BLADE: Auditing Information Use for Failure Attribution in LLM Multi-Agent Systems
BLADE tries to find the agent and step that actually misused information, not just trace where the information came from.
The paper presents a training-free failure attribution framework for LLM multi-agent systems. It audits source-to-decision handoffs, checking whether a receiving agent used prior information faithfully and within task requirements. The method separates error introduction, propagation, and repair before making a global diagnosis across the run. On 1,228 failed runs, it reports best or tied-best step accuracy among eight strategies without reference answers. ArXiv · AI/CL/LG's note
The paper presents a training-free failure attribution framework for LLM multi-agent systems. It audits source-to-decision handoffs, checking whether a receiving agent used prior information faithfully and within task requirements. The method separates error introduction, propagation, and repair before making a global diagnosis across the run. On 1,228 failed runs, it reports best or tied-best step accuracy among eight strategies without reference answers. ArXiv · AI/CL/LG's note
score 5