PRISM trains multi-agent topology generators with privileged sibling graphs
Original title: PRISM: PRIvileged Sibling-guided Multi-Agent Topology Design
The paper says PRISM beat every topology-design baseline across six benchmarks, and code is released.
PRISM is a post-training method for generators that choose which LLM agents participate and how they communicate. A teacher reads a sibling graph that also solved the query, through a zero-initialized low-rank branch unavailable at deployment, and the student matches its per-decision targets on self-sampled trajectories. The authors say this supplies a per-query preference that continued supervised fine-tuning lacks.
- Corpora of successful graphs do not rank graphs for each query.
- Without a preference, generators drift toward the most common topologies.
- Matched-teacher controls are cited as support for the privileged signal.
- The sibling graph is not available when the student is deployed.
From the original article
1 min read at huggingface.co
Multi-agent systems (MAS) based on large language models (LLMs) have become a powerful way to address complex problems, and their effectiveness depends critically on the collaboration topology that decides which agents participate and how they communicate. Recent work learns to generate this topology from the task query, fitting the generator to corpora of execution-verified graphs.
By Phuong Tuan Dat, Qi Li, Xinchao Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.