Learning How Much to Collaborate: Difficulty-Aware Topology Selection for Multi-Agent Code Generation
4.00T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.13890.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryPaper evaluates five communication topologies for multi-agent code generation on 614 problems, finding that hierarchical collaboration's advantage scales with problem difficulty (2.4 pass@1 points on easy problems to 21.1 on hard, at roughly 10x token cost). The proposed Difficulty-Aware Topology Selector, a graph network routing per problem, reaches 77.7% pass@1 at 40% of always-hierarchical cost across four backbones and two domains.
Why it mattersQuantifies when multi-agent code systems actually beat single agents and gives a cost-aware router to pick topology per problem. Useful for anyone deploying multi-agent LLM pipelines.
Cited by
No citations on record.
