KongMen: Real-Time Dialogue Purification via Training-Free Entity-Redundancy Gating

Abstract

Long-context LLMs (128K–1M tokens) reduce truncation loss, but full-length inputs incur proportional latency, cost, and memory overhead. We propose KongMen, a training-free, LLM-free real-time dialogue purifier for multi-turn conversations. KongMen removes conversation-internal redundancy via query-agnostic sentence-level density scoring (§3.3), a role-asymmetric pipeline, and an Entity Redundancy Gate for referential integrity. Progressive ablation identifies density scoring—not the gate or rules—as the primary mechanism behind ~50% token reduction with no statistically detectable QA degradation under Two One-Sided Tests (TOST) criteria (p = 0.023). Cross-domain evaluation on MeetingBank suggests zero-shot generalization with frozen parameters. In ~3 seconds per 188-turn conversation on CPU, KongMen requires no training, no GPU, and no model loading.