ReasoningFlow: Discourse Structures for Understanding LLM Reasoning Traces

ReasoningFlow is a framework that maps the complex, non-linear reasoning traces of large reasoning models into directed acyclic graphs, enabling better understanding and monitoring of AI reasoning processes. Through analysis of 1,260 traces across multiple models and tasks, researchers discovered that LRMs exhibit structurally similar reasoning patterns despite different training origins, while most erroneous steps don't influence final answers.

ReasoningFlow addresses a critical gap in AI interpretability by providing a structured way to visualize and analyze how large reasoning models arrive at conclusions. As LRMs like DeepSeek-R1 and GPT-4o become more capable, understanding their internal reasoning processes becomes essential for deployment in high-stakes domains. The framework transforms unstructured reasoning traces into fine-grained DAGs, capturing behaviors like backtracking, self-correction, and verification steps that traditional step-by-step analysis misses.

The research reveals convergent reasoning architectures across independently developed models, suggesting that post-training procedures for reasoning may naturally gravitate toward similar structural patterns. This finding has implications for model safety and alignment—if reasoning structures are largely inevitable rather than model-specific, interventions targeting these common patterns could broadly improve reasoning reliability. The discovery that mechanistic dependencies don't align with discourse structures is particularly significant, indicating that layer-level causal analysis and language-level reasoning require different interpretability approaches.

For developers building on top of LRMs, ReasoningFlow offers practical tools for monitoring reasoning quality in production systems. The insight that erroneous steps rarely propagate to final answers suggests models have built-in error-correction mechanisms, though this also indicates potential inefficiencies in reasoning paths. The released dataset and code enable further research into reasoning interpretability, advancing the field's ability to audit and improve LRM behavior. This work bridges the gap between mechanistic interpretability and human-readable reasoning analysis.

• →ReasoningFlow captures non-linear LRM reasoning into directed acyclic graphs, enabling fine-grained analysis of reasoning traces across multiple models and tasks.
• →Large reasoning models exhibit structurally similar reasoning patterns despite different training origins, suggesting convergent post-training procedures.
• →Most erroneous reasoning steps do not influence final answers, indicating LRMs have built-in error-correction mechanisms.
• →Mechanistic causal dependencies between steps differ from language-level discourse structures, requiring separate interpretability approaches.
• →The publicly released dataset and code enable developers to build better monitoring and auditing systems for reasoning model outputs.