Uncertainty-Aware Attention Heads: Efficient Unsupervised Uncertainty Quantification for LLMs

2025-05-26

AI Summary

Existing uncertainty quantification (UQ) methods for large language models (LLMs) suffer from high computational overhead or reliance on supervised signals, limiting their practicality in mitigating hallucination. Method: We propose RAUQ—a fully unsupervised, single-forward, sequence-level UQ method. RAUQ is the first to empirically identify and exploit a spontaneous decay pattern in “uncertainty-aware” attention heads of Transformers, wherein these heads progressively diminish attention to preceding tokens. It derives real-time, white-box uncertainty scores via attention-weight analysis, recursive aggregation, and token-level confidence modeling—requiring no labels, fine-tuning, or auxiliary training. Results: RAUQ achieves state-of-the-art performance across 12 diverse tasks on four mainstream LLMs, with computational overhead under 1% of inference latency. It exhibits strong generalization and zero-label dependency, enabling plug-and-play deployment.

Technology Category

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

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• 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

Abstract

Large language models (LLMs) exhibit impressive fluency, but often produce critical errors known as"hallucinations". Uncertainty quantification (UQ) methods are a promising tool for coping with this fundamental shortcoming. Yet, existing UQ methods face challenges such as high computational overhead or reliance on supervised learning. Here, we aim to bridge this gap. In particular, we propose RAUQ (Recurrent Attention-based Uncertainty Quantification), an unsupervised approach that leverages intrinsic attention patterns in transformers to detect hallucinations efficiently. By analyzing attention weights, we identified a peculiar pattern: drops in attention to preceding tokens are systematically observed during incorrect generations for certain"uncertainty-aware"heads. RAUQ automatically selects such heads, recurrently aggregates their attention weights and token-level confidences, and computes sequence-level uncertainty scores in a single forward pass. Experiments across 4 LLMs and 12 question answering, summarization, and translation tasks demonstrate that RAUQ yields excellent results, outperforming state-of-the-art UQ methods using minimal computational overhead (<1% latency). Moreover, it requires no task-specific labels and no careful hyperparameter tuning, offering plug-and-play real-time hallucination detection in white-box LLMs.

Problem

• Unsupervised uncertainty quantification for LLMs
• Detect hallucinations using attention patterns
• Minimal computational overhead in real-time

Innovation

• Unsupervised uncertainty quantification via attention patterns
• Recurrent aggregation of attention weights for efficiency
• Plug-and-play real-time hallucination detection