ST-EVO: Towards Generative Spatio-Temporal Evolution of Multi-Agent Communication Topologies

Xingjian Wu; Xvyuan Liu; Junkai Lu; Siyuan Wang; Xiangfei Qiu; Yang Shu; Jilin Hu; Chenjuan Guo

cs.MAFeb 16, 2026

ST-EVO: Towards Generative Spatio-Temporal Evolution of Multi-Agent Communication Topologies

Authors: Xingjian Wu, Xvyuan Liu, Junkai Lu, Siyuan Wang, Xiangfei Qiu, Yang Shu, Jilin Hu, Chenjuan Guo,

Organizations: East China Normal University

Abstract

LLM-powered Multi-Agent Systems (MAS) have emerged as an effective approach towards collaborative intelligence, and have attracted wide research interests. Among them, ``self-evolving'' MAS, treated as a more flexible and powerful technical route, can construct task-adaptive workflows or communication topologies, instead of relying on a predefined static structue template. Current self-evolving MAS mainly focus on Spatial Evolving or Temporal Evolving paradigm, which only considers the single dimension of evolution and does not fully incentivize LLMs' collaborative capability. In this work, we start from a novel Spatio-Temporal perspective by proposing ST-EVO, which supports dialogue-wise communication scheduling with a compact yet powerful flow-matching based Scheduler. To make precise Spatio-Temporal scheduling, ST-EVO can also perceive the uncertainty of MAS, and possesses self-feedback ability to learn from accumulated experience. Extensive experiments on nine benchmarks demonstrate the state-of-the-art performance of ST-EVO, achieving about 5%--25% accuracy improvement.

Figures & tables

Figure 1: Existing multi-agent systems. Our proposed ST-EVO pioneers the exploration of Spatio-Temporal Evolving multi-agent systems.

Figure 2: The overview of ST-EVO . The framework follows a macro-to-micro logic: Construct & Schedule outlines the core macro-workflow where the Scheduler iteratively refines the communication topology $\mathcal{G}_{i}$ . Supporting this, Materials details the foundational components used to instantiate agents, while the Spatio-Temporal Multi-Agent System illustrates the step-wise topology evolution via graph flow matching. Finally, Optimize formulates the trajectory-retrieval regularized loss and the stability reward that align the Scheduler’s preferences.

PE	VE	System State Description
Low ( $\downarrow$ )	Low ( $\downarrow$ )	High Confidence (Deterministic)
High ( $\uparrow$ )	High ( $\uparrow$ )	Conflicting Uncertainty (Branching)
High ( $\uparrow$ )	Low ( $\downarrow$ )	Uniform Ignorance (Cluelessness)
Low ( $\downarrow$ )	High ( $\uparrow$ )	Overconfident Anomaly (Rare state)

Table 1: Taxonomy of System States based on Predictive Entropy (PE) and VarEntropy (VE).

Method	MMLU	GSM8K	MultiArith	SVAMP	AQuA	HumanEval	DS-1000	HotpotQA	DDXPlus	Avg.
Vanilla	80.47	87.15	93.40	87.25	69.27	73.28	38.40	82.35	56.40	74.22
Single-Agent Systems
CoT	81.85 ↑1.38	87.35 ↑0.20	93.60 ↑0.20	87.90 ↑0.65	72.40 ↑3.13	74.05 ↑0.77	42.35 ↑3.95	81.04 ↓1.31	64.22 ↑7.82	76.08
SC	82.95 ↑2.48	87.66 ↑0.51	94.40 ↑1.00	87.40 ↑0.15	71.85 ↑2.58	75.82 ↑2.54	37.66 ↓0.74	82.24 ↓0.11	66.43 ↑10.03	76.27
Multi-Agent Systems
Chain	82.40 ↑1.93	87.23 ↑0.08	92.88 ↓0.52	87.16 ↓0.09	69.85 ↑0.58	73.40 ↑0.12	38.25 ↓0.15	80.77 ↓1.58	65.28 ↑8.88	75.25

Table 2: Accuracy (%) comparison with four types baselines, including Single-Agent Systems, Multi-Agent Systems, Spatial Evolving Multi-Agent Systems, and Temporal Evolving Multi-Agent Systems. All baselines and ST-EVO are driven by gpt-oss-120b (Vanilla). Red : the best, Blue : the runner-up.

Figure 3: Visualization of performance metrics and token consumption of different MAS across MMLU, HumanEval, GSM8K, SVAMP. The diameters of the circles represent the scales of token consumption.

Figure 4: We utilize the accuracy (%) and

e^{\mathcal{R}_{\text{sta}}}\in[0,1]

to visualize the stability of ST-EVO . Higher values indicate better stabilities.

Figure 5: We compare the accuracy (%) of multiple multi-agent systems before and after prompt attacks on all benchmarks, and report the average results.

Variant	MMLU	HumanEval	SVAMP	GSM8K
ST-EVO	89.9	94.4	94.7	97.6
w/o Scheduling	84.7	89.6	88.6	92.5
w/o $\mathcal{L}_{\text{reg}}$	84.9	88.2	90.3	93.7
w/o Node Encoder	82.0	84.8	87.5	94.8

Table 3: Ablation studies tested on MMLU, HumanEval, SVAMP, and GSM8K benchmarks.

Appendix figures & tables40 assets

Supplementary material from the paper’s appendix.

Appendix

Entropy Estimator	Avg. Acc.	PE Corr.	Extra Calls
Logprob-based	94.15	1.00	0
Sample-based ( $K=5$ )	92.95	0.85	$5\times$
Sample-based ( $K=3$ )	92.23	–	$3\times$
Confidence probe	92.40	0.78	0
w/o PE/VE	91.80	–	0

Appendix

Table 4: Comparison of entropy estimators. Average accuracy is computed over MMLU, HumanEval, SVAMP, and GSM8K. PE Corr. is the Pearson correlation with logprob-based PE, and extra calls are measured per agent and iteration.

Category	Dataset	Answer Type	Metric	#Test	License
General reasoning	MMLU	Multi-choice	Acc.	153	MIT License
General reasoning	HotpotQA	Span	F1	7,405	CC BY-SA 4.0
Math reasoning	GSM8K	Number	Acc.	1,319	MIT License
	MultiArith	Number	Acc.	600	Unspecified
	SVAMP	Number	Acc.	1,000	MIT License
	AQuA	Multi-choice	Acc.	254	Apache-2.0

Appendix

Table 5: Dataset descriptions and statistics.

Method	SWE-Bench	Terminal-Bench	MLE-Bench	WebArena
Vanilla (single)	37.4	32.6	9.2	15.7
Chain (4 agents)	39.1	34.8	10.6	17.9
Complete Graph	40.8	36.3	12.1	20.3
G-Designer	44.2	39.7	15.8	24.6
SafeSieve	45.6	40.9	16.4	25.8
STEER	46.3	41.5	17.2	26.7

Appendix

Table 6: Results (%) on four challenging agentic benchmarks. All methods use gpt-oss-120b and the same communication-based multi-agent protocol.

Backbone	Architecture	Vanilla	Best Baseline	ST-EVO	Gain
gpt-oss-120b	120B Dense	74.22	83.91	88.42	$+4.51$
Qwen3.5-122B-A10B	122B MoE (10B active)	71.99	81.75	86.48	$+4.73$
Qwen3.6-27B	27B Dense	70.05	79.88	84.83	$+4.95$
Qwen3.6-35B-A3B	35B MoE (3B active)	64.16	75.18	81.14	$+5.96$
Qwen3-30B-A3B	30B MoE (3B active)	62.71	73.74	79.59	$+5.85$
Qwen3-Coder-30B-A3B	30B MoE (3B active)	61.67	72.50	78.29	$+5.79$

Appendix

Table 7: Average accuracy (%) across the nine benchmarks with different backbone models. Gain is measured against the strongest baseline for each backbone.

#Agents	5	10	20
Random Graph
Accuracy (%)	82.7	83.3	83.6
Time (min)	14.2	31.4	59.7
#Tokens	431,064	1,317,148	4,245,925
GPTSwarm
Accuracy (%)	83.6	84.2	84.8

Appendix

Table 8: Comparison of accuracy, time, and token consumption across different agent configurations. We use the MMLU benchmark and gpt-oss-120b as the base LLM.

#Iterations	3	6	12
Random Graph
Accuracy (%)	82.7	83.1	83.3
Time (min)	14.2	27.4	52.5
#Tokens	431,064	816,266	1,425,295
GPTSwarm
Accuracy (%)	83.6	84.0	84.2

Appendix

Table 9: Comparison of accuracy, time, and token consumption across different dialogue iterations. We use the MMLU benchmark and gpt-oss-120b as the base LLM.

Variant	MMLU	HumanEval	SVAMP	GSM8K	Avg.
ST-EVO	89.9	94.4	94.7	97.6	94.15
w/ LLM Scheduler	86.8	91.2	91.5	94.9	91.10
w/o Scheduling	84.7	89.6	88.6	92.5	88.85
w/ Random Scheduler	83.8	88.7	87.9	91.6	88.00

Appendix

Table 10: Ablation study on different scheduling strategies. We report the accuracy (%) on MMLU, HumanEval, SVAMP, and GSM8K, as well as the average performance.

Variant	MMLU	HumanEval	SVAMP	GSM8K	Avg.
ST-EVO	89.9	94.4	94.7	97.6	94.15
w/ Diffusion Scheduler	87.6	92.0	92.4	95.5	91.88
w/ VAE Scheduler	87.1	91.6	91.8	95.0	91.38

Appendix

Table 11: Ablation study on different generative mechanisms for topology scheduling. We report the accuracy (%) on MMLU, HumanEval, SVAMP, and GSM8K, as well as the average performance.

Variant	MMLU	HumanEval	SVAMP	GSM8K	Avg.
ST-EVO (full)	89.90	94.40	94.70	97.60	94.15
w/o Entropy Reward	87.23	91.76	92.41	95.82	91.81
w/o Retrieval	84.86	88.24	90.31	93.72	89.28
w/o Both	83.17	86.43	87.94	91.76	87.33

Appendix

Table 12: Combined ablation of the entropy-based stability reward and retrieve-augment module. All values are accuracy (%).

$\alpha$	$\beta$	MMLU	HumanEval	SVAMP	GSM8K / Avg.
0.3	0.7	88.17	92.68	93.14	96.33 / 92.58
0.5	0.5	89.85	94.36	94.72	97.60 / 94.13
0.7	0.3	89.06	93.51	93.85	96.92 / 93.34
1.0	0.0	87.43	91.74	91.96	95.48 / 91.65
0.0	1.0	86.72	91.15	91.37	94.93 / 91.04

Appendix

Table 13: Sensitivity to the stability-reward weights

\alpha

and

\beta

. Results are accuracy (%).

Capacity $L$	MMLU	HumanEval	SVAMP	GSM8K	Avg.
0	84.86	88.24	90.31	93.72	89.28
25	87.74	91.93	92.76	95.81	92.06
50	88.93	93.47	93.85	96.74	93.25
100	89.90	94.40	94.70	97.60	94.15
200	89.82	94.31	94.63	97.54	94.08
500	89.79	94.28	94.58	97.51	94.04

Appendix

Table 14: Sensitivity to trajectory-cache capacity. Results are accuracy (%).

$P$	MMLU	HumanEval	SVAMP	GSM8K	Avg.	Time (ms)
10	86.53	90.74	91.18	94.46	90.73	12
25	88.19	92.83	93.41	96.23	92.67	24
50	89.24	93.79	94.05	96.98	93.52	38
100	89.85	94.36	94.72	97.60	94.13	65
200	89.91	94.38	94.74	97.62	94.16	125
500	89.94	94.42	94.79	97.65	94.20	305

Appendix

Table 15: Accuracy–latency trade-off for the number of Flow-Matching integration steps

P

Method	Jaccard	Std.	Oscillation Rate
Random Scheduler	0.31	0.22	78.4%
G-Designer (fixed)	1.00	0.00	0.0%
ST-EVO w/o smoothing	0.72	0.15	22.3%
ST-EVO w/ smoothing	0.81	0.09	12.8%

Appendix

Table 16: Similarity and oscillation of communication topologies between adjacent dialogue iterations.

Variant	MMLU	HumanEval	SVAMP	GSM8K	Jaccard
w/o smoothing	89.85	94.36	94.72	97.60	0.72
$\lambda=0.01$	89.78	94.29	94.64	97.53	0.78
$\lambda=0.05$	89.71	94.18	94.51	97.42	0.81
$\lambda=0.10$	89.32	93.81	94.03	97.06	0.88
$\lambda=0.50$	87.46	91.53	92.18	95.27	0.95

Appendix

Table 17: Effect of topology-smoothing strength on accuracy (%) and adjacent-topology similarity.

Transfer Setting	Target Domain	Original on Target	Transfer Acc.	Drop
GSM8K $\rightarrow$ SVAMP	Math $\rightarrow$ Math	94.72	93.85	0.87
HumanEval $\rightarrow$ DS-1000	Code $\rightarrow$ Code	58.65	57.42	1.23
GSM8K $\rightarrow$ HumanEval	Math $\rightarrow$ Code	94.36	88.54	5.82
HumanEval $\rightarrow$ GSM8K	Code $\rightarrow$ Math	97.60	90.71	6.89

Appendix

Table 18: Generalization study via cross-dataset transfer.

A\rightarrow B

denotes that the scheduler is few-shot trained on dataset

A

and directly evaluated on dataset

B

. “Original on Target” denotes the performance of ST-EVO trained directly on the target dataset, while “Transfer Acc.” denotes the performance under the transfer setting.

Transfer Setting	Domain	Target-Trained	Zero-Shot	Drop
GSM8K $\rightarrow$ SVAMP	Math $\rightarrow$ Math	94.72	91.38	3.34
GSM8K $\rightarrow$ MultiArith	Math $\rightarrow$ Math	99.05	96.73	2.32
HumanEval $\rightarrow$ DS-1000	Code $\rightarrow$ Code	58.65	53.07	5.58
GSM8K $\rightarrow$ HumanEval	Math $\rightarrow$ Code	94.36	81.92	12.44
HumanEval $\rightarrow$ GSM8K	Code $\rightarrow$ Math	97.60	84.17	13.43
GSM8K $\rightarrow$ MMLU	Math $\rightarrow$ General	89.85	80.26	9.59

Appendix

Table 19: Zero-shot cross-task transfer. Target-Trained denotes a scheduler trained on the target dataset; Zero-Shot directly transfers the source scheduler without target-task queries. All values are accuracy (%).

Hyperparameter	Value
Backbone LLM	gpt-oss-120b
Prompt template	Same as ST-EVO
Decoding temperature	$0.1$
Optimization target	Graph edges and prompts
Graph optimizer	REINFORCE
Initial edge probability	$0.10$

Appendix

Table 20: Key hyperparameter settings of GPTSwarm.

Hyperparameter	Value
Backbone LLM	gpt-oss-120b
Prompt template	Same as ST-EVO
Decoding temperature	$0.1$
Communication rounds	$3$
Node encoder	all-MiniLM-L6-v2
Node embedding dimension	$384$

Appendix

Table 21: Key hyperparameter settings of G-Designer.

Hyperparameter	Value
Backbone LLM	gpt-oss-120b
Prompt template	Same as ST-EVO
Decoding temperature	$0.1$
Search algorithm	Monte Carlo Tree Search
Workflow representation	Code-represented workflow
Optimized components	Workflow edges and node prompts

Appendix

Table 22: Key hyperparameter settings of AFlow.

Hyperparameter	Value
Backbone LLM	gpt-oss-120b
Prompt template	Same as ST-EVO
Decoding temperature	$0.1$
Routing granularity	Reasoning step
Draft model	gpt-oss-120b
Verifier model	gpt-oss-120b

Appendix

Table 23: Key hyperparameter settings of SpecReason.

Hyperparameter	Value
Backbone LLM	gpt-oss-120b
Prompt template	Same as ST-EVO
Decoding temperature	$0.1$
Routing granularity	Reasoning step
Routing signal	Logit-based confidence
Confidence estimator	Maximum logit confidence

Appendix

Table 24: Key hyperparameter settings of STEER.

Hyperparameter	Value
Backbone LLM	gpt-oss-120b
Prompt template	Same as ST-EVO
Decoding temperature	$0.1$
Initial topology	Fully connected graph
Pruning granularity	Communication edge
Initial edge scoring	Semantic compatibility

Appendix

Table 25: Key hyperparameter settings of SafeSieve.

Hyperparameter	Value
Backbone LLM	gpt-oss-120b
Prompt template	Same as ST-EVO
Decoding temperature	$0.1$
Topology generation paradigm	Autoregressive graph generation
Graph initialization	Empty graph
Node generation strategy	Role selection from agent pool

Appendix

Table 26: Key hyperparameter settings of ARG-Designer.

Method	Attack Ratio	MMLU	SVAMP	HumanEval	Avg.
Random Graph	Clean $(0/5)$	83.20	87.10	84.14	84.81
	Attack $(1/5)$	76.48 ( $\downarrow$ 6.72)	79.72 ( $\downarrow$ 7.38)	77.73 ( $\downarrow$ 6.41)	77.98 ( $\downarrow$ 6.83)
	Attack $(2/5)$	72.36 ( $\downarrow$ 10.84)	75.64 ( $\downarrow$ 11.46)	73.97 ( $\downarrow$ 10.17)	73.99 ( $\downarrow$ 10.82)
	Attack $(3/5)$	67.83 ( $\downarrow$ 15.37)	70.82 ( $\downarrow$ 16.28)	68.52 ( $\downarrow$ 15.62)	69.06 ( $\downarrow$ 15.75)
SafeSieve	Clean $(0/5)$	84.65	90.10	90.15	88.30
	Attack $(1/5)$	80.93 ( $\downarrow$ 3.72)	85.15 ( $\downarrow$ 4.95)	84.31 ( $\downarrow$ 5.84)	83.46 ( $\downarrow$ 4.84)

Appendix

Table 27: Robustness analyses under different fractions of compromised agents. We consistently use five agents for all multi-agent systems. We report the accuracy (%) on MMLU, SVAMP, and HumanEval, as well as the average performance. Attack

(k/5)

denotes that

k

out of five agents are injected with an adversarial system prompt. The values in parentheses indicate the absolute performance drop compared with the corresponding clean setting.

Method	Metric	5 Agents	10 Agents	20 Agents	40 Agents
Random Graph	Accuracy (%)	82.7	83.3	83.6	83.8
	Time (minutes)	14.2	31.4	59.7	118.5
	#Tokens	431,064	1,317,148	4,245,925	13,184,762
GPTSwarm	Accuracy (%)	83.6	84.2	84.8	85.1
	Time (minutes)	67.8	201.8	422.5	825.0
	#Tokens	1,672,201	5,103,328	14,054,825	39,487,316

Appendix

Table 28: Scalability comparison on MMLU under different numbers of agents.

Method	Metric	5 Agents	10 Agents	20 Agents	40 Agents
Random Graph	Accuracy (%)	83.6	84.1	84.6	84.9
	Time (minutes)	3.8	7.5	15.2	31.4
	#Tokens	158,426	316,917	641,382	1,296,745
GPTSwarm	Accuracy (%)	86.5	87.1	87.8	88.2
	Time (minutes)	8.9	17.6	35.9	73.8
	#Tokens	189,742	381,653	776,438	1,592,817

Appendix

Table 29: Scalability comparison on HumanEval under different numbers of agents.

Method	Metric	5 Agents	10 Agents	20 Agents	40 Agents
Random Graph	Accuracy (%)	86.6	87.1	87.5	87.8
	Time (minutes)	10.6	21.4	43.8	89.7
	#Tokens	648,372	1,294,816	2,631,457	5,382,904
GPTSwarm	Accuracy (%)	86.7	87.3	87.8	88.1
	Time (minutes)	25.8	52.6	108.4	223.7
	#Tokens	1,006,481	2,018,736	4,126,592	8,482,317

Appendix

Table 30: Scalability comparison on SVAMP under different numbers of agents.

Method	Metric	3 Iterations	6 Iterations	12 Iterations
Random Graph	Accuracy (%)	82.7	83.1	83.3
	Time (minutes)	14.2	27.4	52.5
	#Tokens	431,064	816,266	1,425,295
GPTSwarm	Accuracy (%)	83.6	84.0	84.2
	Time (minutes)	67.8	112.5	199.4
	#Tokens	1,672,201	3,025,162	5,925,102

Appendix

Table 31: Scalability comparison on MMLU under different numbers of iterations.

Method	Metric	3 Iterations	6 Iterations	12 Iterations
Random Graph	Accuracy (%)	84.1	84.5	84.8
	Time (minutes)	5.6	10.8	20.9
	#Tokens	246,382	467,915	834,276
GPTSwarm	Accuracy (%)	87.1	87.5	87.8
	Time (minutes)	18.4	32.7	60.8
	#Tokens	381,624	704,892	1,326,517

Appendix

Table 32: Scalability comparison on HumanEval under different numbers of iterations.

Method	Metric	3 Iterations	6 Iterations	12 Iterations
Random Graph	Accuracy (%)	86.6	87.0	87.2
	Time (minutes)	10.6	20.5	39.2
	#Tokens	648,372	1,227,762	2,143,815
GPTSwarm	Accuracy (%)	86.7	87.1	87.3
	Time (minutes)	25.8	42.8	75.9
	#Tokens	1,006,481	1,820,815	3,566,259

Appendix

Table 33: Scalability comparison on SVAMP under different numbers of iterations.

Methods	MMLU	GSM8K	MultiArith	SVAMP	AQuA	HumanEval	DS-1000	HotpotQA	DDXPlus	Avg.
LLM-Debate	83.28	89.95	95.45	88.30	73.52	85.24	49.26	84.55	71.33	80.10
G-Designer	84.63	93.35	97.60	89.52	76.30	89.40	54.82	87.66	76.42	83.30
ST-EVO (10 Queries)	85.35	93.85	97.45	89.85	77.60	89.70	55.05	88.10	76.95	83.77
ST-EVO (20 Queries)	87.10	95.20	98.20	91.35	81.10	91.55	56.35	89.75	79.15	85.53
ST-EVO (40 Queries)	89.85	96.85	99.05	93.70	84.90	93.30	57.65	91.55	81.40	87.58
ST-EVO (60 Queries)	90.10	97.60	99.12	94.72	86.10	94.10	58.30	92.20	82.25	88.28

Appendix

Table 34: Performance comparison under different numbers of training queries.

Component	Time (ms)	Percentage
LLM inference (4 agents)	8,500	93.9%
Entropy computation	45	0.5%
Flow-Matching scheduling	65	0.7%
Cache retrieval	12	0.1%
Topology parsing and routing	8	0.1%
Inter-agent message passing	420	4.6%

Appendix

Table 35: Inference-latency breakdown per dialogue iteration with four agents.

Metric	Value
Scheduler parameters	2.8M
Training hardware	1 $\times$ NVIDIA 3090 (24 GB)
Training GPU memory	4.2 GB
Training time (40 queries)	$\sim$ 35 min
Training time (80 queries)	$\sim$ 68 min
LLM calls for training (80 queries)	$\sim$ 2,400

Appendix

Table 36: Training and inference cost of the ST-EVO scheduler. Scheduling latency is measured with

P=100

integration steps.

Method	Training Time	GPU Memory	Parameters
G-Designer	$\sim$ 45 min	3.8 GB	2.1M
ARG-Designer	$\sim$ 55 min	4.5 GB	3.2M
SafeSieve	$\sim$ 50 min	3.5 GB	1.8M
ST-EVO	$\sim$ 68 min	4.2 GB	2.8M

Appendix

Table 37: Scheduler training-cost comparison under an 80-query budget.

	MMLU	GSM8K	MultiArith	SVAMP	AQuA
ST-EVO	89.82 $\pm$ 0.18	97.55 $\pm$ 0.21	99.01 $\pm$ 0.12	94.68 $\pm$ 0.24	86.48 $\pm$ 0.36
	HumanEval	DS-1000	HotpotQA	DDXPlus
ST-EVO	94.29 $\pm$ 0.31	58.72 $\pm$ 0.42	92.49 $\pm$ 0.27	82.43 $\pm$ 0.39

Appendix

Table 38: Mean and standard deviation of ST-EVO across three independent runs.

Multi-Agent Communication Topology

See all 34 papers

TacoMAS: Test-Time Co-Evolution of Topology and Capability in LLM-based Multi-Agent Systems
May 10, 2026Chen Xu, Yicheng Hu, Ruizi Wang +4Multi-Agent LLM Systems Multi-Agent System Optimization
1Carnegie Mellon University · University of Science and Technology of China · 3National University of Singapore +1
EvoMAS: Learning Execution-Time Workflows for Multi-Agent Systems
May 9, 2026Chengdong Xu, Kaiqiang Ke, Ziheng Liu +4Multi-Agent LLM Systems Multi-Agent System Optimization
Sun Yat-Sen University · ShanghaiTech University · Zhejiang University +1
Evolve as a Team: Collaborative Self-Evolution for LLM-based Multi-Agent Systems
May 28, 2026Zhezheng Hao, Tianfu Wang, Huanshuo Dong +7Multi-Agent System Optimization Multi-Agent Coordination
Zhejiang University · Hong Kong University of Science and Technology · Tencent

May 10, 2026J/K move · Enter open · S save

TacoMAS: Test-Time Co-Evolution of Topology and Capability in LLM-based Multi-Agent Systems

Chen Xu, Yicheng Hu, Ruizi Wang, Xinyu Lin+3

1Carnegie Mellon University · University of Science and Technology of China · 3National University of Singapore+1

Multi-Agent LLM Systems Multi-Agent System Optimization Multi-Agent Systems Multi-Agent Communication Self-Evolving Agents

Open paper AskPDF arXiv

Multi-agent systems (MAS) have emerged as a promising paradigm for solving complex tasks. Recent work has explored self-evolving MAS that automatically optimize agent capabilities or communication topologies. However, existing methods either learn a topology that remains fixed at inference time or adapt only the topology or capability during inference. We empirically and theoretically show that effective test-time evolution requires jointly adapting both axes, but on different time scales: capabilities should update rapidly to handle emerging subtasks, while the topology should evolve more slowly to preserve coordination stability. We then introduce TacoMAS, a test-time co-evolution framework for dynamic MAS. TacoMAS formulates MAS inference as a task of online graph adaptation, where nodes represent agents with role-specific capabilities and edges define their communication topology. During inference, a fast capability loop updates agent expertise using trajectory-level feedback, while a slow meta-LLM-driven topology loop performs agents' birth-death operations on MAS, including edge edit, agent addition, and agent removal. We further show that this fast-slow design drives MAS evolution toward a task-conditioned stable equilibrium. Experiments on four benchmarks demonstrate that TacoMAS outperforms nearly 20 multi-agent baselines, achieving an average improvement of 13.3% over the strongest baseline. The codes are released at https://github.com/chenxu2-gif/TacoMAS-MultiAgent.

Abstract

Multi-Agent Communication Topology

Explore similar work

TacoMAS: Test-Time Co-Evolution of Topology and Capability in LLM-based Multi-Agent Systems

EvoMAS: Learning Execution-Time Workflows for Multi-Agent Systems

Evolve as a Team: Collaborative Self-Evolution for LLM-based Multi-Agent Systems