AI Paper Atlas

Research map / Large language models / Reasoning and chain of thought

Hallucination Detection from Structural Reasoning Model

Jianbo Sun, Pengkun Yang

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a structural reasoning model that extracts a directed acyclic reasoning graph from chain-of-thought traces to detect hallucinations in LLM reasoning.

Problem

Hallucinations in LLM chain-of-thought reasoning are hard to identify because linear traces hide cross-step dependencies and unsupported intermediate claims.

Approach

Model interactions among local reasoning steps as a structure, extracting a directed acyclic reasoning graph from the trace and using it to detect hallucinated intermediate claims, rather than treating the trace as a linear sequence.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use forward flow aggregation instead of reverse InvFlow (ablation)FLOW · All four benchmarksAP (%): 66.57 -> 73.69Improved
use forward flow aggregation instead of MaxErr (ablation)FLOW · All four benchmarksAP (%): 55.37 -> 73.69Improved
use forward flow aggregation instead of MinErr (ablation)FLOW · All four benchmarksAP (%): 31.70 -> 73.69Improved
use forward flow aggregation instead of MeanErr (ablation)FLOW · All four benchmarksAP (%): 73.28 -> 73.69Improved
add DAG adjacency features to step-error features (FULL) (ablation)logistic regression · MATHAP (%): 74.74 -> 77.56Improved
add DAG adjacency features to step-error features (NOSHORTCUT) (ablation)logistic regression · MATHAP (%): 72.98 -> 76.05Improved
use high-fidelity dependency graph instead of parsed raw DAG (ablation)FLOW · All four benchmarksAP (%): 73.69 -> 80.15Improved
score steps locally on a DAG instead of whole-rationale GPT-5.2 judgeFLOW+ · GSM8KAP (%): 41.55 -> 77.70Improved
score steps locally on a DAG instead of whole-rationale GPT-5.2 judgeFLOW+ · HumanEvalAP (%): 46.95 -> 83.08Improved
use FLOW+ aggregation instead of sampling-based detectorsFLOW+ · HumanEvalAP (%): 25.52 -> 83.08Improved
use FLOW+ instead of PDSFLOW+ · HotpotQAAP (%): 88.69 -> 83.23Hurt
use prefix-step judging vs whole-trace judging, then Step-FLOW with base/high-fidelity DAG (ablation)Step-FLOW · MATH (Grok)AP: 53 -> 65.70/72.87Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy