Research map / Large language models / Language, safety and interpretability
Large language model validity via enhanced conformal prediction methods
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Extends conditional conformal factuality filtering of LLM outputs with level-adaptive thresholds and a differentiable "conditional boosting" procedure that learns better claim scores, improving claim retention while keeping conditional validity.
Problem
Conformal factuality methods for LLM outputs only guarantee correctness marginally over prompts, and imperfect claim scores cause filtering to remove many valid claims, reducing utility of the output.
Approach
Generalize conditional conformal prediction (Gibbs et al.) to monotone losses with prompt-adaptive error levels, and make the conditional cutoff differentiable so claim scores can be boosted via gradient descent to maximize retained claims under conditional coverage constraints.
Key results
- Boosted scores raise mean claim retention on MedLFQA to about 39% versus 24% for unboosted scores at alpha=0.1.
- Level-adaptive conformal with boosted scores retains at least 70% of claims for most prompts, with issued levels ranging roughly 50-85% and empirically calibrated.
- On synthetic heteroskedastic regression, conditional boosting maintains near-perfect conditional coverage where ConfTr undercovers high-variance points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the frequency claim score with learned boosted scores | Conditional conformal filtering · LLM factuality datasets | mean claim retention: 24% vs. 39% | Helped |
| Adapt the conformal level per prompt to target at least 70% claim retention | GPT-3.5-Turbo responses · MedicationQA | claim retention | Helped |
| Relax the claimed correctness level from 90% to 63% using level-adaptive conformal prediction | GPT-3.5-Turbo responses · MedicationQA | claim retention: 90% -> 63% | Helped |
Surprising findings
- Fixed-level filtering at 90% reliability removes nearly all of the response on MedicationQA, while relaxing the level to about 63% preserves almost the entire response.
- Claim scores are only weakly correlated with factuality on Wikipedia biographies, so non-trivial guarantees required allowing up to 3 false claims.
Limitations
- Theory assumes i.i.d. prompt-response tuples; robustness to other distribution shifts is not covered.
- Utility depends on the quality of underlying claim scores; weak scores yield vacuous guarantees.
- Experiments use simplified gradient approximation (optimal basis from D1 only) for computational reasons.
Experimental setup
Datasets: MedLFQA Wikipedia biographies (8516 names) synthetic regression data
Benchmarks: MedLFQA Wikipedia biographies
Metrics: claim retention conditional coverage / calibration of issued levels factuality error rate
Baselines: conformal factuality (Mohri and Hashimoto) marginal boosting / ConfTr fixed-level conditional conformal unboosted frequency score
Builds on: conformal factuality conditional conformal prediction (Gibbs et al.) ConfTr conformal risk control
Closest papers by meaning
- Conformal Language Modeling (ICLR 2024 · similarity 0.93)
- Language Models with Conformal Factuality Guarantees (ICML 2024 · similarity 0.91)
- Geometry-Calibrated Conformal Abstention for Language Models (NeurIPS 2026 · similarity 0.90)
- Conformal Language Modeling via Posterior Sampling (NeurIPS 2026 · similarity 0.90)
- Prune 'n Predict: Optimizing LLM Decision-making with Conformal Prediction (ICML 2025 · similarity 0.90)
- Conformal Information Pursuit for Interactively Guiding Large Language Models (NeurIPS 2025 · similarity 0.89)
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.