Correctness-Optimized Residual Activation Lens (CORAL): Transferrable and Calibration-Aware Inference-Time Steering

2026-02-05

AI Summary

This work addresses the prevalent issue of miscalibration in large language models following instruction tuning and preference alignment, which undermines their reasoning reliability. The authors propose a lightweight, inference-time intervention method that directly optimizes output correctness without requiring retraining. Specifically, they introduce a weight-decay-regularized MLP probe to extract distributed correctness signals from residual activations and apply a regularization mechanism to modulate model behavior. This approach enables effective cross-task transferability. Experimental results demonstrate that the method improves average accuracy by 10% and reduces Expected Calibration Error (ECE) by 50% across multiple 7B-scale models. Furthermore, on four unseen benchmark datasets, it achieves an average accuracy gain of 14% and a 49% reduction in ECE, highlighting its robustness and generalization capability.

Technology Category

Machine Learning: Calibration & Uncertainty QuantificationNatural Language Processing: Safety and RobustnessComputer Vision: Large Vision Models

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingUser Modeling, Personalization and Recommendation: Fairness-aware retrieval and rankingSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Large language models (LLMs) exhibit persistent miscalibration, especially after instruction tuning and preference alignment. Modified training objectives can improve calibration, but retraining is expensive. Inference-time steering offers a lightweight alternative, yet most existing methods optimize proxies for correctness rather than correctness itself. We introduce CORAL (Correctness-Optimized Residual Activation Lens), a regularized inference-time steering method that captures distributed correctness signals from model internal activations using weight-decay MLP probes. We evaluate CORAL across three 7B-parameter models and find that it consistently improves accuracy by 10\% and expected calibration error (ECE) by 50\% on average. We additionally demonstrate that these gains transfer without retraining to the complete published test sets of four held-out benchmarks (ARC-Challenge, HellaSwag, Math-MC, OpenBookQA), averaging 14\% accuracy improvements and 49\% ECE improvements. Our results support the hypothesis that distributed information in model internals can be extracted using regularized probes when individual neurons are insufficient. CORAL thus provides a compute-efficient, transferable, and calibration-aware approach to improve MCQA performance during inference.

Problem

• miscalibration
• large language models
• inference-time steering
• multiple-choice question answering
• expected calibration error

Innovation

• inference-time steering
• model calibration
• correctness optimization
• transferable probing
• residual activation lens