily reports point estimates without confidence intervals or repeated-run variance. Moreover, certain settings exhibit noticeable accuracy degradation, such as GLM-5.3-Flash on AIME25, where accuracy decreases from 100% to 90%. Without repeated evaluations and statistical uncertainty analysis, it is difficult to determine whether the reported accuracy differences are robust. A more systematic analysis of accuracy–efficiency trade-offs would further strengthen the empirical conclusions.

Questions for authors

Q1. Rule Effectiveness and Generalization.

How do the authors verify that extracted rules are genuinely effective and transferable beyond the source instances used during refinement? Could the authors evaluate individual rules on held-out problems and report their validity, applicability, and effectiveness, including whether domain-level consolidation preserves these properties?

Q2. Retrieval Reliability and Robustness.

How reliably does embedding-based similarity identify rules that are actually applicable to a target problem? Could the authors evaluate the impact of retrieving irrelevant or mismatched rules, and compare the current retrieval mechanism with applicability-aware alternatives?

Q3. Statistical Significance and Evaluation Variance.

Are the reported accuracy and token-efficiency improvements statistically significant, particularly when performance differences are small? Could the authors provide confidence intervals, paired statistical comparisons, and robustness analyses across different random seeds or rule repositories?

Q4. Novelty and Comparison with Existing Frameworks.

What are the key methodological differences between TraCE and existing experience-based learning frameworks, such as ExpeL? Could the authors provide more direct comparisons under comparable computational budgets and reasoning-efficiency objectives to better establish the contribution of the proposed framework?

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

This paper proposes TraCE, a training-free framework for reducing redundant reasoning in large reasoning models. TraCE extracts reusable efficiency rules from correct native reasoning trajectories, refines them through iterative execution feedback, consolidates them into domain-level repositories and guidance, and compiles instance-specific strategies at inference time using embedding-based routing and MMR-based rule selection. Experiments on DeepSeek-V4-Pro and GLM-5.3-Flash across mathematical reasoning, scientific reasoning, and planning benchmarks demonstrate substantial reductions in output length.

Critical strengths
  1. Practical motivation and deployment compatibility. TraCE does not require parameter updates or access to internal activations, making it applicable to black-box model APIs.
  2. A coherent framework for experience reuse. The combination of instance-level rule refinement, domain-level consolidation, and hierarchical strategy compilation provides a structured way to reuse problem-solving experience.
  3. The experiments cover two backbone models and multiple reasoning and planning benchmarks, with ablations and case studies that help explain the observed efficiency gains.
Critical weaknesses
  1. Incomplete accounting of the overall computational cost. The evaluation primarily measures output length, while the proposed framework requires substantial offline trajectory generation, iterative rule refinement, and consolidation. The additional input tokens introduced by compiled strategies are also not explicitly accounted for. Consequently, the reported reductions do not establish that TraCE reduces end-to-end token consumption or monetary cost. The paper should report offline construction costs, online input/output tokens, latency, and amortized costs under different query volumes.

  2. Unclear trade-off between efficiency and correctness. Although TraCE substantially reduces output length, accuracy drops from 100% to 90% on GLM-5.3-Flash on AIME25, with smaller regressions on other benchmarks. Since shorter incorrect responses can also improve the output-length metric, it is important to evaluate the accuracy–efficiency frontier and report uncertainty or repeated-run results. The refinement-iteration analysis should also track accuracy, not only output length.

  3. Limited evidence for rule transferability. Candidate rules are accepted based on correctness and shorter outputs on their originating instances. This validates their local utility but does not directly establish that the rules generalize reliably to unseen instances. Despite the encouraging aggregate benchmark results, the paper lacks rule-level analysis of transfer success, overgeneralization, and failure cases. Independent cross-instance rule validation would strengthen the central claim.

  4. Incomplete comparison with closely related training-free approaches. The related-work section discusses CROP, which optimizes prompts using length and accuracy feedback, but this method is not included in the experiments. The paper should justify this omission or provide a comparison. In addition, the cost accounting for baselines involving parallel sampling should be clarified to ensure that all methods are evaluated under comparable cost definitions.

  5. The scope of the generalization claims needs clarification. Each backbone independently performs the generative stages of TraCE, so the results establish that the framework can be instantiated for multiple models, rather than that the learned rules transfer across models. Furthermore, the planning benchmarks primarily evaluate planning-related reasoning rather than interactive agent execution. The claims about cross-model reuse and general agent applicability should be scoped accordingly.

Questions for authors

Refer to the weaknesses.

Reviewer_pP3qOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TraCE, a training-free framework that extracts efficiency rules from correct reasoning traces, refines them through execution feedback, consolidates them by domain, and retrieves instance-specific guidance at inference time. Experiments on 2 models show substantial output-token reductions, with task-dependent accuracy trade-offs.

Critical strengths
  1. Practical deployment through text-based guidance and a single target-model call.

  2. Rule refinement uses correctness and length feedback rather than unvalidated reflection alone.

  3. Substantial output-token reductions across diverse reasoning and planning benchmarks.

Critical weaknesses
  1. Novelty over experiential learning (ExpeL[1], ACE[2]) and prompt optimization (CROP[3]) is insufficiently established, and the comparison is uneven: TraCE uses in-distribution seed problems and extensive offline execution, whereas all baselines are zero-shot prompts. Gains may therefore partly reflect task-knowledge injection (e.g., GLM Meeting accuracy 64.0→73.2) rather than efficiency alone.

  2. Cost evaluation excludes offline construction, added input tokens, and end-to-end latency.

  3. Accuracy preservation is overstated, with notable regressions, e.g., GLM AIME25 100→90, PlanBench-Mystery 96.0→92.0 reported from single runs without uncertainty estimates or decoding/max-length settings.

  4. Execution-feedback gains are not isolated by budget-matched controls, and key details are missing, such as number of selected rules m, embedding model, domain construction, seed/test splits.

[1] A. Zhao, D. Huang, Q. Xu, M. Lin, Y.-J. Liu, and G. Huang, "ExpeL: LLM agents are experiential learners," in Proc. 38th AAAI Conf. Artif. Intell. (AAAI), 2024, vol. 38, no. 17, Art. no. 2188, doi: 10.1609/aaai.v38i17.29936.

[2] Q. Zhang, C. Hu, S. Upasani, B. Ma, F. Hong, V. Kamanuru, J. Rainton, C. Wu, M. Ji, H. Li, U. Thakker, J. Zou, and K. Olukotun, "Agentic context engineering: Evolving contexts for self-improving language models," in Proc. 14th Int. Conf. Learn. Represent. (ICLR), 2026. [Online]. Available: https://openreview.net/forum?id=eC4ygDs02R

[3] D. Shah, S. Badhe, N. Kathrotia, and P. Tiwari, "CROP: Token-efficient reasoning in large language models via regularized prompt optimization," in ICLR 2026 Workshop on Logical Reasoning of Large Language Models, 2026. [Online]. Available: https://openreview.net/forum?id=O2LmrF5Yzq

Questions for authors
  1. How does TraCE compare with CROP, efficiency-adapted ExpeL/ACE, and few-shot prompting with concise seed traces under matched seed data and offline budgets?

  2. What are the total costs, latency, and break-even workload after including offline construction and added inputs?

  3. Do repeated evaluations confirm the reported accuracy trade-offs, particularly on GLM AIME25?

  4. Can you provide a budget-matched no-feedback control, accuracy across refinement iterations, and the missing settings?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.