K-GAT, a multi-agent AI system that retrieves outside evidence before deciding how its agents should work together, reported 78.68% average accuracy across seven reasoning benchmarks. Its inference variant without the external knowledge graph, a linked store of information, reported 72.38%. The abstract also reports a 15.7% accuracy margin over LLM-Debate on GPQA while using less than half the computational tokens.
Evidence comes before the workflow
The paper starts with what it calls Structural Mismatch: workflows are generated from query semantics before external evidence is observed. It asks whether evidence-first topology generation can address the mismatch. In this context, a topology is the pattern of connections that sets out how agents collaborate.
At inference, K-GAT retrieves external evidence, generates a topology conditioned on that evidence, and executes the resulting workflow to synthesize an answer under structural constraints. The defining move is the order: retrieve first, then choose the collaboration structure.
Broad gains, with a clearer split on GPQA
Across the seven primary benchmarks, K-GAT reported 87.71% on MMLU, 66.42% on MMLU-Pro, 50.75% on GPQA, 84.97% on StrategyQA, 91.96% on GSM8K, 84.23% on AQuA and 84.75% on HumanEval. The set spans knowledge-intensive and symbolic or algorithmic reasoning.
In a controlled GPQA component comparison, the Qwen-3-8B base model scored 33.09%. The reported figures were 39.22% with the knowledge graph, 45.34% for G-Designer plus the graph, 50.00% for K-GAT without the verifier and 50.75% for full K-GAT. The paper also reports that K-GAT outperformed ReAct-style reasoning on GPQA, MMLU-Pro and StrategyQA.
Across MMLU-Pro and GPQA, K-GAT showed reported open-book gains of 15.35% and 10.56%, respectively. The paper says those gains were larger than the gains reported for listed single-agent baselines and for query-conditioned G-Designer.
Topology varied across evidence conditions
A manual analysis of 100 StrategyQA instances found an average topology size of 2.02 nodes under highly relevant, consistent evidence, compared with 2.66 nodes under conflicting or low-relevance evidence. The authors present this as a descriptive pattern, and no inferential uncertainty was reported for it.
On GPQA, reported accuracy increased as retrieval depth grew and saturated beyond six retrieved evidence units. The paper identifies a depth of three through five as the best accuracy-cost trade-off. In the reported implementation, the system retrieved the top three relevant documents for each query.
What it takes to train
Curriculum training used a fixed subset of 40 representative instances per dataset. For each instance, the system generated five initial candidate topologies and executed each candidate twice, producing 400 initial multi-agent-system executions per dataset before additional pruning and validation executions.
The reported implementation used Qwen-3-8B as its backbone, a maximum topology size of six, an indegree bound of three, AdamW optimization with a learning rate of 2e-5 and a batch size of four. The indegree bound is simply a cap on the number of incoming links allowed in the workflow.
A focused test with a larger backbone
A supplementary HLE Bio/Chem Gold evaluation used Qwen3-235B-A22B on 149 samples. K-GAT's accuracy was 0.1812, compared with 0.1342 for K-GAT without the knowledge graph, and it exceeded Chain at 0.1140, Star at 0.0939 and Tree at 0.1006.
The limits of the result
The evidence is limited to in-silico benchmark evaluations, primarily with Qwen-3-8B, plus a focused Qwen3-235B-A22B supplement. The study reports point estimates without confidence intervals, inferential tests or other uncertainty estimates. The margins therefore describe the reported runs, not a quantified uncertainty range.
The authors say performance depends on the quality and coverage of the Wikipedia-derived knowledge graph. They also caution that bounded topology size may limit long-horizon scalability, while broader tool-use environments and dynamically evolving knowledge sources were not evaluated. The work used public datasets without private user data or human annotations.
This is an arXiv preprint, version 2 dated 2 Sep 2026. The acknowledgements list support from Hubei Provincial Department of Science and Technology, the National Natural Science Foundation of China and the Natural Science Foundation of Hubei Province, China.
Paper data and sources
Original title: When Evidence Shapes Collaboration: Knowledge-Conditioned Topology Generation for Multi-Agent Systems
Authors: Yangxiao Jiang, Jiarun Fan, Mingcong Xu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text