What Defines Good Reasoning in LLMs? Dissecting Reasoning Steps with Multi-Aspect Evaluation

TL;DR

This work argues that final-answer correctness is an insufficient signal for evaluating LLM reasoning. It formalizes reasoning quality with two step-level dimensions—relevance (grounding in the problem) and coherence (logical progression)—and introduces CaSE, a causal stepwise evaluator that scores steps using only preceding context to avoid hindsight bias. The authors construct expert-annotated benchmarks (MRa-GSM8K, MRa-MATH) and demonstrate that these aspects predict problem-solving success, enabling both improved data curation for supervised fine-tuning and inference-time guidance that enhances reasoning quality. Empirically, CaSE-outcomes align better with human judgments than full-trace baselines and yield tangible improvements across math benchmarks, especially for smaller models, establishing a practical framework for diagnosing, debugging, and improving LLM reasoning.

Abstract

Evaluating large language models (LLMs) on final-answer correctness is the dominant paradigm. This approach, however, provides a coarse signal for model improvement and overlooks the quality of the underlying reasoning process. We argue that a more granular evaluation of reasoning offers a more effective path to building robust models. We decompose reasoning quality into two dimensions: relevance and coherence. Relevance measures if a step is grounded in the problem; coherence measures if it follows logically from prior steps. To measure these aspects reliably, we introduce causal stepwise evaluation (CaSE). This method assesses each reasoning step using only its preceding context, which avoids hindsight bias. We validate CaSE against human judgments on our new expert-annotated benchmarks, MRa-GSM8K and MRa-MATH. More importantly, we show that curating training data with CaSE-evaluated relevance and coherence directly improves final task performance. Our work provides a scalable framework for analyzing, debugging, and improving LLM reasoning, demonstrating the practical value of moving beyond validity checks.

Figures (10)

• Figure 1: Previous evaluations emphasize step- or answer-level correctnesslightman2023letssong2025prmbenchzeng2023mrzeng2024mr. We formalize two causal dimensions at the step-level: relevance (grounding to the question) and coherence (logical consistency with prior steps).
• Figure 2: An example of the proposed fine-grained multi-aspect evaluation. We enable more granular diagnosis of reasoning quality by shifting from answer-level to step-level evaluation and extending the criteria beyond correctness to include relevance and coherence. Corr, Rel, and Coh denote Correctness, Relevance, and Coherence, respectively. Dotted boxes indicate the real datasets; the gray boxes on the right are explanations we added to aid understanding.
• Figure 3: Confusion between solution-level correctness and relevance/coherence labels (left), and breakdown of incorrect solutions in MRa-GSM8K by answer outcome, highlighting the proportion satisfying both relevance and coherence (right).
• Figure 4: The distribution of average step-level Correctness, Relevance, and Coherence across solutions (y-axis), grouped by answer-level correctness (x-axis).
• Figure 5: Average accuracy across three seeds on AIME24 and AIME25 for each method (left), and the prompt example used with QwQ-32B (right). The upper bound of each bar indicates the highest performance across the three runs, while the red dashed line shows the result when only correctness is emphasized (+Corr).