TopoDIM: One-shot Topology Generation of Diverse Interaction Modes for Multi-Agent Systems

AI summary

Overview

Research area: Multi-agent systems (MAS) built on large language models, specifically the design and optimization of communication topologies between agents.

Technical level: Advanced. The paper combines heterogeneous graph neural networks, reinforcement learning with policy gradients, and knowledge distillation.

Scope: A framework called TopoDIM that generates a heterogeneous agent communication topology in a single inference step, replacing iterative multi-round dialogue in LLM-based multi-agent systems.

What This Paper Is About

Existing multi-agent systems mostly connect agents through repeated rounds of dialogue, where each round of interaction triggers further intra-round conversations, compounding token cost and latency. TopoDIM instead lets agents decide, in one shot, which other agents to talk to and in what manner (conditioned, feedback, or debate), using a generated graph rather than iterative conversation. The goal is to improve task performance while cutting communication overhead, and to do so with a decentralized design in which each agent makes its own connectivity decisions.

Key Contributions

1. Observation. The authors argue that replacing iterative topology generation across intra- and inter-round dialogue processes with a one-shot heterogeneous topology formulation improves communication efficiency without a performance trade-off.

2. Framework. They propose TopoDIM, a decentralized framework combining a heterogeneous graph encoder with an autoregressive decoder to generate multi-relational communication topologies in one shot, enabling autonomous agent-level decision-making.

3. Evaluation. Extensive experiments across homogeneous and heterogeneous multi-agent settings show consistent gains in communication efficiency, task performance, and structural robustness, surpassing strong task-adaptive cooperation baselines.

4. Diverse interaction modes. Three collaborative argumentation primitives are formalized: conditioned edges, feedback edges (evaluation and reflection), and debate edges.

Main Findings

• Headline efficiency and accuracy result: TopoDIM reduces total token consumption by 46.41% while improving average performance by 1.50% over state-of-the-art methods.

• Per-model average gains (Table 1): With Gemma-3-it:12B and DeepSeek-V3.2-251201, TopoDIM yields average performance gains of 1.35% and 1.38% over the strongest baselines. On the reasoning model GPT-OSS:120B it produces a 1.50% increase in average predictive accuracy. TopoDIM's overall averages are 62.23 (Gemma-3-it:12B), 90.32 (GPT-OSS:120B), and 88.53 (DeepSeek-V3.2-251201).

• Token savings by benchmark: Using GPT-OSS:120B, TopoDIM saves a total of 1.42M tokens on MMLU-Pro and 2.44M tokens on LiveCodeBench. Appendix D reports maximum prompt and completion token savings of 57.82% and 22.05% compared to the most efficient framework.

• Heterogeneous agent adaptability: In a MAS with three GPT-OSS-20B agents and two GPT-OSS-120B agents (the latter as final decision-maker), TopoDIM reaches SOTA on all datasets with an average improvement of 1.86% over existing homogeneous topology frameworks. Its adaptive sparsification prunes agents making negligible contributions.

• Edge diversity helps: Moving from only conditioned edges (Base.) to adding feedback and debate edges (+ Fbk./Dbt.) improves accuracy/pass@1 by 2.21% on MMLU-Pro and 2.18% on LiveCodeBench.

• Sparsity matters, and more is not better: The best accuracy/pass@1 (80.11% on MMLU-Pro and 87.28% on LiveCodeBench) occurs at sparsity budget ratio α = 0.7. Dense graphs risk redundant communication channels, hallucination, and impaired decisions.

• Data efficiency: Performance improves monotonically with training sample size M, with the most significant gains in the first 40 samples. The decentralized distillation also performs robustly with as few as 40 samples (M' varied from 0 to 160).

• Ablation results: Random heterogeneous edges (w/ Rand) cause average drops of 5.60%; replacing the prior graph with a fully connected graph (w/o Graph) drops 1.99%. Removing the moving-average baseline drops accuracy by 0.89%, and removing entropy regularization drops it by 1.38%.

• Memory footprint: With 1,000 randomly instantiated nodes, the decentralized design (a lightweight network per agent) requires merely 6.36 GB of memory.

• Case study finding: On LiveCodeBench, intra-round dialogue graphs lose information, and intra/inter-round paradigms add token overhead and hallucination risk, while TopoDIM builds a sparse but informative topology that preserves critical context.

Methodology in Plain English

The system treats a multi-agent team as a directed heterogeneous graph: agents are nodes, and edges carry one of three relation types. A conditioned edge means the receiver handles the query using the sender's output. A feedback edge means one agent critiques or validates another's intermediate output and the original agent then re-handles the query, mimicking evaluation and reflection. A debate edge means one agent challenges another's proposition for two rounds, after which the challenger proceeds with the query in the context of the debate.

To generate the topology, a relational graph convolutional network encodes each agent using a prior knowledge graph built from an advanced LLM (GPT-5), combined with role embeddings and a task-query embedding from a sentence encoder. An autoregressive decoder then samples the relation type for each ordered pair of agents, one pair at a time. Three mechanisms shape the output: an acyclic mask prevents circular dependencies for conditioned and debate edges; a TopK adaptive sparsification keeps only the highest-confidence edges under a sparsity budget; and inactive nodes with no valid interaction are filtered out. Execution then follows breadth-first order from the root node.

Training uses reinforcement learning. The reward is a weighted blend of a binary task-success signal and a structural diversity term — the Shannon entropy of the empirical edge-type distribution — so the policy does not collapse into a single interaction pattern. A moving-average baseline and entropy regularization stabilize and diversify training. Finally, the global policy is distilled into lightweight per-agent local networks (MLPs) by minimizing KL divergence, so that agents can pick their own connections at inference time without a central coordinator.

Experiments use Gemma-3-it:12B, GPT-OSS-20B, GPT-OSS-120B, and DeepSeek-V3.2-251201:671B across MMLU-Pro, AIME (2023-2025), MultiArith, GSM8K, LiveCodeBench (v1), and HumanEval, compared against Vanilla, Chain-of-Thought, LLM-Debate, GPTSwarm, G-Designer, and AgentDropout.

Why This Matters

Impact on research: The paper offers an alternative to the dominant spatio-temporal, multi-round dialogue paradigm in LLM-based MAS. It shows that explicitly modeling multiple interaction types in a heterogeneous graph can substitute for iterative dialogue rounds, and it connects argumentation/cognitive mechanisms (evaluation, debate) to graph topology design.

Real-world applications:

• Code generation pipelines, where multiple agents can condition, critique, and debate each other's patches before a final solution — evaluated here on LiveCodeBench and HumanEval.
• Mathematical and quantitative problem solving, where debate and feedback edges are used for tasks such as AIME and GSM8K.
• Privacy-sensitive or decentralized deployments, where no central orchestrator builds the topology, addressing privacy risks in multi-party inference.
• Mixed-capability agent fleets, where strong and lightweight models collaborate and weak contributors are automatically pruned.

Industry relevance: The reported token reduction (46.41%) and savings of 1.42M and 2.44M tokens on two benchmarks translate directly into serving cost and latency, which are primary constraints when deploying multi-agent workflows. The decentralized design's 6.36 GB footprint at 1,000 nodes and its compatibility with proprietary models (a topology-based rather than latent-space approach) support practical deployment. Deployment caveats remain: the authors note that serving heterogeneous LLMs with varying architectures requires engineering support.

Future Directions

• Extending beyond the three interaction modes. The authors state that the current scope does not cover complex organizational methods such as dynamic coalition formation, where subsets of agents spontaneously align on sub-problems.

• Better heterogeneous model mixing. The paper reports that simply introducing random agents yields only small gains when tasks exceed agent capabilities, and that finding the right combination of high-performance and lightweight LLMs is non-trivial.

• Further cost reduction. The authors note that MAS inevitably incurs more tokens and latency than single-LLM approaches, and that future work will explore mechanisms to improve cost-efficiency while improving performance, while staying compatible with proprietary models.

• Open design questions surfaced by the analysis: how edge diversity, sparsity, centralized training sample size, and decentralized distillation sample size interact at larger scales — the paper only reports sweeps over α ∈ {0, 0.3, 0.5, 0.7, 1.0}, M and M' from {40, 80} for optimization and 0 to 160 for distillation, and five initial agents.

Target Audience

Researchers and engineers working on LLM-based multi-agent systems, agent orchestration, and graph-structured communication will get the most from this paper. It also suits practitioners optimizing inference cost in multi-agent deployments, and readers interested in applying reinforcement learning, graph neural networks, or policy distillation to agent coordination. Some familiarity with graph representation learning and policy-gradient methods helps, though the core ideas of conditioned, feedback, and debate edges are accessible without it.

Authors’ abstract

Optimizing communication topology in LLM-based multi-agent system is critical for enabling collective intelligence. Existing methods mainly rely on spatio-temporal interaction paradigms, where the sequential execution of multi-round dialogues incurs high latency and computation. Motivated by the recent insights that evaluation and debate mechanisms can improve problem-solving in multi-agent systems, we propose TopoDIM, a framework for one-shot Topology generation with Diverse Interaction Modes. Designed for decentralized execution to enhance adaptability and privacy, TopoDIM enables agents to autonomously construct heterogeneous communication without iterative coordination, achieving token efficiency and improved task performance. Experiments demonstrate that TopoDIM reduces total token consumption by 46.41% while improving average performance by 1.50% over state-of-the-art methods. Moreover, the framework exhibits strong adaptability in organizing communication among heterogeneous agents. Code is available at: https://github.com/Sundiasy/TopoDIM.