parameter-merging

Tag

Cards List
#parameter-merging

TRACE: Trajectory-Based Safety Patch Learning for LLM Post-Training Realignment

arXiv cs.LG · 2026-07-21 Cached

TRACE proposes a trajectory-based safety patch learning framework for LLM post-training realignment, which learns a plug-in patch that minimally interferes with task-relevant directions while decisively controlling unsafe behaviors, achieving near 100% safety across benchmarks while maintaining utility.

0 favorites 0 likes
← Back to home

Submit Feedback