The problem with multi-agent systems based on LLMs does not end with model selection. There is a second, less visible layer: the collaboration topology, i.e. which agents participate and how they communicate. A wrong topology can turn an agent pipeline into unnecessary cost, multiplying tokens and calls without improving outcomes. PRISM, a method proposed by a research group, addresses a specific limitation of topology generators trained on corpora of execution-verified graphs: those corpora list which graphs solved a query, but they do not rank them by preference. Without per-query rankings, supervised fine-tuning tends to drift the generator toward the most frequent topologies in the corpus, rather than the ones that suit each request. A gain on one task can therefore come at a cost on another.

The core of PRISM is the use of privileged information. During training, a teacher reads a sibling graph that solved the same query but will not be available at deployment. This happens through a zero-initialized low-rank branch. The student, by contrast, does not see the sibling: it must align its per-decision targets on trajectories it samples itself. In practice, the missing per-query signal is recovered without requiring extra annotations or an explicit ranking. Across six diverse benchmarks, PRISM improves over every topology-design baseline evaluated, and matched-teacher controls support the contribution of privileged query information. Implicitly, the losers are generators that settle for the most frequent topologies: corpus frequency ceases to be an acceptable proxy once per-query signal can be recovered.

There is a detail that deserves attention beyond the lab. Topology is not just an accuracy issue: it is a lever on inference costs. In multi-agent systems, every communication edge between agents is a token multiplier and a source of latency. If the generator learns to choose graphs better suited to each query, it reduces superfluous interactions. For teams managing self-hosted deployments or on-premise clusters, where VRAM and compute capacity are limited by choice or constraint, this reduction has a direct impact on TCO. The paper does not explicitly discuss hardware: the saving emerges as a second-order effect, but it is precisely the kind of effect that matters to those who cannot absorb waste through cloud elasticity.

PRISM's mechanism also signals a structural shift in how we think about fine-tuning multi-agent systems. Previously, the problem was the lack of per-query preferences; now it is possible to exploit a signal already present in the corpus, provided one accepts introducing a low-rank branch that disappears at deployment. This approach reduces dependence on expensive preference data and could make topology-generator training more accessible in contexts where annotated data are scarce or constrained by sovereignty requirements. Teams operating in environments where data cannot leave the local perimeter can train on verified corpora without building external annotation pipelines. It is not an automatic solution, but it moves the bottleneck: the constraint is no longer preference collection, but the quality of the verified-graph corpus.

Open questions remain, of course. The paper evaluates six benchmarks, but generalization to dynamic topologies and systems with hundreds of agents is not guaranteed. Moreover, the zero-initialized low-rank branch is an architectural detail that could interact in non-trivial ways with quantization techniques or memory constraints. For those evaluating on-premise deployment, AI-RADAR offers analytical frameworks to weigh these trade-offs, but the decision always depends on the specific constraints of the infrastructure. The point today is that topology design is becoming a separate optimization problem, and methods like PRISM show that useful signal can hide inside apparently flat corpora.