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

Learning to Reason for Hallucination Span Detection

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 1 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.

Trains a 7B/14B LLM with GRPO-style RL and a span-F1 reward (RL4HS) to reason and output hallucinated spans on RAGTruth, adding a class-aware advantage scaling (CAPO) to fix non-hallucination reward imbalance.

Problem

Hallucination span detection (locating unsupported spans in LLM outputs) is treated as a multi-step decision problem, but prior work either uses binary detection or non-reasoning generative/token-level models; it is unclear whether learned reasoning helps span-level detection.

Approach

Apply RL with a verifiable span-F1 reward to a reasoning LLM for hallucination span detection, and introduce class-aware policy optimization that scales advantages of non-hallucination samples by alpha (0.5) to counter GRPO's reward imbalance that over-incentivizes empty predictions.

Key results

• RL4HS-7B reaches avg span-F1 55.9 vs 50.1 for SFT-7B on RAGTruth (summarization, QA, data-to-text).
• RL4HS-14B achieves avg span-F1 58.3, outperforming GPT-5, o3 (avg 51.2) and Qwen3-14B (33.7) on RAGTruth.
• CAPO improves avg F1 from 54.2 (GRPO) to 55.9 with higher recall (51.2 vs 47.3).

Surprising findings

• Reasoning models such as QwQ-32B and Qwen3 underperform much smaller RL-trained 7B models on span detection despite their size.
• Simply rescaling the non-hallucination reward (Dr.GRPO variant) does not fix the imbalance, since GRPO standardization cancels such scaling.

Limitations

• Evaluated only on RAGTruth; generalization limited to its three CNLG tasks.
• CAPO scaling factor alpha=0.5 chosen on a validation set; sensitivity not broadly analyzed.
• Proprietary model results are taken from prompting with limited decoding settings.

Experimental setup

Datasets: RAGTruth

Benchmarks: RAGTruth

Metrics: span-F1 span precision span recall

Baselines: SFT RL4HS-GRPO Multi-View Attention Qwen3-8B Qwen3-14B QwQ-32B GPT-5 o3 GPT-5-mini GPT-4o-mini

Builds on: GRPO Dr.GRPO PPO Chain-of-Thought prompting

Closest papers by meaning

• BALTO: Balanced Token-Level Policy Optimization for Hallucination Mitigation (NeurIPS 2026 · similarity 0.92)
• Reasoning Models Hallucinate More: Factuality-Aware Reinforcement Learning for Large Reasoning Models (NeurIPS 2025 · similarity 0.92)
• Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models (CVPR 2026 · similarity 0.91)
• Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.91)
• Attributive Reasoning for Hallucination Diagnosis of Large Language Models (AAAI 2025 · similarity 0.91)
• Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models (AAAI 2026 · similarity 0.90)