A multi-agent debate framework with four long-term memory designs reduces unsafe response rates on HarmBench and CoSafe, with the strongest results when debate is combined with learned guardrails.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates
A multi-agent debate framework with four long-term memory designs reduces unsafe response rates on HarmBench and CoSafe, with the strongest results when debate is combined with learned guardrails.