RSS feedarxiv.org
LongGuard Analyzes and Mitigates Long-Context Safety Guardrail FailuresLongGuard evaluates safety guardrails from 0.25k to 32k tokens, identifies attention dilution as the core failure mechanism, and introduces training-free mitigations with context-aware routing.