REFLECT: Intervention-Supported Error Attribution for Silent Failures in LLM Agent Traces

Recent progress in large language models (e.g., GPT-4, PaLM) has led to breakthroughs in reasoning, question answering, and tool integration. Prior work such as Chain-of-Thought prompting and Self-Consistency improved reasoning accuracy, but challenges remain in error diagnosis, especially in silent failure scenarios where outputs appear correct but are semantically wrong. Existing error localization methods rely heavily on classifiers or LLM judges, which often hallucinate or lack causal grounding. As models grow more capable, silent failures dominate, emphasizing the need for methods that ground error attribution in actual execution behavior. This evolution underscores the importance of causal verification techniques to improve model interpretability and debugging.

The core problem is that current error localization approaches lack causal grounding, especially in silent failure cases where no explicit signals indicate errors. These methods often produce unreliable attributions, leading to ineffective debugging and reduced trust. The challenge is to develop a method that can accurately identify the earliest step responsible for a failure, grounded in the model's actual execution, and verify this causality through intervention. Achieving this is critical for deploying trustworthy AI systems, particularly in high-stakes domains where understanding failure sources is essential for safety and compliance.

The main innovations include: 1) a diagnosis-guided approach that predicts candidate error steps; 2) a controlled, prefix-preserving replay mechanism that tests hypotheses without introducing unrelated stochastic variations; 3) a causal verification process based on outcome flips, which confirms whether interventions at specific steps change the final result; 4) a feedback loop that refines attribution based on verified outcomes, forming a complete causal attribution cycle. This integrated framework ensures that error localization is grounded in actual model behavior, satisfies all four key requirements, and outperforms existing methods that lack causal verification.

• �� Step 1: Diagnosis — Use an LLM auditor to analyze the completed trace, predict a candidate earliest error step, and generate a structured repair plan.
• �� Step 2: Targeted replay — Construct a repair plan with a specific intervention, select nearby rollback points, and rerun the model from these points while preserving the original prefix.
• �� Step 3: Verification — Check if the replayed trace corrects the output; if successful, compare original and corrected traces to refine the error attribution.
• �� Step 4: Feedback — Use outcome flips as causal evidence to re-localize the error step, iterating until reliable attribution is achieved or no further improvement is possible.
• �� Key algorithms include structured repair planning, constrained prefix replay, and outcome flip validation, ensuring grounded, targeted, and causal error attribution.

The evaluation employs four benchmarks—WTQ, GAIA, SWE-bench, and BBM—covering diverse tasks like table QA, multi-hop reasoning, and software debugging. The models used are GPT-5.2 as auditor and agent, with metrics including localization accuracy (Exact Match, Off-by-1), explanation similarity, and correction rate. Baselines include prompt-based, correction-based, and scoring methods. Ablation studies assess each component’s contribution. Experiments test scenarios with and without ground-truth answers, emphasizing real-world applicability. Results show that REFLECT consistently outperforms baselines, with significant improvements in localization accuracy and explanation quality, validating the causal verification approach.

REFLECT achieves over 20% higher exact match accuracy than previous methods across all benchmarks, with WTQ reaching 70.8% and SWE-bench 19.4%. It maintains correction rates above 85%, and explanation similarity scores exceed 0.6, indicating high fidelity in error attribution. The contrastive analysis reveals that successful repairs correlate strongly with improved explanations, confirming the causal validation mechanism. Ablation results demonstrate that each component—diagnosis, targeted replay, and outcome flip verification—contributes significantly to overall performance. These findings establish REFLECT as a robust, reliable approach for silent failure localization.

This framework is immediately applicable to model debugging, automated testing, and safety-critical AI deployment, where precise error localization is vital. Developers can leverage REFLECT to identify and fix failure points efficiently, reducing debugging time and increasing system trustworthiness. In industry, it can be integrated into CI/CD pipelines for continuous model improvement. Long-term, REFLECT’s causal verification paradigm could underpin autonomous systems, ensuring they operate safely and transparently even in complex, unpredictable environments. Its ability to provide actionable insights without ground-truth answers broadens its utility across various AI applications.

The approach assumes access to expected answers for verification, limiting its use in open-ended tasks. It may struggle with multiple errors acting jointly, as single-step attribution becomes ambiguous. The computational cost of multiple replays and repair planning can be high, especially for long traces or large models. Future work should focus on scaling the method, handling multi-error scenarios, and reducing inference overhead to enable real-time deployment. Additionally, extending the framework to unsupervised settings remains an open challenge.