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

2026-07-08 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, short

Summary

A new paper, arXiv:2607.05861, submitted on 7 Jul 2026, introduces MARGO (Mixed-Mode Advantage Regularization for Grounded Optimization), a reinforcement learning framework designed to mitigate "thinking-induced hallucination" in Large Reasoning Models (LRMs). LRMs typically enhance capabilities by generating explicit thinking traces before providing final answers, often improving performance in factuality-oriented question answering. However, this research identifies instances where explicit thinking can paradoxically overturn correct direct answers, causing factual drift. MARGO addresses this by formulating explicit thinking as a residual over direct-answer tendencies. It constructs mixed-mode rollout groups, comparing thinking and non-thinking trajectories to evaluate the factual value added by explicit thought. This approach suppresses hallucination-prone thinking while retaining beneficial reasoning. Experiments on multiple factuality-oriented QA benchmarks demonstrate MARGO's improved factual reliability over strong baselines, while also preserving general reasoning ability on mathematical benchmarks.

Key takeaway

For Machine Learning Engineers deploying Large Reasoning Models in factuality-oriented question answering, you should consider MARGO to enhance factual reliability. This reinforcement learning framework directly addresses "thinking-induced hallucination" by selectively suppressing detrimental reasoning paths while preserving beneficial ones. Implementing MARGO can significantly reduce factual drift, ensuring your models provide more accurate answers without sacrificing general reasoning capabilities on complex tasks.

Key insights

Thinking traces in LRMs can induce factual hallucinations; MARGO uses mixed-mode regularization to suppress these while preserving beneficial reasoning.

Principles

• Explicit thinking can cause factual drift.
• Compare thinking vs. non-thinking outputs.
• Reinforcement learning can refine reasoning.

Method

MARGO is an RL framework using non-thinking rollouts as same-model references for advantage estimation. It constructs mixed-mode rollout groups to evaluate if explicit thinking adds factual value, suppressing hallucination-prone thoughts.

In practice

• Improve LRM factual reliability.
• Preserve LRM general reasoning.
• Reduce "thinking-induced hallucination".

Topics

• Large Reasoning Models
• Factual Hallucination
• Reinforcement Learning
• Question Answering
• MARGO Framework
• Thinking Traces

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination — Enhanced LLM reasoning, especially via RL, causally increases tool hallucination, revealing a fundamental reliability-capability trade-off.
• Shameless Guesses, Not Hallucinations — AI "hallucinations" are better understood as "shameless guesses" stemming from their reward-driven training process.
• TIGER: Traceable Inference with Graph-Based Evidence Routing for Mitigating Hallucinations in Multimodal Generation — TIGER mitigates multimodal generation hallucinations via graph-based evidence routing for localized, fact-level repair.
• Stop When Further Reasoning Won't Help: Attention-State Adaptive Generation in Reasoning Models — Adaptive generation based on attention-state analysis prevents large reasoning models from overthinking.
• Mitigating Hallucinations in Large Language Models Via Decoder Layer Skipping — DeLask mitigates LLM hallucinations by dynamically skipping deeper decoder layers identified via gradient driftance.
• Detecting Hallucinations for Large Language Model-based Knowledge Graph Reasoning — LUCID detects LLM hallucinations in KG reasoning by fusing LLM attention, KG semantics, and structural information via a GNN.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.