PRISM: PRIvileged Sibling-guided Multi-Agent Topology Design

PRISM. PRIvileged Sibling-guided Multi-Agent Topology Design. By Phuong Tuan Dat and two others, from the National University of Singapore. Published on the twenty-ninth of September. Its code repository is worked out: about one hundred source files, with a readme.

Multi-agent systems built on large language models can take on complex problems, and how well they do depends heavily on their collaboration topology: which agents take part, and how they talk to each other. Recent work learns to generate that topology from the task query, fitting a generator to corpora of graphs that were verified by execution.

Post-training that generator is where things get stuck. A corpus lists the graphs that solved a query, but it never ranks them. Continued supervised fine-tuning therefore has no per-query preference to work with. Without one, the generator slides toward whichever topologies the corpus contains most often, rather than the ones that suit each query. A gain on one task can then be paid for with a loss on another.

PRISM does not go looking for new data or new labels. It recovers a per-query signal from that very same corpus, by learning with privileged information. The privileged information is a sibling graph: a second graph that also solved the query. It is available during training and unavailable at deployment, which is exactly what makes it privileged.

There are two roles. A teacher reads the sibling graph, and it does so through a zero-initialized low-rank branch. A student does not get the sibling. Instead the student samples trajectories of its own, and on those self-sampled trajectories it is trained to match the teacher's targets, one decision at a time. The teacher supplies a judgment informed by a second solution to the same query. The student learns to follow that judgment while acting only on what it will actually have at deployment, which is the query alone.

Across six diverse benchmarks, PRISM improves over every topology-design baseline the authors evaluated. That breadth bears on the original problem, where a gain on one task could come at the expense of another. Matched-teacher controls are reported as well, and they support the contribution of the privileged query information specifically.

Nothing new is collected here. The corpus that seemed to offer no ranking already held a second solution for a query, and treating that sibling as training-time context is what turns a flat list of successes into a signal the generator can learn preferences from. Code is available.