Learning Conformal Abstention Policies for Adaptive Risk Management in Large Language and Vision-Language Models
Sina Tayebati, Divake Kumar, Nastaran Darabi, Dinithi Jayasuriya, Ranganath Krishnan, Amit Ranjan Trivedi
TL;DR
The paper addresses the rigidity of static conformal prediction thresholds in risk-sensitive LLM/VLM applications. It introduces Learnable Conformal Abstention (CAP), which couples reinforcement learning with conformal prediction to adapt two thresholds and produce single-label, set-valued, or abstention outputs, all while maintaining conformal coverage guarantees. Across ten MCQA benchmarks and diverse model families, CAP yields higher accuracy, better hallucination-detection reliability (AUROC), improved uncertainty-guided generation (AUARC), and significantly reduced calibration error (ECE), while preserving at least 90% coverage and producing more informative prediction sets. This approach offers a principled, scalable framework for robust decision-making in safety-critical foundation models, with code and results enabling reproducibility and further exploration in risk-management workflows.
Abstract
Large Language and Vision-Language Models (LLMs/VLMs) are increasingly used in safety-critical applications, yet their opaque decision-making complicates risk assessment and reliability. Uncertainty quantification (UQ) helps assess prediction confidence and enables abstention when uncertainty is high. Conformal prediction (CP), a leading UQ method, provides statistical guarantees but relies on static thresholds, which fail to adapt to task complexity and evolving data distributions, leading to suboptimal trade-offs in accuracy, coverage, and informativeness. To address this, we propose learnable conformal abstention, integrating reinforcement learning (RL) with CP to optimize abstention thresholds dynamically. By treating CP thresholds as adaptive actions, our approach balances multiple objectives, minimizing prediction set size while maintaining reliable coverage. Extensive evaluations across diverse LLM/VLM benchmarks show our method outperforms Least Ambiguous Classifiers (LAC) and Adaptive Prediction Sets (APS), improving accuracy by up to 3.2%, boosting AUROC for hallucination detection by 22.19%, enhancing uncertainty-guided selective generation (AUARC) by 21.17%, and reducing calibration error by 70%-85%. These improvements hold across multiple models and datasets while consistently meeting the 90% coverage target, establishing our approach as a more effective and flexible solution for reliable decision-making in safety-critical applications. The code is available at: {https://github.com/sinatayebati/vlm-uncertainty}.