GRACE: Step-Level Benchmark for Faithful Reasoning over Context

2026-06-14

AI Summary

Current reasoning models are prone to silently deviating from contextual evidence during individual steps of chain-of-thought (CoT) reasoning, yet existing approaches can only detect hallucinations at the final answer level and fail to localize the position or type of errors. To address this gap, this work proposes GRACE—the first step-level faithfulness benchmark for contextually grounded reasoning. GRACE features human annotations of CoT trajectories from multiple models across diverse datasets, with each reasoning step labeled for faithfulness, error category, and natural language explanation. Building on unsupervised clustering, it introduces a data-driven, fine-grained taxonomy of eight error types, formalized as GRACE-Inference and GRACE-Grounding. Experiments reveal substantial deficiencies in current models’ step-level faithfulness, and incorporating these signals into reinforcement learning simultaneously improves downstream task accuracy and reasoning reliability.

Abstract

Many reasoning tasks require models to reason over input context, from document-grounded question answering to rule-based deduction. Chain-of-Thought (CoT) prompting produces traces that appear transparent, yet individual steps can silently deviate from the source evidence, even when the final answer is correct. Existing methods detect hallucinations at the response level but fail to identify where in the chain a failure occurs or what type it is. We introduce GRACE, the first human-annotated step-level faithfulness benchmark with a data-driven error taxonomy for context-grounded textual reasoning. GRACE covers CoT traces from 10 models across 4 source datasets, with each step annotated for faithfulness, error category, and natural language explanation. A data-driven taxonomy, discovered bottom-up via unsupervised clustering, organizes failures into two tracks: GRACE-Inference (deductive errors) and GRACE-Grounding (factual grounding errors), with four categories each. The evaluation set is human-annotated and challenging by design. Our experiments reveal substantial headroom for current models. In addition, integrating step-level faithfulness signals into reinforcement learning pipelines improves both downstream accuracy and reasoning reliability.

Problem

• faithful reasoning
• chain-of-thought
• hallucination detection
• step-level evaluation
• context grounding

Innovation

• step-level faithfulness
• Chain-of-Thought reasoning
• hallucination detection
• error taxonomy
• context-grounded reasoning