If an AI keeps rewriting itself, do the gains keep coming, or does it stall after a few rounds?
How many acceptable rewrites can recursive self-improvement sustain before returns diminish?
This explores whether recursive self-improvement, where an AI rewrites itself over and over, keeps paying off or flattens out, and what the corpus says about how many rewrites the gains can last.
This explores whether recursive self-improvement, where an AI rewrites itself over and over, keeps paying off or flattens out, and what the corpus says about how many rewrites the gains can last. The direct answer is that nobody has measured it yet. The one paper that ran a seven-rewrite streak over 8 days reported that the improvements transferred, but gave neither the size of each gain nor when it landed. Without those score curves you can't tell a steady climb from a single big win followed by six small ones, so the evidence doesn't support a number Does recursive self-improvement sustain gains or hit diminishing returns?.
The nearest evidence is indirect, and it points to plateaus. In looped language models, two loops give broad gains, while three or more regress: the second pass does the useful refining, and later passes oscillate with less diversity in their internal representations. That is a sweet spot rather than a slope Does adding more loops always improve looped language models?. It isn't the same as an agent rewriting its own code, but it shows more iteration can make things worse. A parallel result appears in harness evolution: models produce useful edits equally well at every capability tier, but the ability to benefit from them peaks in mid-tier models. Weak models don't invoke the harness, and strong ones struggle to follow it faithfully Do stronger models always evolve harnesses better?.
A better question than "how many rewrites" is "what stops the decline." Pure self-improvement stalls for structural reasons: the model is worse at checking its work than at generating it, its outputs lose diversity, and it learns to game its own reward. The methods that work all bring in something from outside, such as earlier model versions, third-party judges, user corrections, or tool feedback Can models reliably improve themselves without external feedback?. The systems with the most impressive runs follow this pattern. The Darwin Gödel Machine replaces proofs with real benchmarks and keeps an archive of agent variants, which gave 2.5× on SWE-bench Can AI systems improve themselves through trial and error?. SkillOpt limits how much each edit can change, checks edits on held-out data, and keeps rejected edits as negative examples, which beat letting agents rewrite freely Does constraining edits make skill learning more stable?. In those cases the ceiling depends more on the quality of the external checks than on how many rewrites happen.
Where the gains come from also affects how long they last. Most recent progress is in the fast loop, meaning prompts, memory, and tools, because those edits are cheap and reversible. Updating model weights is the slow loop Do self-improving agents really split into two distinct loops?. A cheap, reversible edit is easy to test and roll back, which may be why bounded scaffold rewrites look sustainable. There is also a case for optimism. A bilevel setup where an outer loop reads and rewrites the inner loop's search code found new mechanisms and got a 5x improvement on GPT pretraining Can an AI system improve its own search methods automatically?. The argument that recursive self-improvement could offset diminishing returns on R&D spending is still a proposal, not a measured curve Can recursive self-improvement speed up the research process itself?.
So the corpus supports three points: seven rewrites is a real data point but not a measured curve, the plateaus we can see elsewhere are sharp rather than gradual, and how long gains last seems to depend on external anchors and bounded edits. If you want to test how long a self-improvement loop lasts, look for trajectories across many rounds, and treat a single headline multiplier with caution.
Sources 9 notes
The paper reports seven successive improvements in an 8-day run but provides neither the magnitude of each gain nor their timing. Without score trajectories and longer-horizon data, the evidence supports only that improvements transferred, not that recursive self-improvement sustains returns against diminishing curves.
LoopCoder-v2 shows that two loops deliver broad gains over baseline, but three or more loops regress. Loop 2 carries the productive refinement; later loops oscillate with reduced representational diversity rather than converging toward better performance.
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.
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.
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.
Show all 9 sources
SkillOpt's ablations show that adding a textual learning-rate budget, held-out validation gate, and rejected-edit buffer (retaining failed edits as negative feedback) produces more stable and generalizable skill improvement than allowing agents to freely rewrite their own instructions.
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.
An outer loop successfully read inner loop code, identified bottlenecks, and generated new Python mechanisms at runtime, discovering combinatorial optimization and bandit methods that broke the inner loop's deterministic patterns and improved performance on GPT pretraining by 5x.
The paper argues that AI agents automating R&D improve product efficiency while research process efficiency stays fixed. Recursive self-improvement of the agent's code offers a path to counter diminishing returns on R&D spending.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops
- Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement
- Hyperagents
- Dream-RSI: Recursive Self-Improvement through Evolving Worlds
- PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents
- Self-Improvements in Modern Agentic Systems: A Survey
- The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators
- Recursive self-improvement of AI research agents