TraceSafe-Bench reveals that LLM guardrail performance on tool-use trajectories depends more on structural data handling than semantic safety alignment, with general models outperforming specialized ones and accuracy improving over longer trajectories.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
TraceSafe: A Systematic Assessment of LLM Guardrails on Multi-Step Tool-Calling Trajectories
TraceSafe-Bench reveals that LLM guardrail performance on tool-use trajectories depends more on structural data handling than semantic safety alignment, with general models outperforming specialized ones and accuracy improving over longer trajectories.