ReasoningFlow: Discourse Structures for Understanding LLM Reasoning Traces

Abstract: Large reasoning models (LRMs) produce reasoning traces with non-linear structures, such as backtracking and self-correction, that complicate the evaluation and monitoring of the reasoning process. We introduce ReasoningFlow, a framework that captures the discourse structures of LRM reasoning traces into fine-grained directed acyclic graphs (DAGs). We develop and validate our annotation schema through careful manual annotation of 31 traces (2.1k steps), achieving high inter-annotator agreement, then scale to automatic annotation of 1,260 traces (247.7k steps) spanning three tasks (math, science, argumentation) and five models (Qwen2.5-32B-Inst, QwQ-32B, DeepSeek-V3, DeepSeek-R1, GPT-oss-120B). By analyzing ReasoningFlow graphs, we find: (1) LRMs exhibit structurally similar traces, despite being trained from different base models and potentially non-overlapping post-training data. (2) ReasoningFlow reveals diverse fine-grained reasoning behaviors (e.g., local verification, self-reflection, and assumptions) that can be used for better reasoning trace monitorability. (3) In LRMs, most of the erroneous steps are not used to derive final answers. (4) Mechanistic causal dependencies between steps do not reflect the language-level discourse structure. We release the dataset and code in: https://github.com/jinulee-v/reasoningflow.

• The paper introduces a DAG-based annotation schema (ReasoningFlow) that segments LLM reasoning traces into atomic discourse units and semantic edges.
• It employs eight node types and fourteen edge relations to capture detailed reasoning behaviors such as local verification, self-correction, and hypothetical reasoning.
• Empirical analysis shows that most errors do not propagate, affirming the efficacy of fine-grained trace evaluation and mechanistic probing for LLMs.

ReasoningFlow: Annotation Schema for LLM Reasoning Trace Structures

Motivation and Background

The paper "ReasoningFlow: Discourse Structures for Understanding LLM Reasoning Traces" (2606.05402) introduces a principled framework for capturing and annotating the semantic discourse structures present in reasoning traces generated by Large Reasoning Models (LRMs). Existing annotation frameworks for LLM reasoning traces were either too coarse—typically operating at paragraph or sentence level without sufficient granularity in node or edge labeling—or lacked rigorous validation. Furthermore, discourse and argumentation parsing schemes, such as RST and PDTB, are incongruent with goal-oriented reasoning phenomena (e.g., self-verification, proof-by-contradiction) prominent in autoregressive LLM outputs.

Recognizing the need for a schema capturing the fine-grained, nonlinear, and self-corrective behaviors in LLM-generated traces, ReasoningFlow systematically annotates traces as directed acyclic graphs (DAGs), establishing a reliable substrate for downstream analysis, monitoring, and evaluation of LLM reasoning.

ReasoningFlow Framework

Graph Structure and Semantic Layering

ReasoningFlow formalizes reasoning traces as DAGs, invoking a left-to-right flow dictated by the autoregressive nature of LLMs. Nodes represent contiguous, atomic reasoning units (sub-sentence to sentence granularity), and edges capture semantic relations critical to reasoning behaviors.

Node Types

The schema defines 8 node types, segmenting reasoning traces by functional role:

• Reasoning: Deductive/inductive inference anchored in context or prior nodes.
• Planning: Direction-setting statements for subsequent steps or subgoals.
• Reflection: Metacognitive evaluation and subjective sentiment on previous nodes.
• Fact: Externally verifiable knowledge independent of prior steps.
• Restatement: Paraphrase or copy of prior information.
• Assumption: Explicit hypothetical premises, often to delimit reasoning scope (e.g., proof-by-contradiction).
• Example: Concrete instantiations of abstract concepts.
• Conclusion: Model's final or intermediate answer assertions.

Edge Types

Edges are classified into 14 granular relation labels spanning four functional categories:

• Reason: Logical infer, execution of plans, restatement, fact elaboration, exemplification.
• Plan: Sequential progression (proceed), verification initiation, decomposition, backtracking.
• Reflect: Sentiment-bearing evaluation (positive, negative, uncertain).
• Validate: Propositional alignment (support, attack).

This schema enables tracing the provenance and semantic motivation of each step, unlocking analysis of behaviors like local verification, DFS branching, and hypothetical reasoning.

Annotation Methodology

Manual annotation was conducted for a seed set of traces across math, science, and argumentation tasks, obtaining Krippendorff's $\alpha > 0.8$ on node and edge labeling, confirming schema reliability and interpretability. For scaling, a multi-stage LLM-powered pipeline (segmentation/classification/edge labeling) was utilized, achieving high F1 scores relative to human benchmarks (best model: Gemini-3.1-Pro, NC $=0.865$, EDC $=0.646$).

Empirical Findings

Structural Regularities

Despite disparate model families (Qwen, DeepSeek, GPT-oss) and non-overlapping post-training data, reasoning models produce structurally similar ReasoningFlow graphs. Triplet distribution analysis (node label, edge label, node label) via PCA and JS-divergence outlines domain-induced clustering, with math, science, and argumentation traces each manifesting domain-specific patterns (e.g., chained Planning in argumentation, local verification in science).

Fine-Grained Reasoning Behaviors

Local Verification and Self-Reflection

The framework enables quantitative characterization of local verification—a process whereby LRMs correct intermediate mistakes via short-range reflective steps—which occurs more frequently and has greater causal effectiveness than global solution-level verification. Corrected nodes are more likely than their original erroneous counterparts to serve as premises for final answers, demonstrating the operational impact of in-trace self-verification.

Reflection node sentiment (positive, negative, uncertain) correlates with node quality (correctness, argument quality); positive reflections indicate higher correctness and stronger arguments. This observation enables monitorability of internal model uncertainty and opens opportunities for automated trace quality assessment.

Assumption and Hypothetical Reasoning

Assumption nodes delineate scope for hypothetical reasoning, supporting proof-by-contradiction and DFS search strategies. The prevalence of assumption-attack and assumption-proceed chains (covering ~42% of Assumption nodes) highlights the expressivity required for evaluating higher-order reasoning.

Error Propagation and Stepwise Evaluation

Empirical analysis reveals that only a minority of erroneous steps (14.4%) causally propagate to incorrect final answers. Most errors are either unused (not involved in downstream derivations) or neglected (do not affect the correctness of final answers even when referenced), undermining the efficacy of naive step-level error detectors for global correctness prediction. Integrating ReasoningFlow structures into error evaluation pipelines enables fine-grained localization and characterization of causality, refining stepwise evaluation protocols.

Mechanistic Causal Dependencies

Comparison to mechanistic interpretability methods (e.g., Thought Anchors via causal masking) indicates poor alignment: causal scores are not substantially more predictive of semantic discourse edges than simple proximity heuristics. The gap between model-internal causal dependencies and language-level discourse relations highlights the challenge in faithful tracing of reasoning provenance. Supervised probing on hidden state differences yields better reconstruction of logical dependencies, suggesting that ReasoningFlow structures offer a necessary semantic lens for understanding and improving LLM interpretability.

Practical and Theoretical Implications

The ReasoningFlow schema and dataset advance the monitorability and evaluation of LLM-generated traces, enabling:

• Fine-grained reasoning behavior discovery: Including self-correction, proof-by-contradiction, DFS, and sentiment-based monitorability.
• Error propagation tracing: Differentiating causal, non-causal, and neglected errors, thereby informing the architecture of process reward models and filtering strategies.
• Facilitation of more faithful process evaluation and mechanistic probing: Grounding empirical analyses in semantic discourse structures rather than proxy metrics.
• Generalization across LLMs: Structural regularities observed suggest applicability to a wide range of reasoning models, including future architectures.

Potential future work includes studying reasoning behavior emergence during training, adaptation under compute constraints, and handling conflicting knowledge by analyzing model adaptation in discourse structure space.

Conclusion

ReasoningFlow delivers a rigorously validated annotation schema for LLM reasoning traces, capturing fine-grained semantic behaviors and causal structures as DAGs. It enables precise monitoring, faithful evaluation, and mechanistic interpretability of reasoning processes in LRMs. The framework both addresses the limitations of prior coarse annotation schemes and bridges the gap between formal logic, discourse parsing, and cognitive modeling in LLM research. As ReasoningFlow and its associated dataset become more widely adopted, they are poised to facilitate advancements in trace-driven LLM evaluation, model debugging, and reasoning research.