Between the Layers Lies the Truth: Uncertainty Estimation in LLMs Using Intra-Layer Local Information Scores

2026-03-17

AI Summary

Large language models often exhibit overconfidence even when incorrect, highlighting the urgent need for efficient and transferable uncertainty estimation methods. This work proposes a lightweight, training-free approach that derives instance-level uncertainty scores by analyzing cross-layer local consistency among internal representations through a single forward pass. The method is computationally efficient, model-agnostic, and readily applicable across diverse architectures and settings. Experimental results demonstrate that it matches or surpasses existing detection techniques on in-distribution data and significantly outperforms baselines under challenging conditions—such as out-of-distribution generalization and 4-bit quantization—with improvements of up to 2.86 points in AUPRC and 21.02 points in Brier score. Additionally, the analysis reveals notable differences across models in how uncertainty is encoded across layers.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Web Mining and Content Analysis: Large pretrained models with web data
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Large language models (LLMs) are often confidently wrong, making reliable uncertainty estimation (UE) essential. Output-based heuristics are cheap but brittle, while probing internal representations is effective yet high-dimensional and hard to transfer. We propose a compact, per-instance UE method that scores cross-layer agreement patterns in internal representations using a single forward pass. Across three models, our method matches probing in-distribution, with mean diagonal differences of at most $-1.8$ AUPRC percentage points and $+4.9$ Brier score points. Under cross-dataset transfer, it consistently outperforms probing, achieving off-diagonal gains up to $+2.86$ AUPRC and $+21.02$ Brier points. Under 4-bit weight-only quantization, it remains robust, improving over probing by $+1.94$ AUPRC points and $+5.33$ Brier points on average. Beyond performance, examining specific layer--layer interactions reveals differences in how disparate models encode uncertainty. Altogether, our UE method offers a lightweight, compact means to capture transferable uncertainty in LLMs.

Problem

• uncertainty estimation
• large language models
• internal representations
• transferability
• model confidence

Innovation

• uncertainty estimation
• large language models
• intra-layer information
• cross-layer agreement
• model robustness