Summary
这篇 HKUST/UIUC/CUHK/HKU/PKU 联合 survey 把 co-evolution(多个组件相互施加 adaptive pressure 的多组件 self-evolution)作为中心组织轴,提出渐进三阶段 taxonomy:Agent-Agent、Agent-Environment、Meta Co-Evolution。全文主线是”演化自由边界(boundary of evolutionary freedom)逐步扩大、系统逐步摆脱 human-engineered constraints”,并统一形式化为演化算子 Ω 作用域的扩张。
Problem & Motivation
Agentic system 部署后需要持续改进,但 single-entity self-evolution 受限于 static learning context(固定的 task 与 feedback),改进曲线有上界。既有 self-evolving agents surveys(Gao et al. 2026、Xiang et al. 2026 等)以单 agent 内部组件(model / memory / tool / harness)为组织轴,只把 co-evolution 当 subtheme。本文以 co-evolution 为 central inclusion criterion,追问三个问题:哪些组件在共演、演化自由边界如何跨组件扩张、这一过程的终极形态是什么。
Method
形式化框架(Sec. 2):系统 S=(A,E),每个 agent a_i=(m_i, h_i) 由 model backbone 与 harness 构成,A=({a_1..a_n}, Π) 中 Π 编码角色与通信拓扑。演化以算子 S^(t+1)=Ω(S^t, τ^t) 描述,Ω 规定 what/when/how/where evolve 与质量评估方式。co-evolution 的判别标准:至少两个 evolving units 相互适应并持续重塑彼此的后续演化,仅交换信息或交互不算。Appendix A 据此与十个相邻概念逐一划界(multi-agent interaction、agent loops、harness engineering、evolutionary optimization、continual learning、self-play、self-evolution、meta-evolution、recursive self-improvement、open-endedness)。
Stage 1 — Agent-Agent Co-Evolution(Sec. 3):A^(t+1)=Ω(A^t, E, τ^t),环境 E 固定,耦合演化发生在 agent collective 内部。三类:
- Adversarial:pairwise(GAN、ACE-Safety 的 MCTS 搜攻击、MAGIC 多轮攻防博弈)与 multi-source(AlphaStar league training、TriPlay-RL 攻击者/防御者/评估者三方共演);
- Collaborative:parallel(MARL、CoMAS 从讨论中生成 reward、CORAL 经共享 memory 扩散尝试与 skill)与 role-differentiated(CORY 先答后改、CoVerRL 从多数投票自举 verifier、EvoScientist 研究者/工程师双角色积累经验);
- Evolving organizations:角色与团队结构本身可演化(R3DM 从行为中发现角色、SkillMAS 联合更新 skill 与团队结构)。
Stage 2 — Agent-Environment Co-Evolution(Sec. 4):(A^(t+1), E^(t+1))=Ω(A^t, E^t, τ^t),环境状态进入更新,按被演化的子空间分三类:
- Task-space:exposure/selection(curriculum、PORTAL 选中间任务)与 adaptive generation(GenEnv 难度对齐共演、Tool-R0 按 solver 成功率校准难度、Search Self-Play 生成 multi-hop 可验证问题);
- Feedback-space:preference-driven(PEBBLE、DUO 挑 reward 分歧大的样本对)、outcome-driven(ROSKA 共搜 reward 候选与 policy 变体、CURE 从执行失败演化 unit test)、consistency-augmented(ARCO 步级分数与最终结果对账、ECHO 按建议是否改进 policy 更新 critic);
- Interaction-space:executable world construction(POET 演化 environment-agent pair 种群、OpenAI ADR 随 policy 变强扩大随机化范围、LLM-POET)与 model-based world construction(WebEvolver 共训 web world model 与 agent policy、COMAP on-policy 文本 world model、EvolvingAgent)。
Stage 3 — Meta Co-Evolution(Sec. 5):Ω^(t+1)=Γ^t(S^t, Ω^t, τ^t),下层共演系统通过自生成的 revision process Γ 修订演化机制本身。作者明确指出现有多数工作只是 single-entity precursor(PromptBreeder 演化 mutation prompt、MemEvolve 演化 memory 架构);更接近 true meta co-evolution 的例子仅 RQGM 一类——在 task agents 与 evaluators 共演之上,meta-agent 利用 joint feedback 引导后续演化。
Key Results
本文是 survey,无自有实验;核心产出为 taxonomy 与两项 survey 层面的证据整理:
- Figure 4 cross-paper evidence:汇总被综述论文的结果论证 co-evolution 的 effect / consistency / convergence——Panel A/B 为各论文内部 static-vs-evolving 的 matched 对比(均值/配对提升),Panel C 为归一化 performance 轨迹;数据选取与聚合方法在 Appendix C 说明。
- Appendix B Table 1 与 Gao et al. 2026、Xiang et al. 2026(self-evolving agents surveys)、Guo et al. 2024(LLM multi-agent survey)等逐项对比,声称的差异是它们均未以 co-evolution 为中心轴。
- 挑战三条(Sec. 6):dynamic evaluation(固定 benchmark 需辅以 historical cross-play、component ablations、held-out evaluators 等过程级测试);scaling(agents/harness/environment 联动更新时决定”谁变、更新如何互相影响、如何保持演化压力 productive 而非 unstable”);safety & governance(演化出的攻击策略与通信协议可能超出人类理解,需 sandboxed deployment、continuous monitoring、rollback to verified states、human intervention points)。
Evidence Ledger
| Claim ID | Claim | Type | Source locator | Evidence excerpt | Status |
|---|---|---|---|---|---|
| C1 | 首个以 co-evolution 为中心组织轴的 agentic systems survey;既有 self-evolving surveys 只把它当 subtheme | sota-novelty | Sec. 1; Appendix B | ”Surveys on self-evolving agents (Gao et al., 2026; Xiang et al., 2026) are closest to our scope, but treat co-evolution as a subtheme” | source-verified |
| C2 | Stage 1 定义:A^(t+1)=Ω(A^t,E,τ^t),E 固定,耦合在 agent collective 内部 | causal-mechanism | Sec. 2.3; Sec. 3 | ”A^{t+1}=Ω(A^t, E, τ^t)”; “mutual adaptation among evolving peers within a fixed environment” | source-verified |
| C3 | Stage 2 定义:(A^(t+1),E^(t+1))=Ω(A^t,E^t,τ^t),环境进入更新,分 task/feedback/interaction 三子空间 | causal-mechanism | Sec. 2.3; Sec. 4 | ”(A^{t+1},E^{t+1})=Ω(A^t, E^t, τ^t)”; “extends the process to the environment itself” | source-verified |
| C4 | Stage 3 定义:Ω^(t+1)=Γ^t(S^t,Ω^t,τ^t),演化机制由自生成 revision process 修订 | causal-mechanism | Sec. 2.3; Sec. 5 | ”revises its evolution mechanism through a self-generated revision process Γ^t: Ω^{t+1}=Γ^t(S^t,Ω^t,τ^t)“ | source-verified |
| C5 | co-evolution 判别标准:≥2 个 evolving units 相互重塑彼此后续演化;Appendix A 与十个相邻概念划界 | comparison | Sec. 2.2; Appendix A | ”at least two evolving units jointly adapt and continually reshape each other’s further evolution, rather than merely exchanging information or interacting” | source-verified |
| C6 | Appendix B 对比对象含 Gao et al. 2026、Xiang et al. 2026、Guo et al. 2024 等,差异为中心轴不同 | comparison | Appendix B, Table 1 | ”we use co-evolution as the central inclusion criterion” | source-verified |
| C7 | 归类:AlphaStar → multi-source adversarial;POET/ADR → executable world construction;WebEvolver → model-based world construction | comparison | Sec. 3.1.2; 4.3.1; 4.3.2 | ”AlphaStar…league training”; “POET…evolves a population of environment-agent pairs”; “WebEvolver…co-trains a web world model with the agent policy” | source-verified |
| C8 | PromptBreeder/MemEvolve 被归为 single-entity precursors;RQGM 被作为更接近 true meta co-evolution 的例子 | comparison | Sec. 5.2 | ”Most related methods remain single-entity precursors… RQGM (Iacob et al., 2026) goes beyond by co-evolving task agents and evaluators” | source-verified |
| C9 | Figure 4 汇总跨论文证据(Panel A/B matched 对比、Panel C 归一化轨迹),构建方法见 Appendix C | benchmark-setting | Figure 4; Appendix C | ”Cross-paper evidence for the effect, consistency, and convergence of co-evolution. See Appendix C for data selection and aggregation details.” | source-verified |
| C10 | 论文未提供本 survey 自己的 centralized GitHub repo / paper-list 链接 | license-code | abs 页 + 全文检索 | —(verifier 全文与 abs 页检索均无本 survey 的 repo 链接) | source-verified |
Strengths & Weaknesses
亮点:
- 以”演化算子 Ω 作用域扩张”为轴的渐进 taxonomy,比按组件罗列的 self-evolution survey 更有判别力——它自带明确的 inclusion criterion(mutual adaptive pressure),Appendix A 的十项概念划界直接回应了该领域 self-play / open-endedness / meta-evolution 术语混用的问题。
- Figure 4 的 cross-paper evidence 汇总(static-vs-evolving matched 对比 + 归一化轨迹)在 survey 中少见,是”survey 也应给证据”的好实践。
- 对 Stage 3 保持诚实:作者明确承认现有工作绝大多数只是 single-entity precursor,真正的 meta co-evolution 例证极薄。
局限:
- Stage 3 实质上只有 RQGM 一个像样例子,“三阶段”中的第三阶段更接近 aspiration 而非成熟文献簇;taxonomy 的渐进叙事隐含”演化自由更大 = 更接近终极形态”的方向性预设,而 Sec. 6 的 safety 讨论自己也承认更大自由带来失控风险——两者之间的张力文中未正面处理(推测性评价)。
- Figure 4 的跨论文汇总受各论文 benchmark、setting 差异限制,Appendix C 的 data selection 本身可能引入幸存者偏差(作者未报告失败/负面案例的纳入比例;推测)。
- 三条挑战(evaluation / scaling / safety)停在方向性建议,未给出可操作的评测协议或 governance 机制设计。
Mind Map
mindmap root((CoEvolutionSurvey)) Problem 单实体 self-evolution 受 static context 上界 既有 survey 把 co-evolution 当 subtheme Method 形式化 S=(A,E), 算子 Ω Stage 1 Agent-Agent Adversarial / Collaborative / Organizations Stage 2 Agent-Environment Task / Feedback / Interaction space Stage 3 Meta Ω 自身可演化, Γ revision Results Figure 4 cross-paper evidence Appendix B 对比既有 survey 挑战: evaluation / scaling / safety
Notes
- 与 vault 已有 2507-SelfEvolvingAgentsSurvey、2508-SelfEvolvingAIAgentsSurvey 正交互补:那两篇以单 agent 内部组件为轴,本篇以多组件耦合为轴;三篇合看可覆盖 self-evolution 文献的两种切法。
- Stage 2 的 task-space / interaction-space 分类与近期消化的 environment-evolution 工作(EnvACE、EvoHarnessRL 一类)直接相关,可作为该方向的定位地图;GUI agent 的环境/harness 共演问题在本文框架下属于 Stage 2 interaction-space + harness 演化的交叉,文中覆盖较少,或是空白点。