Mitigating Factual Hallucination in Large Reasoning Models via Mixed-Mode Advantage Regularization

2026-07-07

AI Summary

This work addresses “reasoning-induced hallucination” in large reasoning models—where explicit chain-of-thought (CoT) reasoning overrides an initially correct direct answer, leading to factual errors. To mitigate this, the authors propose the MARGO framework, which treats explicit reasoning as a residual correction to the model’s default answering tendency. MARGO constructs an intra-model reference by fusing CoT and non-CoT response trajectories and introduces a mixed-mode advantage estimator in reinforcement learning to assess whether reasoning yields factual improvement. This approach provides the first formal characterization and targeted mitigation of reasoning-induced hallucination, significantly enhancing answer reliability on multiple factual question-answering benchmarks while preserving or even improving general reasoning capabilities on mathematical tasks.

Abstract

Large reasoning models (LRMs) improve language model capabilities by generating explicit thinking traces before final answers. In factuality-oriented question answering (QA), such thinking often improves overall performance by helping the model recover relevant knowledge and refine its answers. However, we find that this benefit is not uniform at the instance level: explicit thinking can also overturn correct non-thinking answers and lead to factual drift. We refer to this failure mode as \emph{thinking-induced hallucination}. To explain this phenomenon, we formulate explicit thinking in factuality QA as a thinking residual over the model's direct-answer tendency, which can either recover missing knowledge or introduce unsupported associations. Based on this formulation, we propose MARGO, \underline{\textit{M}}ixed-Mode \underline{\textit{A}}dvantage \underline{\textit{R}}egularization for \underline{\textit{G}}rounded \underline{\textit{O}}ptimization, a reinforcement learning framework that uses non-thinking rollouts as same-model references in advantage estimation. By constructing mixed-mode rollout groups with both thinking and non-thinking trajectories, MARGO evaluates whether explicit thinking adds factual value beyond direct answering, thereby suppressing hallucination-prone thinking while preserving beneficial thinking behaviors. Experiments across multiple factuality-oriented QA benchmarks demonstrate that MARGO improves factual reliability over strong baselines, while evaluations on mathematical benchmarks show that it preserves general reasoning ability.

Problem

• factual hallucination
• large reasoning models
• thinking-induced hallucination
• factuality-oriented QA
• explicit thinking

Innovation

• thinking-induced hallucination
• mixed-mode advantage regularization
• factual reliability
• reasoning residual
• grounded optimization