Research map / Large language models / Language, safety and interpretability
Harnessing Reasoning Trajectories for Hallucination Detection via Answer-agreement Representation Shaping
Jianxiong Zhang, Bing Guo, Yuming Jiang, Haobo Wang, Bo An, Xuefeng Du
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 |
|---|---|---|---|
| shape embeddings with ARS instead of vanilla LRM embeddings (CCS) | CCS detector · TruthfulQA, Qwen3-8B | AUROC: 66.85 -> 86.64 | Improved |
| shape embeddings with ARS (CCS) | CCS detector · TriviaQA, Qwen3-8B | AUROC: 59.24 -> 88.54 | Improved |
| shape embeddings with ARS (HaloScope) | HaloScope · TruthfulQA, Qwen3-8B | AUROC: 57.78 -> 71.03 | Improved |
| shape embeddings with ARS (EigenScore) | EigenScore · TriviaQA, Qwen3-8B | AUROC: 55.26 -> 56.09 | Improved |
| use latent trace-boundary intervention instead of token deletion (ablation) | ARS (CCS) · TruthfulQA, Qwen3-8B | AUROC: 75.25 -> 86.64 | Improved |
| use latent intervention instead of trace paraphrasing (ablation) | ARS (CCS) · TruthfulQA, Qwen3-8B | AUROC: 50.47 -> 86.64 | Improved |
| use learned ARS shaping instead of PCA projection (ablation) | ARS (CCS) · TruthfulQA, Qwen3-8B | AUROC: 73.26 -> 86.64 | Improved |
| use learned ARS shaping instead of random projection (ablation) | ARS (CCS) · TruthfulQA, Qwen3-8B | AUROC: 68.36 -> 86.64 | Improved |
| perturb only the last token instead of last 1% tokens (ablation) | ARS (CCS) · TruthfulQA, Qwen3-8B | AUROC: 81.70 -> 86.64 | Improved |
| use greedy instead of multinomial sampling (ablation) | ARS (CCS) · TruthfulQA, Qwen3-8B | AUROC: 84.49 -> 86.64 | Improved |
| use embeddings f instead of Attn(f) (ablation) | ARS (CCS) · MATH-500, Qwen3-8B | AUROC: 67.79 -> 78.66 | Improved |
| increase shaping temperature from 0.1 to 0.5 (ablation) | ARS (CCS) · TruthfulQA, Qwen3-8B | AUROC: 86.64 -> 63.97 | Hurt |
Closest papers by meaning
- Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models (AAAI 2026 · similarity 0.93)
- Thinking in Uncertainty: Mitigating Hallucinations in MLRMs with Latent Entropy-Aware Decoding (CVPR 2026 · similarity 0.89)
- Learning to Reason for Hallucination Span Detection (ICLR 2026 · similarity 0.89)
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.89)
- Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.89)
- Token-Guard: Towards Token-Level Hallucination Control via Self-Checking Decoding (ICLR 2026 · similarity 0.88)
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.
