ProbGuard uses early output distributions and Monte Carlo sampling to estimate calibrated unsafe-generation probability, beating 13 baselines on Brier/ECE and limiting jailbreak success to at most 1%.
In Proceedings of the 2025 Annual Meeting of the Association for Computational Linguistics, 31983–32016
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
ProbGuard: Calibrated Safety Risk Estimation from LLM Output Distributions
ProbGuard uses early output distributions and Monte Carlo sampling to estimate calibrated unsafe-generation probability, beating 13 baselines on Brier/ECE and limiting jailbreak success to at most 1%.