Research map / Large language models / Language, safety and interpretability
TRACEDET: HALLUCINATION DETECTION FROM THE DECODING TRACE OF DIFFUSION LARGE LANGUAGE MODELS
Shenxu Chang, Junchi Yu, Weixing Wang, Yongqiang Chen, Jialin Yu, Philip Torr, Jindong Gu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add sub-trace masking (IB-based sub-instance extraction) to the detector (ablation) | TraceDet, LLaDA-8B-Instruct · Average of 3 QA datasets | AUROC (%): 69.1 -> 72.0 | Improved |
| add sub-trace masking (ablation) | TraceDet, Dream-7B-Instruct · Average of 3 QA datasets | AUROC (%): 78.4 -> 80.8 | Improved |
| replace average stepwise entropy with the learned TraceDet detector (ablation) | TraceDet, LLaDA-8B-Instruct · Average of 3 QA datasets | AUROC (%): 62.8 -> 72.0 | Improved |
| replace average stepwise entropy with the learned TraceDet detector (ablation) | TraceDet, Dream-7B-Instruct · Average of 3 QA datasets | AUROC (%): 65.3 -> 80.8 | Improved |
| add masking on TriviaQA with 64 generation steps (ablation) | TraceDet, Dream-7B-Instruct · TriviaQA (64) | AUROC (%): 87.1 -> 86.7 | No clear effect |
| use TraceDet versus strongest baseline EigenScore | TraceDet, LLaDA-8B-Instruct · Average of 3 QA datasets | AUROC (%): 63.2 -> 72.0 | Improved |
| use TraceDet versus strongest baseline EigenScore | TraceDet, Dream-7B-Instruct · Average of 3 QA datasets | AUROC (%): 69.8 -> 80.8 | Improved |
| use TraceDet versus multi-sampling baselines for detection cost | TraceDet, LLaDA-8B-Instruct · Inference time, 100 samples | time (s): 801.35 -> 147.52 | Improved |
| increase generation length to 128 from moderate lengths (ablation) | TraceDet, Dream-7B-Instruct · TriviaQA | AUROC | Hurt |
| vary step length over 1, 2, 4, 8 (ablation) | TraceDet, Dream-7B-Instruct · TriviaQA/HotpotQA/CommonsenseQA | AUROC | No clear effect |
| use random remasking instead of low-confidence/entropy/top-k (ablation) | TraceDet, Dream-7B-Instruct · TriviaQA | AUROC | Hurt |
| vary remasking among low-confidence, entropy and top-k (ablation) | TraceDet, Dream-7B-Instruct · TriviaQA and CommonsenseQA | AUROC: 75-85% AUROC | No clear effect |
Closest papers by meaning
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.92)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.92)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.91)
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.91)
- Entropy Distribution as a Fingerprint for Hallucinations in Generative Models (NeurIPS 2026 · similarity 0.91)
- Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations (NeurIPS 2025 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
