Sentinel: Decoding Context Utilization via Attention Probing for Efficient LLM Context Compression

Le résumé fourni par la source

Retrieval-augmented generation (RAG) often suffers from long and noisy retrieved contexts. Existing context compression methods typically rely on heuristic relevance estimation or supervised compression models rather than on how LLMs utilize retrieved context during inference. We propose Sentinel, a lightweight sentence-level compression framework that decodes inference-time contextual utilization behaviors from head-wise attention patterns of frozen LLMs. To ground supervision in retrieval-dependent answering behavior, Sentinel trains a lightweight probe using QA examples where the model succeeds only when retrieved context is available. Sentinel performs compression using only a single non-autoregressive forward pass without dedicated compression training or autoregressive scoring. Empirically, we find that effective contextual utilization signals remain accessible even in compact proxy models. On LongBench, Sentinel with a 0.5B proxy model achieves up to 5$\times$ compression while attaining question-answering performance competitive with compression methods built on 7B-scale models. Despite being trained only on English QA data, Sentinel also generalizes effectively to Chinese and out-of-domain settings.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.