Context Dependence and Reliability in Autoregressive Language Models
Poushali Sengupta, Shashi Raj Pandey, Sabita Maharjan, Frank Eliassen
TL;DR
This work tackles the challenge of explaining autoregressive LLM behavior in the presence of redundant context by introducing RISE, a dependence-aware scoring method. RISE quantifies the unique information each structured context unit provides about the next-token distribution, using conditional mutual information and normalization to suppress redundancy. The authors formalize theoretical properties, analyze computational aspects, and present an integrated lightweight selector alongside an optional learned module to retain only informative anchors. Empirical results on open-weight models demonstrate that RISE reduces redundancy-driven attribution inflation, improves interpretability, and yields actionable monitoring signals for prompt auditing, retrieval validation, and safety in practical deployments. Overall, RISE advances trustworthy AI explanations by prioritizing true dependence over frequency or overlap in long, structured contexts.
Abstract
Large language models (LLMs) generate outputs by utilizing extensive context, which often includes redundant information from prompts, retrieved passages, and interaction history. In critical applications, it is vital to identify which context elements actually influence the output, as standard explanation methods struggle with redundancy and overlapping context. Minor changes in input can lead to unpredictable shifts in attribution scores, undermining interpretability and raising concerns about risks like prompt injection. This work addresses the challenge of distinguishing essential context elements from correlated ones. We introduce RISE (Redundancy-Insensitive Scoring of Explanation), a method that quantifies the unique influence of each input relative to others, minimizing the impact of redundancies and providing clearer, stable attributions. Experiments demonstrate that RISE offers more robust explanations than traditional methods, emphasizing the importance of conditional information for trustworthy LLM explanations and monitoring.