Research map / Large language models / Language, safety and interpretability

Scalable Token-Level Hallucination Detection in Large Language Models

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Data and hallucination detection · 0 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 TokenHD, a pipeline that synthesizes token-level hallucination labels from multiple critic LLMs (adaptive weighted ensemble) and trains small Qwen3-based detectors with importance-weighted BCE to detect hallucinated tokens in free-form reasoning text.

Problem

Detecting hallucinations in reasoning-intensive LLM outputs at fine granularity; existing process reward models require explicit step segmentation and MCTS-based methods are costly, and token-level hallucination labels are scarce.

Approach

Generate token-level hallucination labels by having multiple critic LLMs flag erroneous spans, restore them to exact text spans, and combine critics with a learned weighted ensemble; train a lightweight token classifier on these soft labels using an importance-weighted BCE loss to counter sparsity of hallucinated tokens.

Key results

• TokenHD-0.6B reaches 63.64 S_incor on Math-500 versus 55.05 for QwQ-32B, and detection scales with size to TokenHD-8B (64.55 on Math-500, 68.45 on Olym-Math).
• TokenHD-8B achieves AUROC 0.8739-0.9004 across four math benchmarks, beating the strongest critic (o4-mini) by about 0.04-0.09.
• TokenHD-8B outperforms Qwen2.5-Math-PRM-72B by 30 to 57 points of S_incor across five benchmarks.

Surprising findings

• A 0.6B detector outperforms the 32B reasoning critic QwQ-32B on hallucination detection F1.
• Math-only detectors transfer zero-shot to code only partially (e.g., 41.30 S_incor on LiveCodeBench-Lite), and Gemini-2.0-Flash outputs are a harder distribution.

Limitations

• Ground-truth evaluation labels depend on labeler models (GPT-5, o3), and the training data is from critic models rather than human labels.
• Generalization to out-of-distribution policy models (e.g., Gemini-2.0-Flash) and code domains is weaker without additional training data.

Experimental setup

Datasets: Math AceReason-Math Big-Math OpenCodeReasoning Math-500 AIME-2024 AIME-2025 OlympiadBench-Math GPQA OlympiadBench-Phy FinQA Code-Elo

Benchmarks: Math-500 AIME-2024 AIME-2025 OlympiadBench-Math GPQA OlympiadBench-Phy FinQA Code-Elo LiveCodeBench-Lite

Metrics: token-level F1 (S_incor) recall on non-hallucinated samples (S_cor) AUROC AUPRC

Baselines: GPT-5 o3 R1-Qwen3-8B QwQ-32B GPT-4.1 o4-mini Qwen2.5-Math-PRM-7B Qwen2.5-Math-PRM-72B

Builds on: process reward models Qwen3 TIES-Merging DARE-Merging

Closest papers by meaning

• LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.95)
• HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.95)
• ANAH-v2: Scaling Analytical Hallucination Annotation of Large Language Models (NeurIPS 2024 · similarity 0.94)
• Robust Hallucination Detection in LLMs via Adaptive Token Selection (NeurIPS 2025 · similarity 0.93)
• Beyond the Global Scores: Fine-Grained Token Grounding as a Robust Detector of LVLM Hallucinations (CVPR 2026 · similarity 0.93)
• Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback (AAAI 2025 · similarity 0.93)