DyTopo: Dynamic Topology Routing Lets an 8B Model Beat a 120B Model in Multi-Agent Reasoning

Forum topic · 小凯 · 2026-03-28

Summary

DyTopo (arXiv:2602.06039) is a dynamic topology routing framework for multi-agent LLM systems that replaces static communication structures with per-round, semantic-matching-based graph construction. Each agent emits a lightweight Query (what it needs) and Key (what it can provide); sentence embeddings (all-MiniLM-L6-v2) and cosine similarity decide which directed edges form each round, while a greedy cycle-breaking heuristic prevents deadlock in O(n²) time. On HumanEval code generation with five worker agents, DyTopo driven by Llama-3-8B reached 92.07% accuracy, beating the strongest baseline powered by GPT-OSS-120B by 6.3 points while using only 48% of the tokens and 56% of the latency, converging in 2.6 rounds versus fixed 5-round baselines. The framework also yields interpretable communication graphs for failure diagnosis. Known limitations include dependence on descriptor quality, hallucination propagation risk, and task-sensitive hyperparameters such as sparsity thresholds and communication budgets.

Overview

Paper: DyTopo: Dynamic Topology Routing for Multi-Agent Reasoning via Semantic Matching arXiv: 2602.06039 Authors: Yuxing Lu et al. Headline finding: An 8B-parameter Llama-3 driven DyTopo system beat a 120B-parameter GPT-OSS baseline on code generation.

Key points

Counterintuitive result: On HumanEval, DyTopo with Llama-3-8B achieved 92.07% accuracy — a 6.3% improvement over the strongest GPT-OSS-120B baseline — at roughly half the token cost (9,453 vs 19,520) and latency (22.3s vs 39.8s).
The problem — context pollution: In conventional "group chat" multi-agent systems, every agent broadcasts to everyone, flooding each context window with mostly irrelevant information that derails reasoning chains.
The problem — static topologies: Chain, star, fully-connected, and tree structures are fixed in advance, but different task phases (brainstorming, implementation, verification) need different collaboration patterns. Fully-connected designs scale at O(n²) and pollute context the most.
Core idea — semantic matching routing: Each round, every agent outputs a lightweight Query (what information it needs) and Key (what it can provide). A 384-dim sentence embedding model (all-MiniLM-L6-v2) encodes these; cosine similarity above a threshold creates a directed edge between agents.
Dynamic graph construction: The communication graph is rebuilt every round — sparse during exploration, denser during integration, sparse again during verification. Example evolution on a HumanEval task: Round 1: 2 edges among 5 agents; Round 2: 7 edges forming subgroups; Round 3: 3 precise connections.
Greedy cycle-breaking heuristic: Query-Key matching can create circular dependencies (deadlock). DyTopo sorts all candidate edges by similarity and adds them greedily, discarding any edge that would form a cycle, producing a high-quality sparse topology in O(n²) time.

Benchmark comparison (HumanEval, 5 worker agents)

| Method | Scale | Accuracy | Total tokens | Avg latency | |---|---|---|---|---| | Single-round agent | 4× | 88.41% | 2,835 | 6.7s | | Random topology | 4×5 rounds | 88.17% | 15,783 | 34.2s | | AgentScope | 4×5 rounds | 90.24% | 19,520 | 39.8s | | DyTopo | 5×2.6 rounds | 92.07% | 9,453 | 22.3s |

Why a small model team can beat a large single model

Precise information flow: Each 8B agent only processes highly relevant context instead of being drowned in noise, letting the team approach its performance ceiling. The analogy: five students with focused notes beat one person hauling an entire library into an exam.
Efficiency: DyTopo converged in 2.6 rounds on average versus fixed 5-round baselines, cutting token use to 48% and latency to 56% of AgentScope.
Interpretability: Every round yields an explicit communication graph. Failed tasks can be diagnosed — e.g., a debugging agent's Query for input-format documentation matched no Key, revealing a missing role in the team.

Limitations

Descriptor quality dependence: Inaccurate Query/Key descriptions lead to wrong routing decisions.
Hallucination propagation: Errors can spread multi-hop through the dynamic network; the authors recommend safety filters and application-specific guardrails.
Task sensitivity: Communication budgets and sparsity thresholds must be tuned per task (code vs. math), with no automatic tuning yet.

Broader takeaway

The paper's central insight is that intelligence is not only a property of individuals but of the connection structure. A well-organized team of small models can outperform a much larger monolithic model working in a polluted context — an AI echo of Amazon's "two-pizza team" principle, suggesting organizational designs based on need-driven, dynamically formed collaboration networks rather than fixed hierarchies.

Paper link: https://arxiv.org/abs/2602.06039