Search papers, labs, and topics across Lattice.
This paper introduces Decoupled Update Dynamics (DUD), a novel framework for improving uncertainty quantification in Large Language Models (LLMs) by separating the contributions of Feed-Forward Networks and Attention mechanisms. By employing noise-induced causal interventions, the authors reveal how memory-context misalignments serve as critical indicators of uncertainty, which traditional methods often overlook. Experimental results show that DUD not only enhances uncertainty estimation and calibration but also achieves better generalization across datasets compared to existing approaches.
Decoupling memory and context in LLMs reveals hidden fragilities that traditional uncertainty metrics miss, leading to significantly improved reliability in uncertainty quantification.
Accurate Uncertainty Quantification (UQ) is critical for reliable deployment of Large Language Models (LLMs), yet traditional probability-based metrics often fail to capture the model's true epistemic state. While recent mechanistic approaches leverage hidden state dynamics, they typically aggregate residual stream updates, conflating the distinct roles of parametric memory (Feed-Forward Networks) and contextual processing (Attention). We argue that this aggregation obscures fine-grained mechanistic conflicts, such as memory-context misalignment, that are fundamental indicators of uncertainty. To address this, we introduce \textbf{D}ecoupled \textbf{U}pdate \textbf{D}ynamics \textbf{(DUD)}, a framework that explicitly decouples FFN and Attention contributions via noise-induced causal interventions. By quantifying the independent restoration capabilities of each module, we construct a dual-stream dynamic profile that captures the model's internal fragility. Extensive experiments demonstrate that DUD significantly outperforms state-of-the-art baselines in both uncertainty estimation and calibration, while exhibiting superior cross-dataset generalization, validating decoupled dynamics as a robust proxy for model faithfulness.
