Research map / Large language models / Language, safety and interpretability
Differentiable Conformal Training for LLM Reasoning Factuality
Nathan Hittesdorf, Marco Salzetta, Lu Cheng
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 hand-crafted frequency scoring with learned differentiable conformal scorer (DCF) | Coherent Factuality · MATH (alpha=0.03) | retention: 0.73 -> 1.76 | Improved |
| learned DCF scorer vs CF | Coherent Factuality · MATH (alpha=0.10) | retention: 3.33 -> 5.99 | Improved |
| learned DCF scorer vs CF at low alpha | Coherent Factuality · FELM (alpha=0.01) | retention (%): 11.1 -> 17.9 | Improved |
| learned DCF scorer vs CF at mid-range alpha | Coherent Factuality · FELM (alpha=0.08) | retention (%): 61.3 -> 52.7 | Hurt |
| learned DCF scorer vs CF across alpha | Coherent Factuality · FELM | retention (%): +61.3 at 0.01, -14.0 at 0.08 | Mixed |
| add graph-aware learning to combine features instead of single feature nx_reachability (ablation) | CF · MATH (alpha=0.06) | retention: 2.87 -> 3.56 | Improved |
| combine features vs claim_index only (ablation) | CF · MATH (alpha=0.06) | retention: 1.86 -> 3.56 | Improved |
| combine features vs frequency only (ablation) | CF · FELM (alpha=0.01) | retention (mean claims): 0.51 -> 0.72 | Improved |
| combine features vs frequency only at alpha 0.08 (ablation) | CF · FELM (alpha=0.08) | retention (mean claims): 2.56 -> 2.11 | Hurt |
| add learning without graph structure (Boosted Independent) versus non-learned Independent (ablation) | Independent Factuality · MATH/FELM | retention | Hurt |
| add graph structure to Independent Factuality (CF) (ablation) | Independent Factuality · MATH/FELM | retention | Improved |
| replace differentiable conformal training with independently trained XGBoost scorer (ablation) | DCF · MATH/FELM | retention | Hurt |
Closest papers by meaning
- Language Models with Conformal Factuality Guarantees (ICML 2024 · similarity 0.91)
- Inference-Time Conformal Reasoning with Valid Factuality Control for Large Language Models (ICML 2026 · similarity 0.90)
- Large language model validity via enhanced conformal prediction methods (NeurIPS 2024 · similarity 0.89)
- Conformal Language Model Reasoning with Coherent Factuality (ICLR 2025 · similarity 0.89)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.88)
- A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning (NeurIPS 2025 · 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.
