Can an AI get better by trying out versions of itself and keeping what works, instead of proving each change is safe?
Do evolutionary archives let agents improve themselves without formal proof?
This explores whether an agent can improve itself by keeping an evolving archive of its own variants and testing them empirically, instead of formally proving that each self-edit is an improvement.
This explores whether an agent can get better by keeping an archive of its own variants and testing them, rather than proving each change is an improvement. The corpus's clearest answer is yes, with a catch. The Darwin Gödel Machine drops the requirement to formally prove that a self-edit helps. It runs the edited agent on a benchmark and keeps the variant if it scores better. Every variant goes into an evolutionary archive, and the process reached 2.5× on SWE-bench and 2.2× on Polyglot. Along the way it discovered better code editing and context management Can AI systems improve themselves through trial and error?. Almost no self-edit can be proven good in advance, but trying one is cheap.
The catch is that dropping proof doesn't drop outside grounding. One note argues that pure self-improvement is circular. It stalls on the gap between generating and verifying, on collapsing diversity, and on reward hacking. The methods that work all smuggle in an external anchor, such as past model versions, third-party judges, user corrections, or tool feedback Can models reliably improve themselves without external feedback?. Read that way, the benchmark is the Darwin Gödel Machine's anchor. The archive swaps a proof for an outside scoreboard, not for the agent's own confidence.
That raises the question of what happens when no scoreboard exists. The Red Queen Gödel Machine puts the evaluator inside the improvement loop, so it evolves alongside the agent. This lets agents improve at writing and proof generation without a static verifier, and it matches fixed-evaluator performance with fewer tokens Can evaluators improve alongside the agents they score?. SAGA goes further and evolves the goal itself, turning natural-language objectives into scoring functions the search can use immediately Can agents evolve their own objectives during search?. A survey frames this as three stages of shedding human engineering: first the peers, then the environment and feedback, and finally the evolution mechanism Can agents evolve beyond the constraints humans engineer?.
These archives mostly evolve the code around the model, not its weights. Self-improving agents split into a slow loop that updates weights and a fast loop that updates prompts, memory, and tools. Recent progress sits in the fast loop because scaffold edits are cheaper and reversible Do self-improving agents really split into two distinct loops?. That suits an archive, since a bad variant costs little and can be dropped. It also seems to generalize. AIDE85 was evolved through seven accepted rewrites in 8 days, and it equals or beats the human-built AIDE on four held-out benchmarks, including out-of-distribution ones Does automated evolution match human-built agent performance?.
Archives aren't the only proof-free route. Reflexion stores verbal self-diagnoses in episodic memory and improves across episodes. It works because clear success or failure feedback stops the agent from rationalizing Can agents learn from failure without updating their weights?. AgentFly reached 87.88% on GAIA validation using memory operations alone Can agents learn continuously from experience without updating weights?. VOYAGER grows a library of executable skills Can agents learn new skills without forgetting old ones?. Those refine one agent's memory or skills, while an archive keeps many whole agents alive to branch from. Self-editing doesn't automatically scale with model strength either. Models produce useful harness edits about equally well at every tier, but the ability to benefit from them peaks in mid-tier models. Weak models fail to invoke the harness, and strong ones struggle to follow it faithfully Do stronger models always evolve harnesses better?. So proof isn't what makes self-improvement work. What matters is who runs the scoreboard.
Sources 11 notes
DGM replaces formal proofs with empirical benchmarking and maintains an evolutionary archive of agent variants, achieving 2.5× improvement on SWE-bench and 2.2× on Polyglot by discovering capabilities like better code editing and context management.
Pure self-improvement stalls due to the generation-verification gap, diversity collapse, and reward hacking. Reliable improvement methods succeed by smuggling in external anchors: past model versions, third-party judges, user corrections, or tool feedback.
Red Queen Gödel Machine makes evaluation part of the improvement loop, allowing agents to optimize writing and proof generation without a static verifier. Co-evolved systems match fixed-evaluator performance while using fewer tokens, suggesting shared learning drives efficiency.
SAGA's bi-level architecture closes a feedback loop from optimization results back to goal design by having an outer LLM loop propose new objectives and compile them into code the inner loop can immediately use, enabling objective formulation as part of discovery rather than a fixed input.
A survey framework organizes co-evolving systems into three stages that progressively remove human engineering: dynamic peers first, then adaptive environments and feedback, finally the evolution mechanism itself. Single-entity self-improvement stalls in static contexts; co-evolution supplies adaptive pressure across multiple components.
Show all 11 sources
A survey framework organizes self-improving agents into two update mechanisms: slow parametric loops updating foundation model weights, and fast non-parametric loops updating prompts, memory, and tools. Recent progress concentrates in the fast loop because scaffold updates are cheaper and reversible than weight updates.
AIDE85, evolved through seven accepted rewrites in 8 days, equals or surpasses AIDEhuman on four held-out benchmarks spanning in- and out-of-distribution tasks including weather forecasting. The result shows automated design iteration can match human-driven R&D on generalization.
Reflexion demonstrates that unambiguous environmental feedback (success/failure) enables agents to write useful self-diagnoses and improve across episodes without parameter updates. The binary signal prevents rationalization, and keeping reflections uncompressed preserves their usability.
AgentFly formalizes agent learning as a Memory-augmented MDP with three memory modules (case, subtask, tool) that enable credit assignment and policy improvement entirely through memory operations. The approach achieved 87.88% on GAIA validation without modifying LLM parameters.
VOYAGER demonstrates that storing executable skills in an embedding-indexed library and composing complex skills from simpler ones allows agents to learn continuously while avoiding the forgetting that occurs with weight-update-based methods. Environmental feedback refines skills while an automatic curriculum drives continual exploration.
Model capability to produce useful harness edits stays constant across tiers, but capacity to actually benefit from those edits follows an inverted U-shape, peaking in mid-tier models. Weak models fail to invoke harnesses; strong models struggle with faithful instruction-following.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Self-Improvements in Modern Agentic Systems: A Survey
- The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators
- Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement
- Co-Evolution in Agentic Systems: Toward Self-Directed Evolution Beyond Human Design
- Hyperagents
- DarwinX: Evolving Agent Harnesses Through Natural Selection
- Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents
- Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops