Failed Reasoning Traces Tell You What Is Fixable (But Not by Reading Them)

Problem
This work addresses a gap in the understanding of failed reasoning in post-trained language models, particularly in how to leverage failed inference attempts for improved performance. The authors highlight that current practices often disregard failed traces, treating them as irrelevant, which overlooks valuable information about the nature of failures. This paper is a preprint and has not undergone peer review.

Method
The authors propose a novel approach that extracts three problem-level trajectory features from the distributional signature of failed rollouts, rather than from the text of the failures themselves. These features are designed to capture the recoverability structure of failures, indicating which test-time interventions could potentially rectify a given failure. The method clusters failures into stable regimes and characterizes the failure landscape across different post-training methods. The authors also introduce a training-free routing rule that utilizes these features to enhance the model’s ability to recover from failures. The approach was evaluated on the Steerable-Hard subset, which consists of failures where simple retries are insufficient, and bounded interventions are applicable.

Results
The proposed method achieved an accuracy of $84.3{\pm}4.3\%$, representing a $20\%$ improvement over a majority-class baseline. Additionally, the training-free routing rule provided a $+12.2\%$ lift in performance on the Steerable-Hard subset, demonstrating its effectiveness in identifying and applying appropriate interventions based on the failed traces. The features and routing rule were shown to transfer effectively across two cross-family probes, indicating robustness and generalizability of the approach.

Limitations
The authors acknowledge that their method relies on the assumption that the failed traces contain sufficient information to infer recoverability structure, which may not hold in all scenarios. They also note that the approach does not involve retraining the model, which may limit the potential for deeper structural improvements. Furthermore, the reliance on specific trajectory features may not generalize to all types of reasoning tasks or model architectures.

Why it matters
This research has significant implications for the field of AI, particularly in enhancing the robustness of language models in reasoning tasks. By transforming failed traces into diagnostic objects, the proposed method allows for more informed test-time routing and post-training analysis, potentially leading to more efficient deployment strategies. This work opens avenues for further exploration into the structural characteristics of model failures and their implications for model design and training methodologies, as published in arXiv cs.AI.