DyTopo: Dynamic Topology Routing Lets an 8B Model Beat a 120B Model in Multi-Agent Reasoning
> Paper: DyTopo: Dynamic Topology Routing for Multi-Agent Reasoning via Semantic Matching > arXiv: 2602.06039 > Authors: Yuxing Lu et al. > Key finding: An 8B-parameter Llama-3 system beats a 120B-parameter GPT-OSS system on code generation
Key points
- Counterintuitive result: On the HumanEval benchmark, a DyTopo multi-agent system powered by Llama-3-8B outperformed the strongest baseline powered by GPT-OSS-120B — +6.3% accuracy with roughly half the token cost and latency.
- Problem: context pollution. In conventional multi-agent systems where every agent broadcasts to every other agent (a "group chat"), each agent's context window fills with mostly irrelevant information, breaking reasoning chains.
- Problem: static topologies. Fixed structures (chain, star, fully-connected, tree) each have fatal flaws and cannot adapt to different task phases (brainstorming, implementation, verification).
- Solution: semantic-matching routing. Each agent outputs a lightweight Query (what it needs) and Key (what it offers). A 384-dim sentence embedding model (all-MiniLM-L6-v2) computes cosine similarity; when A's Query matches B's Key above a threshold, a directed edge B → A is added. The graph is rebuilt every round, evolving from sparse (exploration) to denser (integration) and back to sparse (verification).
- Cycle handling: A greedy cycle-breaking heuristic sorts all Query-Key pairs by similarity and adds edges highest-first, discarding any edge that would create a cycle — avoiding deadlock in O(n²) time.
- Descriptor quality: inaccurate Query/Key descriptions lead to wrong routing.
- Hallucination propagation: errors can spread through multi-hop paths; the authors recommend safety filters and application-specific guardrails.
- Task sensitivity: communication budgets and sparsity thresholds differ per task; no automatic tuning yet.
Benchmark results (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 |
DyTopo converged in only 2.6 rounds on average, used 48% of AgentScope's tokens, and 56% of its latency.
Why a small model can win
A large single model operating alone must检索 through an overload of context, while a well-organized team of small models exchanges only highly relevant, targeted information. Precise information flow lets each 8B agent operate near its effective ceiling. DyTopo's routing graphs also provide interpretability: failures can be traced to missing information matches (e.g., a debugging agent whose Query for input-format documentation found no matching Key).
Limitations
Broader implication
The paper's core thesis is that intelligence is not only a property of individuals but of connection structure. A properly organized network of small models can outperform a much larger monolithic model — a principle the author likens to Amazon's "two-pizza team" philosophy and dynamic, need-based human organizations.
Paper link: https://arxiv.org/abs/2602.06039