Weakly supervised early failure alerting for LLM agents

Arxiv CS.CL (NLP/language models)

Paper introduces weakly supervised method for early failure alerting in dialogs and LLM-agent trajectories, using only trajectory-level success/failure labels. The approach handles sparse supervision by leveraging partial trajectory data.

llm

Arxiv CS.CL (NLP/language models)·Avinash BaidyaXinran LiangRuocheng GuoXiang GaoKamalika Das

6 days ago