Unsupervised Conformal Inference: Bootstrapping and Alignment to Control LLM Uncertainty

2025/09/26 by Ling-Yu Pang, Pang, Lingyou, Lei Huang +11
Computer Science · Engineering · #Statistical and Computational Modeling #Fluid Dynamics and Mixing #Natural Language Processing Techniques

paper · pdf · doi:10.48550/arxiv.2509.23002

Abstract

Deploying black-box LLMs requires managing uncertainty in the absence of token-level probability or true labels. We propose introducing an unsupervised conformal inference framework for generation, which integrates: generative models, incorporating: (i) an LLM-compatible atypical score derived from response-embedding Gram matrix, (ii) UCP combined with a bootstrapping variant (BB-UCP) that aggregates residuals to refine quantile precision while maintaining distribution-free, finite-sample coverage, and (iii) conformal alignment, which calibrates a single strictness parameter τ so a user predicate (e.g., factuality lift) holds on unseen batches with probability ≥ 1-α. Across different benchmark datasets, our gates achieve close-to-nominal coverage and provide tighter, more stable thresholds than split UCP, while consistently reducing the severity of hallucination, outperforming lightweight per-response detectors with similar computational demands. The result is a label-free, API-compatible gate for test-time filtering that turns geometric signals into calibrated, goal-aligned decisions.

Citations

• LLM Web Dynamics: Tracing Model Collapse in a Network of LLMs
• Semantic Volume: Quantifying and Detecting both External and Internal Uncertainty in LLMs
• Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs
• Large language model validity via enhanced conformal prediction methods
• Language Models with Conformal Factuality Guarantees
• Generating with Confidence: Uncertainty Quantification for Black-box Large Language Models
• Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation
• ASQA: Factoid Questions Meet Long-Form Answers
• Self-Consistency Improves Chain of Thought Reasoning in Language Models
• Conformal prediction beyond exchangeability
• A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification
• Adaptive Conformal Inference Under Distribution Shift
• AmbigQA: Answering Ambiguous Open-domain Questions
• Latent Retrieval for Weakly Supervised Open Domain Question Answering
• BERTScore: Evaluating Text Generation with BERT
• HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering
• Least Ambiguous Set-Valued Classifiers With Bounded Error Levels
• Distribution-Free Predictive Inference For Regression
• Risk-Sensitive and Robust Decision-Making: a CVaR Optimization Approach
• Distribution Free Prediction Bands

Related

• Taming Variability: Randomized and Bootstrapped Conformal Risk Control for LLMs
• Conformal Cascade: Distribution-Free Accuracy Guarantees for Multi-Tier LLM Inference
• COLEP: Certifiably Robust Learning-Reasoning Conformal Prediction via Probabilistic Circuits
• The Benefit of Being Bayesian in Online Conformal Prediction
• Another Fit Bites the Dust: Conformal Prediction as a Calibration Standard for Machine Learning in High-Energy Physics