Cross-Session Decomposition Attacks: Scaling Risk and Intent-Aligned Retrieval Defense
IntentAlign-MiniLM, the authors' 22M-parameter intent-aligned retriever, outperforms much larger embedding models on held-out intent retrieval and yields the best learned-retriever harmful recall across tested guardrails.
Disen Liao, Yihan Wang, Freda Shi et al.
· 0 citations