Research map / Large language models / Language, safety and interpretability
Factuality Enhanced Language Models for Open-Ended Text Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| scale LM from 126M to 530B | top-p 0.9 LM · FactualityPrompts factual | NE error (%): 63.69 -> 33.30 | Helped |
| switch top-p 0.9 to greedy decoding | 1.3B LM · FactualityPrompts factual | NE error (%): 52.4 -> 39.9 | Helped |
| switch top-p 0.9 to greedy decoding | 1.3B LM · FactualityPrompts factual | diversity: 0.88 -> 0.05 | Hurt |
| use factual-nucleus sampling instead of top-p 0.9 | 1.3B LM · FactualityPrompts factual | NE error (%): 52.4 -> 42.1 | Helped |
| use factual-nucleus sampling instead of top-p 0.9 | 1.3B LM · FactualityPrompts factual | diversity: 0.88 -> 0.55 | Hurt |
| add omega-bound to lambda-decay + p-reset (ablation) | factual-nucleus · FactualityPrompts factual | diversity: 0.52 -> 0.55 | Helped |
| add p-reset to lambda-decay (ablation) | factual-nucleus 0.9|0.9 · FactualityPrompts factual | repetition (%): 30.7 -> 10.3 | Helped |
| continued Wikipedia training without TopicPrefix (ablation) | 1.3B LM · FactualityPrompts factual | NE error (%): 42.1 -> 42.7 | No clear effect |
| add TopicPrefix to continued training (ablation) | 1.3B LM · FactualityPrompts factual | NE error (%): 42.1 -> 27.6 | Helped |
| add half-pivot sentence completion loss to TopicPrefix (ablation) | 1.3B LM · FactualityPrompts factual | entailment ratio (%): 8.7 -> 17.4 | Helped |
| add half-pivot sentence completion loss to TopicPrefix (ablation) | 1.3B LM · FactualityPrompts factual | NE error (%): 27.6 -> 23.6 | Helped |
| use ROOT pivot instead of half-sentence pivot (ablation) | 1.3B LM · FactualityPrompts factual | NE error (%): 24.7 vs 23.6 | No clear effect |
Closest papers by meaning
- Fine-Tuning Language Models for Factuality (ICLR 2024 · similarity 0.89)
- Is Factuality Enhancement a Free Lunch For LLMs? Better Factuality Can Lead to Worse Context-Faithfulness (ICLR 2025 · similarity 0.89)
- Precise Information Control in Long-Form Text Generation (NeurIPS 2025 · similarity 0.89)
- Long-form factuality in large language models (NeurIPS 2024 · similarity 0.89)
- Attention Satisfies: A Constraint-Satisfaction Lens on Factual Errors of Language Models (ICLR 2024 · similarity 0.87)
- Scalable Influence and Fact Tracing for Large Language Model Pretraining (ICLR 2025 · similarity 0.87)
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.