作者位次 #7/8SIGIR 2026Oral· Information Retrieval (Industry)
WeSEAL: Well-calibrated Search for Eliminating Attention-sink Leakage
Juyuan Wang · Chenxing Wang · Aolin Li · Huiyun Hu · Yuchen Fang · Haijun Wu · Jin Xu · Dongliang Liao
摘要
Large Language Models (LLMs) have revolutionized relevance ranking, yet their deployment in high-concurrency search systems is stifled by a critical pathology: ''Attention Sink Drift.'' In standard causal attention, models disproportionately allocate weight to the initial document regardless of r...