Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2608.10646.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

A research paper proposing ASCon, a model for failure attribution in LLM-based multi-agent systems. It uses direction-aware graph attention and reciprocal agent-step contextualization to identify faulty agents, erroneous steps, and failure modes. Experiments report 5.83% improvement in faulty-agent detection, 10.63% in faulty-step detection, and 14.73% in failure-mode detection, with out-of-domain generalization for LLM-based methods.

Provides a unified framework for diagnosing failures in multi-agent LLM systems across three attribution targets. Concrete benchmark gains and out-of-domain results make it directly relevant to anyone building or debugging agent pipelines.

Cited by

No citations on record.