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

Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 RACE, a black-box hallucination detector for large reasoning models that combines reasoning-trace consistency, SINdex answer uncertainty, reasoning-answer alignment, and reasoning coherence over distilled chain-of-thought steps.

Problem

Existing sampling-based hallucination detectors rely on final-answer uncertainty and miss hallucinations or logical inconsistencies in the explicit reasoning traces of large reasoning models.

Approach

Jointly scores four signals (reasoning consistency via NLI on distilled CoT steps weighted by attention, SINdex answer uncertainty, LNPE-based reasoning-answer alignment, and entity-based internal coherence) after distilling reasoning traces with a fine-tuned CoT extractor.

Key results

• RACE achieves the best AUROC across most LRMs; e.g., on HotpotQA with DS-8B it reaches 78.56 vs 74.98 for SINdex.
• On HotpotQA with QwQ-32B, RACE attains 81.01 AUROC versus 78.19 for SINdex.
• Applied to Qwen2.5-14B-Instruct with CoT, RACE reaches 79.87 on HotpotQA versus 77.19 for the best baseline SINdex.

Surprising findings

• Reasoning-consistency score alone (S_CC) is the strongest single component on HotpotQA, outperforming answer-level SINdex alone.
• A simple reasoning-trace baseline (S_RR) performs poorly, below answer-level baselines in many settings, while RACE's extracted-CoT version is strong.

Limitations

• Requires sampling multiple outputs (about 40% extra inference time) plus an extra 8B CoT extractor and NLI model, adding overhead.
• Equal component weights are used in main results; optimized weights (RACE+) gain depends on tuning on a held-out split and varies across models and datasets.
• Evaluated only on closed-book/short-answer QA datasets with LLM-as-judge labels.

Experimental setup

Datasets: TriviaQA SQuAD NQ-Open HotpotQA 2WikiMultihopQA

Benchmarks: TriviaQA SQuAD NQ-Open HotpotQA

Metrics: AUROC

Baselines: LNPE P(true) Semantic Entropy SEU SelfCheckGPT-NLI SINdex S_RR

Builds on: SINdex Semantic Entropy SelfCheckGPT LNPE

Closest papers by meaning

• Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.92)
• Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.92)
• MIRAGE: Assessing Hallucination in Multimodal Reasoning Chains of MLLM (NeurIPS 2025 · similarity 0.92)
• FactTest: Factuality Testing in Large Language Models with Finite-Sample and Distribution-Free Guarantees (ICML 2025 · similarity 0.92)
• Attributive Reasoning for Hallucination Diagnosis of Large Language Models (AAAI 2025 · similarity 0.91)
• The Labeling Problem in Hallucination Detection Benchmarks: An Empirical Evaluation (NeurIPS 2026 · similarity 0.91)