How Context Attribution Handles What the Model Already Knows

2026-07-26

AI Summary

Existing context attribution methods struggle to disentangle the contributions of in-weight knowledge—encoded in model parameters—from those of the input context, leading to unreliable attributions. This work presents the first systematic evaluation of such methods under knowledge overlap scenarios, where both sources contain similar information. To address this gap, we introduce four novel evaluation metrics—BCS, CAC, APS, and SSP—and a benchmark dataset, WMDP-Cyber++, annotated with ground-truth source labels. Empirical results demonstrate that prevailing attribution methods produce unfaithful attributions in these settings, failing to accurately distinguish between in-weight and in-context knowledge sources. Our findings reveal fundamental limitations of current approaches and establish a rigorous foundation for future research on reliable knowledge attribution in large language models.

Abstract

Context attribution methods for large language models (LLMs) identify which input context contributes to the model response. Recent works show the initial success in attributing the con- tributive score of the contexts. However, we observe that when the context overlaps with the training data, these methods can- not disentangle in-context from in-weight (IW) contributions, producing unreliable scores. Based on this observation, in this work, we introduce: 1) an evaluation protocol that relies on four new metrics (base-model context attribution score (BCS), cross-model context attribution consistency (CAC), attribution preservation score (APS), source separation pre- cision (SSP)) and 2) a benchmark dataset (WMDP-Cyber++) with ground-truth provenance labels to systematically assess attribution under IW overlap. In our experiments across four well-known context attribution methods, we demonstrate that they provide unfaithful attribution when the knowledge from the context also exists in the weights. Finally, we adapt these methods for source separation (IW vs. in-context learning (ICL)) and show that they cannot do the disentanglement based on the contributive score

Problem

• context attribution
• in-weight knowledge
• in-context learning
• attribution reliability
• knowledge overlap

Innovation

• context attribution
• in-weight knowledge
• source separation
• evaluation metrics
• benchmark dataset