Reward-Guided Topology Generation

The multi-agent communication graph is emitted by a trained autoregressive generator, and that generator is then fine-tuned RLHF-style against a learned reward model that jointly scores task correctness and structural compactness, so the topology designer itself is optimised to produce sparse graphs that keep accuracy while cutting token cost.

Properties

family: Dynamic Topology
maturity: frontier
diagram: dynamic-graph
whyMultiAgent: The object being generated is the wiring among several distinct agents; with one agent there is no communication graph to design, and the reward signal (task correctness of the resulting team versus token cost of its message edges) only exists once a team actually executes under the generated topology. Distinct from the already-tracked dynamic-topology-routing (DyTopo), which rebuilds the graph each round at inference time via a semantic need/offer matcher, and from topology-task-routing (AdaptOrch), which classifies a task into one of a few canonical orchestration shapes: here the topology is the output of a learned generative model, produced edge by edge per task, and the novelty is closing the loop with a reward model so sparsity is explicitly incentivised rather than emerging by accident. Also distinct from two-timescale-coevolution, which edits a live team's topology slowly during execution: the generator designs the topology up front, before the team runs.
useCases: [Token-budgeted multi-agent system design (reported average 20.5 percent token reduction versus ARG-Designer at matched accuracy)][Automatic per-task team wiring in agent orchestration frameworks][Cost-aware search over communication structures for reasoning benchmarks]
source: RGA-Designer · arXiv:2608.20099

Provenance

Source: https://arxiv.org/abs/2608.20099

Reward-Guided Topology Generation

Properties

family

Dynamic Topology

maturity

frontier

diagram

dynamic-graph

whyMultiAgent

The object being generated is the wiring among several distinct agents; with one agent there is no communication graph to design, and the reward signal (task correctness of the resulting team versus token cost of its message edges) only exists once a team actually executes under the generated topology. Distinct from the already-tracked dynamic-topology-routing (DyTopo), which rebuilds the graph each round at inference time via a semantic need/offer matcher, and from topology-task-routing (AdaptOrch), which classifies a task into one of a few canonical orchestration shapes: here the topology is the output of a learned generative model, produced edge by edge per task, and the novelty is closing the loop with a reward model so sparsity is explicitly incentivised rather than emerging by accident. Also distinct from two-timescale-coevolution, which edits a live team's topology slowly during execution: the generator designs the topology up front, before the team runs.

useCases

[Token-budgeted multi-agent system design (reported average 20.5 percent token reduction versus ARG-Designer at matched accuracy)][Automatic per-task team wiring in agent orchestration frameworks][Cost-aware search over communication structures for reasoning benchmarks]

source

RGA-Designer · arXiv:2608.20099