The Shape of Reasoning: Topological Analysis of Reasoning Traces in Large Language Models

A new framework applies topological data analysis to evaluate reasoning traces from large language models, moving beyond manual rubrics and simple graph-based proxies, according to research posted on arXiv [1]. The study, submitted in October 2025 and last revised in May 2026, addresses what its authors describe as an understudied and labor-intensive area: assessing the quality of step-by-step reasoning produced by large language models [1]. Current evaluation methods rely on expert rubrics, manual annotation, and slow pairwise judgments, while automated approaches typically use graph-based proxies that measure structural connectivity but do not clarify what constitutes high-quality reasoning [1]. The researchers argue that such abstractions can be overly simplistic for inherently complex processes [2]. The introduced framework uses topological data analysis to capture the geometry of reasoning traces, enabling label-efficient, automated assessment [2]. Topological data analysis draws on concepts from geometry, a branch of mathematics concerned with properties of space such as distance, shape, size, and relative position of figures [4]. The field has expanded dramatically since the 19th century, with subfields including differential geometry, algebraic geometry, and algebraic topology [4]. Mathematics more broadly uses logical reasoning and proof to establish properties of abstract concepts, often expressed as theorems and equations [5]. In an empirical study, topological features yielded substantially higher predictive power for assessing reasoning quality than standard graph metrics [1]. This finding suggests that effective reasoning is better captured by higher-dimensional geometric structures rather than purely relational graphs [2]. The authors further identified a compact, stable set of topological features that reliably indicates trace quality, which they propose as a practical signal for future reinforcement learning algorithms [1]. The submission history shows the paper was first uploaded on 23 October 2025 at a size of 15,472 KB, with subsequent revisions in May 2026 reducing the file to 15,387 KB [1].