Research map / Large language models / Language, safety and interpretability
Conditional Factuality Controlled LLMs with Generalization Certificates via Conformal Sampling
Kai Ye, Qingtao Pan, Shuo Li
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 |
|---|---|---|---|
| replace global ICP threshold with feature-conditional CFC threshold | conformal factuality · Synthetic data (alpha=0.10) | GSC (%): 57.4 -> 88.7 | Improved |
| add exact conformal correction on top of a learned conditional threshold (CFC vs Learnt CP) (ablation) | conditional threshold · Synthetic data (alpha=0.10) | GSC (%): 84.3 -> 88.7 | Improved |
| replace ICP with CFC | Llama-3-8B-Instruct · GSM8K (alpha=0.05) | APSS: 4.73 -> 2.35 | Improved |
| replace ICP with CFC | Llama-3-8B-Instruct · TriviaQA (alpha=0.30) | GSC (%): 49.3 -> 62.8 | Improved |
| add PAC stability adjustment (CFC-P-F instead of CFC) (ablation) | CFC · TriviaQA (alpha=0.30) | ECR toward target 70%: 68.4 -> 70.0 | Improved |
| add PAC stability adjustment (CFC-P-F instead of CFC) (ablation) | CFC · GSM8K (alpha=0.05) | APSS: 2.35 -> 4.59 | Hurt |
| increase sample budget from 5 to 20 candidates (ablation) | CFC on GSM8K · GSM8K (alpha=0.10) | GSC (%): 86.36 -> 87.73 | No clear effect |
| increase sample budget from 5 to 20 candidates (ablation) | CFC on GSM8K · GSM8K (alpha=0.10) | APSS: 1.49 -> 3.79 | Hurt |
| use chosen entropy-and-loss feature map instead of entropy-linear (ablation) | CFC on TriviaQA · TriviaQA (alpha=0.30) | GSC (%): 45.1 -> 62.8 | Improved |
| use chosen feature map instead of max-loss-linear (ablation) | CFC on TriviaQA · TriviaQA (alpha=0.30) | GSC (%): 56.5 -> 62.8 | Improved |
| apply CFC-P-F to a vision-language model (Qwen2-VL) instead of ICP | Flickr8k captioning · Flickr8k VLM (alpha=0.03) | GSC (%): 85.21 -> 95.21 | Improved |
| use plain CFC instead of CFC-P-F on easy Flickr8k (ablation) | CFC · Flickr8k VLM (alpha=0.03) | ECR toward target 97%: 97.27 -> 95.81 | Hurt |
Closest papers by meaning
- Large language model validity via enhanced conformal prediction methods (NeurIPS 2024 · similarity 0.89)
- Language Models with Conformal Factuality Guarantees (ICML 2024 · similarity 0.89)
- Conformal Language Modeling via Posterior Sampling (NeurIPS 2026 · similarity 0.89)
- Inference-Time Conformal Reasoning with Valid Factuality Control for Large Language Models (ICML 2026 · similarity 0.89)
- FactTest: Factuality Testing in Large Language Models with Finite-Sample and Distribution-Free Guarantees (ICML 2025 · similarity 0.88)
- FLAME : Factuality-Aware Alignment for Large Language Models (NeurIPS 2024 · 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.
