AgentAbstain tests whether agents know when not to act. Across 263 task pairs and 17 frontier models, none exceeds 60% paired accuracy. Project lead Xun Liu of @IllinoisCS will be on hand at Frontier Data Summit, where AgentAbstain is one of 25+ posters.https://t.co/uCDxsTKM8C pic.twitter.com/CokH8YGvba
— Snorkel AI (@SnorkelAI) 6 octobre 2026
AgentAbstain tests whether agents know when not to act. Across 263 task pairs and 17 frontier models, none exceeds 60% paired accuracy. Project lead Xun Liu of @IllinoisCS will be on hand at Frontier Data Summit, where AgentAbstain is one of 25+ posters. https://
frontierdatasummit.ai