EvolveMAS: A Unified Self-Evolution Framework for Decentralized Multi-Agent Systems
Authors: Yalun Wu, Qiwei Ma, Bingbing Xu, Dan Zhang, Long Bai, Boyang Wang, Jiatong Li, Zheqi Lv, Shaozu Yuan, Peiying Wang, Jiawei Wang, An Zhang, Tat-Seng Chua
Organizations: National University of Singapore · Hunan University · Institute of Computing Technology, Chinese Academy of Sciences · Xingchen AGI, China Telecom Artificial Intelligence Technology Co., Ltd. · Cornell University · Meituan · University of Science and Technology of China
Decentralized multi-agent systems (MAS) enable large language model agents to interact without centralized control. Agents need to adapt to other agents at test time and reuse accumulated experience to refine their interaction strategies for future interactions across cooperative, mixed-motive, and adversarial settings. However, existing approaches often develop adaptation mechanisms for particular interaction settings. This raises a central question: how can decentralized MAS support test-time adaptation and reliable experience reuse across interaction settings? We propose EvolveMAS, a unified self-evolution framework that supports experience-driven adaptation within and across episodes. EvolveMAS pairs each main agent with a dedicated MemoryAgent for test-time memory governance, which maintains isolated Structured Memory outside the main agent's reasoning context. EvolveMAS organizes adaptation into two complementary processes: Within-Episode Online Adaptation and Cross-Episode Knowledge Evolution. During Within-Episode Online Adaptation, MemoryAgent updates episodic, semantic, and working memory from interaction evidence to inform subsequent actions. During Cross-Episode Knowledge Evolution, the main agent uses memory and episode outcomes to distill reusable interaction knowledge comprising partner profiles and interaction strategies for subsequent episodes. Across five benchmarks and six evaluation configurations spanning the three interaction settings, EvolveMAS achieves the highest normalized average score of 77.2 among the compared methods. Ablations show that test-time memory governance and cross-episode knowledge reuse both improve performance. Interaction traces further illustrate how updated memory and distilled knowledge guide subsequent interactions. These results support experience-driven adaptation in decentralized MAS without updating model parameters.
Figures & tables
Figure 1: Three representative interaction regimes for decentralized MAS.
Figure 2: Results at a glance: the seven methods on the six configurations of Table 1 . Compare bar lengths only within a sector; EvolveMAS is dark blue.
Figure 3: Overview of EvolveMAS. The upper lane is the main agent’s Execute–Evaluate–Evolve cycle and the lower lane the MemoryAgent that maintains Structured Memory. Blue arrows are within-episode updates, green arrows cross-episode evolution.
Cooperative
Mixed-motive
Adversarial
Avg.
Method
BBH
HumanEval
SOTOPIA
Deliberation
WW-V
WW-W
AgentNet
79
95.7
0.554
2.60
51.0
45.0
61.6
ChatDev
73
95.7
0.509
2.06
53.0
49.0
59.3
SOTOPIA
72
95.7
0.618
3.14
55.0
57.0
65.6
LLM-Deliberation
74
93.9
0.450
4.92
64.0
46.0
67.5
Werewolf/DPIns
72
92.7
0.588
3.41
55.0
44.0
63.2
Table 1: Cross-setting results. WW-V and WW-W report village win and wolf deception success in one-sided Werewolf audits. Shading identifies source settings. Bold and underlined values denote the best and second-best results in each column, respectively, including ties. Avg. is the mean of six percentage-scaled metrics: SOTOPIA is multiplied by 100 and Deliberation by 100/6 .
Figure 4: Within-session cumulative evolution. The cooperative (a–b) and mixed-motive (c–d) blocks are stacked side by side, and the adversarial row reports symmetric Werewolf (e) with the one-sided ablations, scored as village win (f) and as wolf deception success (g). Lines are cumulative means; the thicker green line is the paired difference (EvolveMAS − native), on the right-hand axis, with its zero dotted and its scale set independently in each panel. Shading marks each panel’s early, unstable tasks or games.
Figure 5: LLM-Deliberation ablation. Rows vary memory governance; columns vary the availability of cross-episode Knowledge Ke=(Ge,Pe) . Each cell reports one run of 50 sessions; color scales are panel-specific. The effect summary reports average marginal effects on unanimity and the total gain.
Figure 6: Instrumented SOTOPIA example illustrating Within-Episode Online Adaptation (a) and Cross-Episode Knowledge Evolution (b). In (a), entries written at the current turn are green and carried-over entries are yellow.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Family
Native protocol
Native benchmark
ChatDev
Planner → developer → tester pipeline (3 agents)
HumanEval
AgentNet
5-agent heterogeneous parallel or routed reasoning
BBH
SOTOPIA
Buyer/seller/mediator bargaining (3 agents)
SOTOPIA
LLM-Deliberation
6-party multi-issue proposal voting
LLM-Deliberation
Werewolf/DPIns
5-player hidden-role deception and detection
Werewolf
Appendix
Table 2: Source method families for the transfer rows, with native protocol and benchmark.
Question
Check
Result
Team size †
BBH acc. at 1/2/4/5/6 agents
83/83/86/84/85; stable across sizes (Fig. 7 a)
Aggregation
EvolveMAS vs. static MoA (5 agents)
84.0 vs. 81.0 (Fig. 7 b)
Short horizon
GSM8K accuracy
94.0 vs. No Memory 92.0; maintained
Cost
SOTOPIA $/ep at ep 0 / 7
0.0084 / 0.0105 vs. No Memory 0.0071 / 0.0341 (Fig. 7 c)
Context
Occupancy at episode 7
0.62% vs. No Memory 6.54% (Fig. 7 d)
Appendix
Table 3: Systems checks. The team-size and aggregation rows report EvolveMAS only. Cost and occupancy compare EvolveMAS with the No Memory ablation, which carries the raw interaction history in context.
Figure 7: Systems stability checks. (a) Team-size scaling. (b) Structured cooperation versus static MoA aggregation. (c–d) SOTOPIA cost and context occupancy over eight episodes.
Model (tier)
Sem. raw
Sem. unique
Rule strategies
Abstraction level
Gemini-2.5-Pro (strong)
27
11
40
general rules
Gemini-3.5-Flash (mid)
4
4
40
subtask features
GLM-5.2 (mid)
11
10
40
boundary conditions
GPT-4.1 (mid)
27
27
40
procedural advice
Qwen3 (weak)
56
11
0
bare answers
Appendix
Table 4: Knowledge sedimentation under EvolveMAS on the heterogeneous BBH team ( n=100 tasks).
Figure 8: Knowledge intake and role-level gains under EvolveMAS. (a) Unique Semantic Memory entries per model. (b) Per-role SOTOPIA goal-score change on the 0–10 scale (EvolveMAS − No Memory).
System
Strong lineup
Weak lineup
AgentNet (native)
84.0
79.0
Self-Memory
84.0
82.0
EvolveMAS
82.0
83.0
ChatDev
81.0
73.0
SOTOPIA
84.0
72.0
LLM-Deliberation
83.0
74.0
Appendix
Table 5: BBH under the weak and the strong lineup. The weak lineup is the main-table setup; the strong lineup is the five-model team used in the appendix analyses, on which the native AgentNet baseline reaches 84.0%. The transfer families keep their native role structure — the three-role ChatDev pipeline, the SOTOPIA bargaining trio, the five-model LLM-Deliberation parallel loop, and the five-player Werewolf/DPIns team — populated with strong-lineup models, and answer the same 100 items under the same cooperative loop and highest-confidence selection. Answers are scored by option-letter extraction for multiple-choice items and exact match on the remaining subtasks. The strong-lineup runs form a separate batch from the ablation of Table 6 .
MemoryAgent
Self-Memory
Benchmark
+Ke
K=∅
+Ke
K=∅
LLM-Deliberation
5.40
5.36
5.24
4.90
BBH
85.0
83.0
84.0
83.0
HumanEval
97.6
96.3
95.7
97.0
SOTOPIA
0.714
0.687
0.697
0.698
Appendix
Table 6: 2×2 factorial on four benchmarks, one designated execution per condition. Bold marks the highest value in each row. BBH answers are scored as in Table 5 .
Table 8: Werewolf memory configurations on a single village-win scale ( n=100 per row). Rows list the village-side policy first.
Figure 9: LoCoMo ablation (gpt-4.1 judge). (a) Overall accuracy. (b) Per-category comparison of Self-Memory and EvolveMAS. (c) Memory model sensitivity, with the dashed line marking Self-Memory.