Safety-aligned LLMs often refuse harmful single-turn requests but comply when the same goal is spread across multiple turns. The authors introduce TRACE, a token-level training objective that uses refusal-attributable advantages and contrastive erasure to control risk on unseen multi-turn trajectories. Across five open-weight models and seven multi-turn attacks, TRACE achieves the lowest attack success rate in all 35 model-attack pairs, while utility on MMLU and HellaSwag drops by at most 1.23 points. This matters because it offers a principled way to harden models against multi-turn attacks without significant performance loss.
TRACE reduces multi-turn attack success while largely preserving model utility
A token-level safety objective outperformed alternatives across 35 model and attack combinations, with benchmark utility falling by no more than 1.23 points.
PRADA Lab · King Abdullah University of Science and Technology · State Key Laboratory of Internet of Things for Smart City · University of Macau · RIKEN Center for Advanced Intelligence Project
In one line
TRACE reduces multi-turn attack success rates across five models and seven attacks, with utility drop at most 1.23 points.
Before you read this
Three ideas this paper builds onThe authors first develop a theoretical bound describing when suppression of harmful behavior in supervised, single-turn contexts can reduce risk over multi-turn trajectories.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.