acceptodds
Withdrawn from ICLR 2027

Reasoning Change-Points (RCP): Training-Free First-Error Localization from Hidden-State Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Hidden-state representations provide a rich substrate for diagnosing multi-step reasoning, yet when reasoning failures leave a reliably localizable boundary in representations remains insufficiently understood. Our key insight is to separate the location of a reasoning error from the representational change that makes that error detectable. We find that boundary strength varies systematically across reasoning regimes: controlled perturbations often induce sharp and reproducible transitions, while naturally occurring failures exhibit subtler structure under passive inspection. Importantly, this structure is not intrinsic to the failure alone: re-evaluating reasoning prefixes under an explicit verification instruction substantially strengthens natural localization, showing that first-error localizability depends not only on the reasoning failure itself but also on the computation used to inspect it. Building on this observation, we introduce Reasoning Change-Points (RCP), a reliability-aware framework that detects local transitions in step-level hidden states, estimates the reliability of candidate error boundaries, and strengthens ambiguous natural evidence through verification conditioning, without training an auxiliary verifier. Across six 4B–8B models spanning multiple model families on GSM8K and MATH-500, controlled failures achieve 0.51–0.90 exact localization, and a prospectively specified Qwen3-32B evaluation reproduces the predicted depth structure. Across four ProcessBench domains, verification conditioning improves exact natural localization by 17.5–22.8 percentage points over passive hidden-state trajectories, indicating confidence calibrated on controlled failures transfers to natural settings, substantially enriching for correctly localized examples. Downstream experiments show that informative boundaries can support targeted regeneration and motivate adaptive reasoning control. Together, these findings characterize a soft reliability boundary for hidden-state first-error localization and show how that boundary can be strengthened and selectively exploited when the resulting evidence is reliable.

Primary area:
interpretability and explainable AI
Keywords:
Chain-of-Thought Reasoning, Failure Localization, Change-Point Detection, Hidden-State Dynamics, Geometric Innovation, Targeted Repair, Large Language Models

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

The paper presents Reasoning Change-Points (RCP), a training-free approach for identifying the first erroneous step in LLM reasoning by detecting shifts in hidden-state representations. The authors demonstrate that artificially injected errors yield more distinct hidden-state transitions than naturally occurring errors, and introduce verification conditioning to amplify these signals. Experiments across multiple models and mathematical reasoning benchmarks show improved first-error localization, with confidence-based selection and targeted regeneration explored as downstream applications.

Critical strengths
  1. The proposed approach is training-free and is relatively straightforward to understand and implement.
Critical weaknesses
  1. Limited novelty and missing baselines. Prior work has already explored hidden-state-based step-level correctness detection (e.g. https://arxiv.org/abs/2511.06209), as well as step-level uncertainty quantification (https://arxiv.org/pdf/2502.11250), and resampling at detected error positions (https://arxiv.org/abs/2507.14958) for reasoning models. The proposed approach seems to primarily combine existing ideas, but lacks direct comparisons with these related methods. Without such comparisons, it is difficult to assess its practical advantages over potentially simpler alternatives. Comparisons with some standard Process Reward Models (PRMs) would also provide useful reference points.

  2. Limited experimental evaluation. Experiments rely on off-policy reasoning traces (e.g., ProcessBench). It remains unclear whether the method generalizes to other model families, domains, or errors in the model's own generations.

Questions for authors
  1. Most importantly, could the authors compare against relevant hidden-state probing and uncertainty-based verification methods mentioned in the Weaknesses section? I see that the proposed method is training-free, but I would argue that requiring a small amount of training data is not necessarily a major practical limitation, so a comparison with change-point detection based on hidden-state probing would be beneficial.

  2. Since the method already uses verification conditioning, could the authors report how accurately the backbone LLM judges the correctness of its own reasoning steps? This would provide another useful baseline for the method.

  3. What motivates the specific weighting 0.4P+0.3L+0.3D? The choice seems somewhat arbitrary.

  4. It would be valuable to include on-policy evaluations, where models detect errors in their own reasoning, as well as experiments across additional domains. Detecting errors in another model's reasoning may differ from detecting one's own mistakes.

Reviewer_64GXOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Reasoning Change-Points (RCP), a training-free framework for locating the first error in a reasoning trace using changes in step-level hidden states. It studies controlled corruptions and naturally occurring errors, and uses verification-conditioned prefix representations to strengthen the natural-error signal. A confidence score combines transition prominence, layer agreement, and detector agreement to select predicted boundaries. Experiments cover controlled mathematical reasoning, ProcessBench localization, and targeted regeneration.

Critical strengths
  • Reproducing a teacher’s reasoning text does not necessarily mean that the student uses the described reasoning to produce its answer. The paper addresses this distinction with a regularizer that can be added to existing distillation objectives without circuit tracing inside the training loop.

  • CAD improves many combinations of teacher, student, and training objective, including transfer to related evaluation datasets. For example, standard CoT distillation on GSM8K improves from 18.94% to 38.50%. The comparison with typo augmentation is also useful: it shows that simply adding perturbed training examples does not reproduce the full benefit.

Critical weaknesses
  • ProcessBench results primarily compare verification-conditioned representations with passive hidden-state inspection. This establishes an improvement over the paper’s own passive method, but not competitiveness with process reward models or explicit step-verification baselines. Appendix O briefly mentions a supervised PRM baseline without supplying an identifiable model and corresponding comparative results. Since verification conditioning evaluates every accumulated prefix, a same-model explicit verifier with comparable inference cost is particularly important. Without these controls, it is unclear whether the gain comes from the proposed hidden-state readout or simply from asking the model to verify the reasoning.

  • The main text introduces several signals, detectors, layers, and confidence components, but leaves important operational details scattered across the appendix or external artifacts. Table 1 does define its abbreviated configurations, and Appendix I formalizes the detectors; nevertheless, there is no unified algorithm covering verification conditioning, boundary prediction, and confidence transfer. In particular, the normalization and calculation of P,L,DP,L,D in C=0.4P+0.3L+0.3DC=0.4P+0.3L+0.3D, and their application to the fixed natural-localization protocol, need an explicit account. Appendix R.2 supplies only the verification prompt’s semantics rather than its exact text. Tables also mix configurations, metrics, and evaluation regimes in small, heavily abbreviated layouts, making the main evidence unnecessarily hard to audit.

  • The controlled calibration procedure targets a 95% lower confidence bound of at least 0.80 accuracy. After transfer, accepted-case exact accuracy on the four natural domains is only 0.346–0.512 at 0.138–0.396 coverage (Table 2). This supports selective enrichment, but does not demonstrate preservation of the controlled reliability target. The frozen three-seed study is useful, but covers one controlled Qwen3-8B/GSM8K setting. It does not establish robustness of the central natural-verification or repair results to prompt and sampling choices. The paper should distinguish these scopes explicitly. I do not regard the available evidence as establishing cherry-picking, given the reported held-out selection and complete grids.

Questions for authors
  1. Can you compare natural first-error localization with established PRMs and a same-model explicit step verifier under the same candidate population and comparable inference budgets?
  2. Can you provide the exact verification prompt, hidden-state readout position, confidence-component definitions and normalization, and an end-to-end algorithm for the natural setting? How are first-step errors and entirely correct traces handled?
  3. What reliability claim is intended after transfer, given that accepted-case natural accuracy remains far below the controlled 0.80 target?
  4. Can you report sensitivity to verification-prompt wording and repeated natural-repair runs?
Reviewer_4LPkOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors introduce a framework named Reasoning Change-Points (RCP) where this pipeline can detect local transitions in step-level hidden states and scores how reliable the resulting boundary is for first-error localization in a model’s reasoning. They evaluated this framework across six 4B-8B models in math problems (GSM8K and MATH500) with injected errors and natural errors on Qwen3-8B across four domains in ProcessBench and Qwen3-32B on GSM8K.

Critical strengths

S1. The problem setup explored is well-motivated since being able to find the first erroneous reasoning step is useful for more insightful inspection of a model’s reasoning.

Critical weaknesses

W1. The experiment setup of the controlled failure has a potential confound that RCP may be picking up odd-looking text (since it is injected) rather than broken reasoning. For instance, quantity swaps in table 20, show low change in final answer, 14%, yet for most models, they are localized more accurately than arithmetic corruptions in table 32.

W2. A lot of concepts introduced here are loosely defined. What exactly is a “boundary”? It sometimes implies the true first-error step, sometimes the predicted step or the strength of the signal. Also, “boundary strength” is not really measured directly but rather inferred. The components of the confidence score (P,L,D) are also loosely introduced but not concretely defined.

W3. There is limited evaluation because there aren’t comparisons with existing error-localization methods. The authors can also do another evaluation wherein they ask the model with every prefix if the error exists and choose the first “yes” to see if the hidden state adds anything beyond this. The natural error results also don’t have a random baseline.

Questions for authors

Q1. For W1, can you compare RCP with per-token surprisal baseline and also report localization accuracy? This could help rule out the confound.

Also consider W2 and W3 as questions of clarification and additional requests for additional evaluation.

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

Related papers

Discussion (0)

Sign in to comment.