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 reasoning that matches agents via semantic similarity between lightweight Query (need) and Key (supply) descriptors using 384-dimensional sentence embeddings (all-MiniLM-L6-v2). Instead of fixed chain, star, fully-connected, or tree communication structures that cause context pollution and bottlenecks, DyTopo rebuilds the communication graph each round and breaks dependency cycles with a greedy O(n²) cycle-breaking heuristic. On the HumanEval code generation benchmark with 5 worker agents, a DyTopo system powered by Llama-3-8B outperformed the strongest baseline driven by GPT-OSS-120B, achieving 92.07% accuracy (+6.3%), while consuming only 48% of the tokens and 56% of the latency of AgentScope, and converging in 2.6 rounds versus 5. The framework also offers interpretability: per-round communication graphs enable post-hoc diagnosis of failures. Limitations include dependence on descriptor quality, hallucination propagation risk in dynamic networks, and task-sensitive hyperparameters. Beyond AI, the results suggest organization design lessons: precise, demand-driven collaboration among smaller units can beat large monolithic ones.

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-driven baseline on HumanEval code generation — +6.3% accuracy at roughly half the token cost and latency.

The Problem: Why Multi-Agent Collaboration Fails

Traditional multi-agent LLM systems suffer from two core issues:

1. Context pollution — In broadcast-style "group chat" communication, every agent's output goes to everyone, so each agent's context window fills with mostly irrelevant information, breaking reasoning chains. 2. Rigid static topologies — Fixed communication structures each have fatal flaws:

| Topology | Structure | Fatal flaw | |---------|---------|----------| | Chain | A → B → C → D | Slow propagation; upstream errors cascade | | Star | All → hub | Hub bottleneck, single point of failure | | Fully-connected | Everyone ↔ everyone | Worst context pollution, O(n²) complexity | | Tree | Hierarchical | Cross-branch collaboration is hard; information distortion |

Different task phases (brainstorming, implementation, verification) need different collaboration patterns — something static topologies cannot provide.

DyTopo's Approach: From "Group Chat" to "Free Market"

DyTopo replaces central planning with demand-driven matching:

Query-Key Semantic Matching

Each round, every agent outputs two lightweight descriptors:

Query: what information it currently needs
Key: what information it can currently provide

A 384-dimensional sentence embedding model (all-MiniLM-L6-v2) converts these into vectors; cosine similarity above a threshold creates a directed edge (B → A: B sends to A whose Query matched B's Key).

Dynamic Graph Construction

The communication graph is rebuilt every round:

Exploration phase: sparse topology, broad information gathering
Integration phase: denser topology as information converges
Verification phase: sparsifies again into precise verification

Greedy Cycle-Breaking

Semantic matching can create circular dependencies (A needs B, B needs C, C needs A). DyTopo's greedy cycle-breaking heuristic: 1. Sorts all Query-Key pairs by similarity 2. Adds edges from high to low 3. Discards any edge that would form a cycle 4. Stops when all legal edges are added or a sparsity threshold is reached

This avoids deadlock and produces high-quality sparse topologies in O(n²) time.

Results: Why 8B Beats 120B

HumanEval benchmark (5 worker agents):

| Method | Model 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 |

Key numbers:

DyTopo converges in 2.6 rounds vs. fixed 5 rounds for baselines
Token cost is only 48% of AgentScope
Latency is 56% of AgentScope

The explanation: each small model processes only highly relevant context instead of being flooded with noise — precise information flow lets a well-organized team of small models exceed a large monolithic one.

Interpretability Bonus

Every round yields a clear communication graph (who talked to whom and why). Failed tasks can be diagnosed by tracing where information flow broke down — e.g., a debugging agent's Query finding no matching Key reveals a missing role in the team.

Limitations

Descriptor quality dependency: inaccurate Query/Key descriptions lead to wrong routing
Hallucination propagation: errors can spread through multi-hop links; the paper recommends safety filters and application-specific guardrails
Task sensitivity: communication budget and sparsity thresholds differ per task type; no automatic tuning yet

Implications for Human Organizations

DyTopo doubles as a lens on organization design: move from rigid hierarchies (static star topologies) toward demand-driven dynamic collaboration networks — no fixed reporting lines, no departmental walls, teams reassembled per project phase, and transparent "what I offer / what I need" mechanisms.

The core insight: intelligence is not just an attribute of individuals, but of how they are connected. An 8B model in the right topology can unleash collective wisdom beyond its individual capacity, while a 120B model in the wrong topology drowns in information overload.

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