acceptodds
Under review as a conference paper at ICLR 2027

DyTopo: Recursive Self-Improvement through Topology Evolution for Multi-Agent Reasoning

Abstract

Multi-agent systems built from prompted large language models can improve multi-round reasoning, yet most existing pipelines rely on fixed, trajectory-wide communication patterns that are poorly matched to the stage-dependent needs of iterative problem solving. We introduce DyTopo, a manager-guided multi-agent framework that reconstructs a sparse directed communication graph at each round. Conditioned on the manager's round goal, each agent outputs lightweight natural-language query (need) and key (offer) descriptors; DyTopo embeds these descriptors and performs semantic matching, routing private messages only along the induced edges. Across code generation and mathematical reasoning benchmarks and four LLM backbones, DyTopo consistently outperforms over the strongest baseline (avg. +6.2). Beyond accuracy, DyTopo yields an interpretable coordination trace via the evolving graphs, enabling qualitative inspection of how communication pathways reconfigure across rounds.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Multi-agent, Large Language Model, Dynamic Topology, Semantic Matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_eD44OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DyTopo. Multiple worker agents first generate messages and then describe what they need and what they can provide. The system uses these descriptions to build a communication graph and sends selected messages into the next round. A manager updates the shared goal and decides when to stop.

Critical strengths
  1. DyTopo obtains the best result in all 16 model-task settings in Table 1. This suggests that the method is not effective only in one isolated case.
  2. The random-routing, low-similarity, and fixed top-k comparisons support the claim that message selection matters. Under a fixed five-round budget, DyTopo also reports lower token usage than several baselines.
Critical weaknesses
  1. The most important fixed-topology comparisons are missing. The paper does not compare against LLM Debate, nor does it include a fixed Solver-Verifier structure for the mathematical tasks. The paper lacks a same-framework control that freezes the semantic graph constructed in the first round. Therefore, it is unclear whether the gains come from semantic routing or from rebuilding the graph at every round.

  2. The comparison with direct dynamic methods is missing, and generalization is difficult to assess. The paper does not compare against methods such as DyLAN. The mathematical experiments use only 70-problem subsets of MATH-500 and Omni-MATH, while APPS and Omni-MATH are also used for threshold selection. Results with four models and threshold transfer provide some evidence, but they are not enough to establish stable performance on independent distributions.

Questions for authors
  1. Under the same model, roles, message budget, and total token budget, could you compare DyTopo with LLM Debate, a fixed Solver-Verifier structure for mathematics, and a version that freezes the semantic graph after the first round? These comparisons are needed to determine whether dynamic rewiring provides an independent gain.
  2. Could you include a directly relevant dynamic collaboration baseline such as DyLAN? Please report accuracy, number of model calls, and token usage, and explain how the costs of team selection and early stopping are handled fairly.
  3. Could you select the roles, prompts, and thresholds on an independent development set, then evaluate on a complete mathematical benchmark or a sufficiently large held-out test set that was not used for selection? Please report paired confidence intervals. The current 70-problem subsets and partial evaluation-set reuse make generalization difficult to assess.
Reviewer_9bdDOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents DyTopo, a training-free multi-agent communication framework that reconstructs a directed communication graph at each round based on semantic matching between agents' self-reported needs and offers. Private messages along selected edges are aggregated into downstream agent contexts, while a centralized Manager updates shared subgoals and determines task termination. The approach is evaluated across four coding and mathematical reasoning benchmarks using four LLM backbones.

Critical strengths

S1: Formulating communication via need–offer semantic matching offers an intuitive, interpretable routing mechanism without requiring reinforcement learning or dedicated controller training. Decoupling message generation from subsequent routing gives the recursive graph update a clear operational meaning.

S2: Evaluations across four LLM backbones demonstrate consistent accuracy improvements over the evaluated agent baselines. Notably, edge-matched counterfactuals confirm the utility of semantic feedback selection, while DyTopo maintains around 35% lower token consumption compared to AgentScope at five execution rounds.

Critical weaknesses

W1: While the comparison between high- and low-similarity routing in Section 5.2 shows that semantic feedback selection improves task performance by 14.02–25.61 percentage points, both conditions reconstruct the communication graph at every round. Consequently, this ablation cannot isolate the specific advantage of per-round dynamic rewiring from the benefit of constructing a task-specific semantic topology. A semantic graph frozen after the first round might capture much of the performance gain. Therefore, the additional benefit of repeated graph reconstruction remains unclear.

Questions for authors

Q1: Could you provide experimental results where the semantic graph generated in the first round is frozen across all subsequent rounds, keeping the Manager, agent prompts, stopping criteria, and token budgets identical? What is the performance gap between this frozen-graph baseline and fully dynamic DyTopo?

Q2: The similarity thresholds (0.3 for code, 0.4 for math) were selected directly on APPS-Competition and Omni-MATH. How do these benchmarks perform when thresholds are determined strictly on independent validation sets, to rule out potential tuning bias?

Reviewer_2pQROpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces DyTopo, a manager-guided framework that reconstructs per-round sparse communication graphs via natural-language query–key semantic matching. Across four reasoning benchmarks and four LLM backbones, DyTopo achieves strong performance gains with significantly reduced token consumption and runtime.

Critical strengths
  1. DyTopo constructs per-round directed graphs via semantic need–offer matching with a frozen encoder. This enables stage-adaptive context pruning across LLM backbones without training task-specific routers.
  2. The framework consistently outperforms the strongest baselines across four coding/math benchmarks and four LLM backbones, while reducing token usage and runtime.
Critical weaknesses
  1. Model weights, roles, and matching rules remain strictly frozen, and extra rounds can degrade accuracy. Transient within-task context routing is not persistent self-improvement, making the conceptual contribution thin.
  2. Thresholded cosine similarity on prompted need–offer strings is vulnerable to phrasing mismatches and misdiagnosed errors ("unknown unknowns"), offering limited algorithmic depth.
  3. Because a central Manager rewrites shared round goals and controls early stopping, gains from dynamic edge routing are entangled with centralized orchestration.
  4. With only 3–4 predefined roles per task, complex embedding-based graph reconstruction offers questionable value over simpler routing.
Questions for authors
  1. The framework relies heavily on a central Manager to update shared round goals, track global state, and control early stopping. Could the authors clarify how much of the end-to-end performance and token efficiency gain stems specifically from the dynamic query–key edge rewiring versus the centralized orchestration and stopping control of the Manager? Was an ablation conducted where the Manager is replaced with a fixed static rule while keeping DyTopo's dynamic routing enabled?
  2. Workers generate natural-language query (need) and key (offer) descriptors, which are then matched via thresholded cosine similarity. When an agent fails or misunderstands its own error ("unknown unknowns"), its generated query may be fundamentally misaligned with what it actually needs. How robust is DyTopo to such flawed or misleading descriptors, and did the authors observe failure cases where poor queries trapped agents in unproductive communication loops?
  3. In tasks such as mathematical reasoning, the predefined worker pool consists of only 3 to 4 roles. Given such a tiny graph scale, what is the quantitative advantage of running a full embedding-based semantic matching and graph reconstruction pipeline every round compared to a well-designed static sparse topology or a simple fixed routing schedule?
Reviewer_PJnvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces DyTopo, an adaptive multi-agent communication framework that evolves outputs and communication topology during reasoning. Workers tells the community the information they need and can provide, and semantic matching dynamically constructs directed communication graphs to route relevant messages. A Manager coordinates the shared objective and stopping criteria. The key innovation is making communication topology an integral part of recursive self-improvement rather than fixing it before reasoning begins.

Areas for improvement:

1/ Mechanism analysis the contribution of dynamic topology evolution from semantic message selection and other components through more comprehensive ablations. 2/ Scalability: Evaluate communication and computational overhead as the number of agents. Robustness: Investigate sensitivity to inaccurate worker descriptors, noisy messages, and imperfect semantic matching. 3/ Convergence: Better characterize why additional reasoning rounds can reduce accuracy and develop more reliable stopping or topology-update strategies. 4/ Generality: Extend evaluation beyond coding and mathematical reasoning to broader tasks and heterogeneous multi-agent settings.

Critical strengths

evolves outputs and communication topology during reasoning.

Critical weaknesses

Cost and scalability need more study to show it is useful in practice

Questions for authors

Thanks for the great write up, interesting idea. Some questions on this framework are: how does Scalability look like, e.g. Evaluate communication and computational overhead as the number of agents? how to improve and evaluate robustness? e.g. sensitivity to inaccurate worker descriptors, noisy messages, and imperfect semantic matching. how is its Generality? have it been tried to extend evaluation beyond coding and mathematical reasoning to broader tasks and heterogeneous multi-agent settings.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Area Chair and All Reviewers

We sincerely thank the Area Chair and all reviewers for their time and constructive feedback. We are encouraged by your recognition of our intuitive, training-free need–offer semantic routing (PJnv, 2pQR, 9bdD), top performance across all 16 settings (2pQR, 9bdD, eD44), substantial token/runtime savings (2pQR, 9bdD, eD44), routing ablations (9bdD, eD44), and clear presentation (9bdD, eD44).

We briefly recap our core contributions: DyTopo pioneers within-task structural self-improvement by (1) formulating reasoning as a coupled update of agent state and topology G(t)G^{(t)}; (2) proposing training-free query–key matching with verification-gated halting; and (3) ranking 1st in all 16 settings (+5.2 pts mean gain; paired bootstrap +6.31 pts, p<0.01p<0.01), while cutting tokens by 35%–56% and latency by up to 53%.

Summary of New Experiments & Responses to Core Concerns

Below we summarize how our new experiments address each core concern (see individual replies for full tables):

1. Isolating dynamic rewiring from Frozen Round-1 topology & Manager (9bdD W1/Q1, 2pQR W3/Q1, eD44 W1/Q1, PJnv W1). On Qwen3-8B: (i) Frozen Round-1 Graph (G(t)≡G(0)G^{(t)}\equiv G^{(0)}) drops accuracy from 89.63% to 42.68% on HumanEval ($-46.95pts)and51.4346.95 pts) and 51.43% to 21.43% on Omni-MATH (-$30.00 pts), as Round-1 memories Hi(0)=∅\mathcal{H}_i^{(0)}=\emptyset lack draft solutions and miss late-stage verification edges (Developer<->Tester); (ii) Static Manager + Dynamic Routing reaches 78.05% on HumanEval (+35.37 pts over Frozen Round-1 and +31.10 pts over Single-turn [46.95%]), while dynamic Manager subgoals (89.63%/51.43%) synergize on multi-step tasks. With halting disabled (5 rounds), DyTopo still cuts tokens by 35% (MiMo) and 20%–31% (Qwen: 8,594 vs. 10,720–12,371).

2. Comparisons with LLM Debate, Fixed Solver–Verifier, and DyLAN (eD44 W1/W2/Q1/Q2, 2pQR W4/Q3). On Qwen3-8B (fairly counting all peer-rating and Manager costs): DyTopo (89.63% HumanEval, 75.71% MATH-500, 4,736 tokens, 15.3s) beats 5-round LLM Debate (68.29%/54.29%, 17,934 tokens) by +21.34/+21.42 pts with 73.6% fewer tokens, matched-budget LLM Debate ($\sim$8,500 tokens: 59.76%/48.57%) by +29.26/+27.14 pts, Fixed Solver–Verifier (Single-turn: 46.95%/57.14%; 5-round AgentScope: 80.00%/72.86%, 10,720 tokens), and DyLAN (77.44%/68.57%, 11,378 tokens, 35.4s) by +12.19/+7.14 pts with 58.4% fewer tokens and 56.8% lower latency.

3. Disjoint Dev/Test threshold validation (9bdD Q2, eD44 W2/Q3). We only coarsely selected τedge\tau_{\text{edge}} on one dataset per domain to avoid extreme channel density/sparsity. A strict 50% Dev →\to 50% Test Split yields +7.64 to +11.43 pts gains (21.36% vs. 13.72% on APPS Llama3; 37.14% vs. 25.71% on Omni-MATH Llama3; 48.57% vs. 40.00% on Omni-MATH Qwen3), alongside 100% held-out HumanEval (+5.75 pts, p=0.003p=0.003) and MATH-500 (+7.58 pts, p=0.001p=0.001).

4. Scalability, noise robustness, and generality (PJnv W2/W4/Q1–Q3, 2pQR W2/W4/Q2/Q3). (i) Scaling NN: Adding math agents into HumanEval (N=5,6,7N=5,6,7, Qwen3) shows DyTopo isolates irrelevant agents (89.63%, 89.63%, 89.02%) at 1.8–2.3 ms/round (<0.03%<0.03\% overhead). (ii) Noise: Under Gaussian score noise and 10% descriptor corruption, DyTopo reaches 88.41%/42.86% and 84.15%/40.00%, still beating unperturbed MasRouter (80.98%/38.72%). (iii) Generality: DyTopo beats AgentScope by +4.49 pts on MMLU-Pro (67.24% vs. 62.75%), and by +7.31/+4.29 pts (64.02% vs. 56.71%; 50.00% vs. 45.71%) in a mixed Qwen3+Llama3 pool.

With only about three days left in the discussion period, we warmly welcome any further questions or suggestions for prompt discussion!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.