Research map / Large language models / Language, safety and interpretability
Building Reliable Long-Form Generation via Hallucination Rejection Sampling
Lin Li, Georgia Channing, Suhaas M Bhat, Gabriel Davis Jones, Yarin Gal
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 |
|---|---|---|---|
| apply SHARS (Ours-Prec) to greedy decoding | Qwen3-32B · FactScore | factual precision (%): 52.4 -> 78.4 | Improved |
| apply SHARS (Ours-Prec) to greedy decoding | Llama3.1-8B · FactScore | factual precision (%): 53.7 -> 78.4 | Improved |
| apply SHARS (Ours-Prec) to greedy decoding | Qwen3-4B · FactScore | factual precision (%): 50.0 -> 74.0 | Improved |
| apply SHARS (Ours-Prec) to greedy decoding | Llama3.1-8B · FactScore | supported claims: 6.7 -> 5.0 | Hurt |
| apply SHARS (Ours-Prec) to greedy decoding | Qwen3-32B · FactScore (200-word constraint) | factual precision (%): 58.0 -> 77.9 | Improved |
| use semantic entropy instead of token-level naive entropy (ablation) | SHARS · FactScore | factual precision (%): 70.1 -> 78.4 | Improved |
| add SHARS (Ours-Prec) to FactAlign | FactAlign Llama3-8B-Instruct · FactScore (no length constraint) | factual precision (%): 53.1 -> 80.6 | Improved |
| add SHARS (Ours-Prec) to FactAlign | FactAlign Llama3-8B-Instruct · FactScore (200-word constraint) | factual precision (%): 55.4 -> 79.1 | Improved |
| add SHARS to greedy decoding | Qwen3-32B · LongFact (no length constraint) | factual precision (%): 93.0 -> 94.6 | Improved |
| add rewriting to Following sampling (ablation) | SHARS · FactScore ablation | response rate (%): 54.4 -> 91.8 | Improved |
| add rewriting to Following sampling (ablation) | SHARS · FactScore ablation | factual precision (%): 73.5 -> 69.4 | Hurt |
| use Following sampling instead of Temperature sampling (with rewriting) (ablation) | SHARS · FactScore ablation | factual precision (%): 64.8 -> 69.4 | Improved |
Closest papers by meaning
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.91)
- Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling (NeurIPS 2025 · similarity 0.91)
- Zero-resource Hallucination Detection for Text Generation via Graph-based Contextual Knowledge Triples Modeling (AAAI 2025 · similarity 0.90)
- ANAH-v2: Scaling Analytical Hallucination Annotation of Large Language Models (NeurIPS 2024 · similarity 0.90)
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.90)
- Token-Guard: Towards Token-Level Hallucination Control via Self-Checking Decoding (ICLR 2026 · similarity 0.90)
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.
