New technique grafts latent knowledge to improve LLM agent failure attribution

Researchers have identified a "knows-but-won't-say gap" in large language model (LLM)-based agentic systems, where models possess the internal information to correctly attribute failures but do not explicitly output this information. This paper proposes a technique called "grafting" that leverages this latent knowledge to improve automated failure attribution without needing ground-truth labels during inference. Evaluations show that grafting consistently outperforms direct-prompting methods, highlighting its potential for more reliable failure diagnosis in complex LLM agent systems. AI

IMPACT This research could lead to more reliable debugging and performance improvements in complex LLM agentic systems.

RANK_REASON The cluster contains a research paper detailing a new technique for LLM agent systems. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Agentic Systems
• alphaXiv
• arXiv
• Automated Failure Attribution
• CatalyzeX
• DagsHub
• Gotit.pub
• grafting
• Hugging Face
• LLM-based Agentic Systems
• ScienceCast

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →