SliceGraph: Mapping Process Isomers in Multi-Run Chain-of-Thought Reasoning

2026-05-14 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

SliceGraph is a novel post-hoc problem-model-cell graph designed to map process isomers in multi-run chain-of-thought (CoT) reasoning. This graph is constructed using mutual-kNN over sparse activation-key Jaccard similarity between CoT slices, treating it as a measurement object for process geometry. Across sampled CoT ensembles from three primary 4B/8B models on math and science benchmarks, SliceGraph revealed that biconnected components represent shared reasoning-state units and process families signify strategy-coherent route units. The research found that in 85.5% of 954 problem-model cells, correct CoTs with the same normalized answer split into multiple process families. Among cells with at least two such runs, 76.6% of run pairs were cross-family on average, indicating the presence of "process isomers"—same-answer, family-divergent correct trajectories. This structured multi-route process geometry is often overlooked by final-answer aggregation.

Key takeaway

For research scientists evaluating large language model reasoning, you should consider analyzing the underlying process geometry of chain-of-thought outputs. Relying solely on final-answer aggregates can obscure critical insights into how models arrive at solutions, potentially missing diverse and equally valid reasoning paths. Implementing tools like SliceGraph can reveal process isomers, offering a richer understanding of model capabilities and informing more nuanced model development or fine-tuning strategies.

Key insights

Multi-run CoT reasoning exhibits "process isomers," where diverse reasoning paths lead to identical correct answers.

Principles

• Reasoning trajectories are not monolithic.
• Final answers obscure process diversity.

Method

SliceGraph constructs a problem-model-cell graph using mutual-kNN over sparse activation-key Jaccard similarity to map CoT process geometry and identify process isomers.

In practice

• Analyze CoT diversity beyond final answers.
• Identify distinct reasoning strategies.

Topics

• SliceGraph
• Chain-of-Thought Reasoning
• Process Isomers
• Multi-run CoT Analysis
• Reasoning Process Geometry

Best for: Research Scientist, AI Scientist

Related on AIssential

• Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models — Large language models exhibit a "commitment boundary" in Chain-of-Thought reasoning, after which subsequent steps are epiphenomenal.
• Reasoning Structure of Large Language Models — Evaluating Large Reasoning Models requires analyzing their internal logical flow, not just final accuracy or token count.
• AILS-NTUA at SemEval-2026 Task 12: Graph-Based Retrieval and Reflective Prompting for Abductive Event Reasoning — AILS-NTUA's three-stage system won SemEval 2026 Task 12 for Abductive Event Reasoning, achieving a first-place rank with an accuracy score of 0.95.
• Livecoding: ICE and the Factored Cognition Primer by Ought — Iterated decomposition and execution tracing significantly enhance language model reliability and debugging.
• Correct Chains, Wrong Answers: Dissociating Reasoning from Output in LLM Logic — LLMs can produce correct reasoning steps but still output wrong final answers, revealing a reasoning-output dissociation.
• CoT-Space: A Theoretical Framework for Internal Slow-Thinking via Reinforcement Learning — CoT-Space reframes LLM reasoning as continuous optimization in a semantic space, explaining optimal CoT length.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.