An AI method for designing communication graphs—diagrams of which language-model agents exchange information—was reported to use 20.5% fewer tokens, the units used in the paper’s count of model text, than ARG-Designer while reporting benchmark accuracy at a similar level. The findings come from a version-one arXiv preprint dated 20 Aug 2026. Across six benchmark rows, RGA-Designer averaged 87.80±0.60% accuracy, compared with 87.55±0.49% for ARG-Designer. Average use was 3,032±205 tokens per task for RGA, versus 3,815±124 for ARG. The figures are means with standard deviations over 10 independent runs.
A graph with a reason to stay small
RGA-Designer treats the arrangement of agents and communication links as a graph-design problem. It assigns each candidate graph an explicit, rule-based reward that favors two goals at once: completing the task and remaining structurally compact. Because graph quality is treated as programmatically verifiable, the method uses that reward in place of human feedback. The central question is whether this compactness incentive can reduce token consumption without giving up task accuracy.
The reward is weighted toward correctness. Task completion contributes 0.6 of the score, agent count contributes 0.3 and edge count contributes 0.1. That ordering gives the task result more weight than the graph’s size, while still making the number of agents and links part of the design decision.
The generator is policy-optimized with an on-policy GRPO variant. During inference, it produces five candidate graphs for each query and selects the one with the highest reward. Its reward model uses role and query embeddings of 384 dimensions, five structural features and two GraphSAGE layers with residual connections and layer normalization.
The evaluation compared RGA-Designer with five baselines: Vanilla, G-Designer, AgentPrune, AgentDropout and ARG-Designer. The comparison was designed to test whether adding an explicit compactness signal could change token use without a reported loss of task accuracy.
Six tests, repeated across ten runs
The test set included 500 instances from GSM8K, 214 from AQuA, 500 from MultiArith, 500 from SVAMP, 500 from MMLU and 121 from HumanEval. Accuracy was the reported measure for the first five benchmarks; HumanEval was assessed with Pass@1.
The datasets were randomly split into 15 training samples, 25 fine-tuning samples and a remaining test portion capped at 500 samples. Each experiment was repeated over 10 independent runs, allowing the paper to report variation around its mean results rather than relying on a single run.
Welch’s t-test was used to compare mean accuracy and token usage between methods. For GSM8K, AQuA, MultiArith, SVAMP, HumanEval and MMLU, the reported accuracy p-values were 0.320, 0.735, 0.814, 0.160, 0.437 and 0.365 respectively. The paper therefore reports no significant accuracy degradation versus ARG-Designer in those comparisons.
The token gap was clearer than the score gap
The accuracy finding supports a narrow conclusion: within this experiment, no significant degradation was reported. It does not establish that the two methods are identical. The token comparison showed a larger numerical separation. RGA-Designer was reported to use 20.5% fewer tokens on average than ARG-Designer, with statistically significant reductions on five of the six benchmarks. MultiArith was the exception.
The paper also reports ablations, comparing the complete method with altered settings. Among those settings, the complete method had the lowest token-use mean for GSM8K, at 3,863 tokens; HumanEval, at 1,715; and MMLU, at 3,875. The ablation values are means over 10 runs.
Taken together, the reported pattern is specific: a graph generator that rewards both task completion and compactness achieved a similar six-benchmark accuracy average while using fewer tokens than the main comparator. The result does not show that every communication reduction will preserve performance in a different task or system.
What the benchmark cannot answer
The evidence is bounded by the six benchmark datasets, the stated splits and the 10-run protocol. The reported endpoints are benchmark task performance and average token use per task. The supplied evaluation does not provide direct measurements of latency, monetary cost, energy use or real-world deployment outcomes.
That boundary matters when interpreting the 20.5% figure. A lower token count is the result reported in this computational comparison; it is not, by itself, a demonstration that an operating system would be faster, cheaper or less energy-intensive. The study also does not establish how the same comparison would look beyond the tested benchmark setting.
Further testing would be needed across different backbone language models, heterogeneous multi-agent systems and open-ended tasks without explicitly verifiable answers. The supplied analysis also leaves open whether graph and preference data would need to be rebuilt when the underlying base model changes.
An early result from a preprint
The document is an arXiv preprint, version 1, dated 20 Aug 2026. No journal publication or peer-review status is reported in the supplied record. Its acknowledgment says the experiments used the University of Luxembourg HPC facilities.
For now, the finding is best read as a benchmark result for teams designing LLM-based multi-agent systems: RGA-Designer reported similar accuracy with lower token use than ARG-Designer in the tested setting. Whether that balance holds across other models, agent combinations, tasks and deployment conditions remains the central open question.
Paper data and sources
Original title: Reward-Guided Autoregressive Graph Generation for Efficient Multi-Agent Communication Topology Design
Authors: Poomphob Suwannapichat, Boonyarit Changaival, Caesar Wu, Pascal Bouvry
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text