When Evidence Shapes Collaboration: Knowledge-Conditioned Topology Generation for Multi-Agent Systems

2608.27984 ↗ · 2026-08-28 · cs.AI

Yangxiao Jiang, Jiarun Fan, Mingcong Xu, Yanxi Guo, Jiwen Feng, Shanqing Xu, Mengchen Qian, Wei Chen, Xiaojin Zhang

The finding

Headline claim

Conditioning multi-agent collaboration topology on retrieved external evidence (K-GAT) improves accuracy on knowledge-intensive benchmarks while cutting token consumption, beating LLM-Debate on GPQA by +15.7% with less than half the tokens.

The test

K-GAT, a trained evidence-conditioned autoregressive topology generator over a Qwen-3-8B backbone, was evaluated on seven benchmarks (MMLU, MMLU-Pro, GPQA, StrategyQA, GSM8K, AQuA, HumanEval) against static topologies (chain/star/tree), MAS baselines (LLM-Debate, AgentPrune, AgentDropout, AFlow), and larger single models, with curriculum training from execution-scored candidate workflows and ablations, retrieval-depth sensitivity, and efficiency analysis.

Effect size & scope

The headline number is +15.7% accuracy over LLM-Debate on GPQA at under half the token cost; the paper reports consistent average gains over Qwen-3-8B-backbone baselines across all seven tasks, but the full text is truncated before the complete main-results table, so per-benchmark deltas beyond GPQA cannot be verified; results are per-task QA/code benchmarks with a single 8B open backbone and task-specific role inventories.

Limitations

The paper itself concedes training requires expensive execution-based candidate mining (multi-run execution and pruning per training instance), reliance on a structured knowledge source with provenance (KG construction is nontrivial), and task-specific topological constraints per benchmark. Unstated: single backbone model, fixed retrieval top-K, structural cost weights (lambda_V, lambda_E, lambda_C) are hand-tuned, and generalization to open-ended or non-QA work is untested; no human-in-the-loop or real enterprise workload evaluation.

Counterpoint

A skeptic would note the comparison is against MAS baselines sharing one small 8B backbone while larger single models (Qwen-3-32B averages 74.99, Llama-3.1-70B 73.56) sit in the same table and may already rival or beat the MAS gains on several tasks, making the whole multi-agent apparatus potentially unnecessary with a stronger model; benchmark accuracy on fixed-answer QA sets is also exactly the setting where topology search can overfit, training cost may erase the inference-token savings, and the result table provided is incomplete so the headline claim rests on one benchmark's margin.

Economy implication

Little or nothing yet for AI economics: this is an architecture/methods paper on orchestration efficiency. At most, if the token-halving with accuracy gains replicates, it points toward falling inference cost per unit of task accuracy in agentic workflows, which is directionally relevant to who captures savings, but the evidence base (one small backbone, curated benchmarks) is far too thin to inform pricing, capex, or labor decisions in the next 12 to 24 months.

Who this matters for

Technical Capability: MAS/agent-systems researchers and engineers should care because it directly attacks over- and under-planning in dynamic topology generation and shows evidence-conditioning is a trainable, measurable lever on accuracy-per-token.

Market & Valuation: Only weakly relevant: token-efficiency claims in agentic orchestration bear on inference cost curves, but this paper's lab benchmarks do not yet establish anything about commercial agent products or vendor margins.

Abstract

Multi-Agent Systems (MAS) have recently moved from static workflows toward dynamically generated collaboration topologies. However, existing topology generation methods rely primarily on the parametric knowledge of large language models, with external search or retrieval used only as a reactive tool rather than an explicit determinant of collaboration structure. This leads to structure-knowledge misalignment, where systems exhibit redundant interactions or insufficient verification in knowledge-intensive tasks. We propose K-GAT (Knowledge-Guided Agent Topology Generator), a neuro-symbolic framework that formulates collaboration topology design as a knowledge-conditioned structure learning problem, integrating external evidence directly into autoregressive graph generation. Extensive experiments on knowledge-intensive benchmarks demonstrate K-GAT's efficiency and effectiveness: notably on the expert-level GPQA dataset, K-GAT outperforms the LLM-Debate baseline by a substantial margin of +15.7% in accuracy, while consuming less than half the computational tokens.