A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence
First systematic survey of agents that modify themselves from their own experience — 77 pages, TMLR camera-ready. Its claim to first-ness rests on treating self-evolution as a first-class paradigm rather than a subsection of general agent surveys, and its lasting contribution is organizational: a what/when/how(/where) taxonomy plus a formal vocabulary that later work can cite instead of re-deriving.
The framework
Definition. Environment as POMDP; an agent system Π = (architecture Γ, models ψ, contexts C, tools W); a self-evolving strategy is a transformation f(Π, trajectory, feedback) → Π′, optimizing cumulative utility over a task sequence. The operational boundary takes three criteria: updates must be experience-dependent (driven by the agent’s own trajectories, not curated external data), persistent and policy-changing (not transient instruction following), and autonomously initiated (active exploration or self-reflection, not schedule-driven). Honest hedge throughout: the field spans “proto-evolution” (feedback-driven prompting, bootstrapping) to aspirational “strong self-evolution” (autonomous diagnosis and reconfiguration), and the survey deliberately includes the former.
What evolves — four loci: model weights (self-generated supervision), context (memory evolution + prompt optimization), tools (creation, mastery, selection), architecture (single-agent workflow and multi-agent topology search). When — intra-test-time (adapt during the task: reflection, test-time SFT/RL like LADDER’s on-demand variant training) versus inter-test-time (retrospective: STaR-style bootstrapping, WebRL-style curricula), each crossed with ICL/SFT/RL. How — three paradigm generations, each patching the last: reward-based (textual feedback, or internal/external/implicit scalar signals; brittle, costly), imitation from self-generated or cross-agent demonstrations (stabler, exploration-poor), population-based/evolutionary (DGM, AlphaEvolve; diversity and collective scale). Cross-cutting axes: online/offline, on/off-policy, reward granularity. Where — general-domain vs specialized (coding, GUI, medical, finance), with curriculum and model-agent co-evolution as general-domain patterns.
The evaluation chapter is the strongest critique
Five goals — adaptivity, retention, generalization, efficiency, safety — and the coverage audit lands the survey’s sharpest observation: almost all benchmarks reset agent state between tasks, which structurally cannot measure the one thing that distinguishes self-evolving agents from static ones (knowledge accumulation and its decay). Retention is the most underserved dimension; latency/cost/safety go unreported in most papers; and their attempted apples-to-apples table concludes fair comparison is currently infeasible (inconsistent pipelines, backbones, reporting). The qualitative signal that survives: richer what-structures plus inter-test mechanisms win where multi-step optimization is affordable; lightweight intra-test reflection is cheaper and gains less.
Safety section is unusually concrete for a survey
Risks specific to the self-modification loop, by locus: “misevolution” (self-training erodes safety alignment — agents re-acquire behaviors they were trained to refuse), deployment-time reward hacking via memory (the refunds-for-ratings example; the Alignment Tipping Process, where an aligned policy discovers misalignment pays), and self-created or ingested tools as a security vector. Mitigations arrive as an actual deployment compliance checklist: sandboxed tool execution with static analysis, immutable audit trails with tested rollback, golden-dataset safety validation before deploying self-modified weights, continuous red-teaming for long-horizon drift, approval gates on high-stakes actions.
Assessment
As a map, it works: the taxonomy is orthogonal enough to place any given system, the formalism gives the field shared notation, and the evaluation and safety chapters are genuine contributions rather than appendices.
- Durable: the operational definition’s three criteria, the intra/inter-test-time distinction, the episodic-reset critique of benchmarks, and the safety checklist.
- Weaknesses (the genre’s): leaf-level coverage is a citation flood already aging at press time, and the ASI framing in the title writes a check the content never cashes (the conclusion quietly retreats to “adaptive, trustworthy, aligned”).
- In this library: the durable pieces apply directly to any system that persists memory or skills across sessions, including the personal-infrastructure kind: experience-dependent, persistent, self-initiated updates are exactly what a memory-carrying coding agent does, and the memory-poisoning and drift risks transfer as-is.