一作ACL 2026Poster· Natural Language Processing
Mitigating Structural Knowledge Collapse in Domain-Specific LLM s via Morpheme-Aware KV -Aggregation
Yuxuan Si · Zheqi Lv · Chengxi Zang · Zhengyu Chen · Fei Wu
摘要
Standard tokenizers over-fragment domain terms, disrupting morpheme semantics. We characterize this representational misalignment as Structural Knowledge Collapse (SKC), where attention mechanisms fail to reconstruct coherent concepts from fragmented inputs. While existing input-centric solutions...