On the Effect of Uncertainty on Layer-wise Inference Dynamics

2025-07-09

AI Summary

This study investigates how large language models (LLMs) represent and process predictive uncertainty across layers to improve uncertainty detection and hallucination mitigation. Addressing the gap in prior work—which fails to characterize how uncertainty dynamically shapes hidden state evolution—we employ the Tuned Lens method to track token-level layerwise probability trajectories across five mainstream LLMs and eleven benchmark datasets, using erroneous predictions as proxies for epistemic uncertainty. Key findings are: (1) Confidence trajectories for correct and incorrect predictions exhibit striking similarity across layers—including synchronized sharp increases at comparable depths—indicating that standard hidden state dynamics are largely insensitive to uncertainty; (2) High-performing models adopt distinct uncertainty-processing strategies. These results challenge the prevailing paradigm of detecting uncertainty directly from static hidden states, revealing that effective uncertainty modeling must focus on dynamic mechanisms rather than instantaneous state snapshots.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations
• Natural Language Processing: (Large) Language Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Understanding how large language models (LLMs) internally represent and process their predictions is central to detecting uncertainty and preventing hallucinations. While several studies have shown that models encode uncertainty in their hidden states, it is underexplored how this affects the way they process such hidden states. In this work, we demonstrate that the dynamics of output token probabilities across layers for certain and uncertain outputs are largely aligned, revealing that uncertainty does not seem to affect inference dynamics. Specifically, we use the Tuned Lens, a variant of the Logit Lens, to analyze the layer-wise probability trajectories of final prediction tokens across 11 datasets and 5 models. Using incorrect predictions as those with higher epistemic uncertainty, our results show aligned trajectories for certain and uncertain predictions that both observe abrupt increases in confidence at similar layers. We balance this finding by showing evidence that more competent models may learn to process uncertainty differently. Our findings challenge the feasibility of leveraging simplistic methods for detecting uncertainty at inference. More broadly, our work demonstrates how interpretability methods may be used to investigate the way uncertainty affects inference.

Problem

• Analyze how uncertainty affects layer-wise inference dynamics in LLMs
• Compare probability trajectories for certain and uncertain predictions
• Evaluate feasibility of simple uncertainty detection methods

Innovation

• Uses Tuned Lens for layer-wise probability analysis
• Compares certain and uncertain prediction trajectories
• Challenges simplistic uncertainty detection methods