Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.29292.

Pipeline notes

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

Proposes CATS, a method that searches for compact agent teams within an LLM multi-agent supergraph by using an LLM attributer to compare execution traces and assign credit to individual agents. On BigCodeBench, improves mean accuracy by 5.3 points over score-only search with 54% fewer tokens; prunes all adversarial agents on HotpotQA.

Concrete method for pruning multi-agent topologies using per-agent credit assignment, with measured token savings and adversarial-agent exclusion across three benchmarks.

Cited by

No citations on record.