Is an AI that rewrites its own code really different from one where a second AI rewrites the first?
Does AIDE2's single loop differ from bilevel autoresearch's nested loops?
This explores whether AIDE2, an agent that rewrites itself, is structurally a different kind of self-improvement from bilevel autoresearch, where an outer loop rewrites the search machinery of an inner research loop.
This explores whether AIDE2, an agent that rewrites itself, is structurally a different kind of self-improvement from bilevel autoresearch, where an outer loop rewrites the search machinery of an inner research loop. The corpus has no head-to-head comparison, so what follows is a reading across the two. The main difference is who improves whom. In AIDE2, the agent proposes changes to its own code, benchmarks the modified versions, and keeps the ones that score best. Each accepted rewrite aide2-proposes-changes-to-its-own-code-benchmarks-the-modified-versions-on-ai-rd|becomes the agent for the next iteration, so the improver and the thing improved are the same object. Bilevel autoresearch splits them. An inner loop does the research work, and an outer loop bilevel-autoresearch-enables-meta-optimization-where-an-outer-loop-autonomously|reads the inner loop's code, spots its bottlenecks, and writes new Python mechanisms at runtime.
The bilevel split has a concrete payoff. The outer loop stands outside the inner loop's habits, so it can see patterns the inner loop can't. It found that the inner loop's deterministic search was stuck, and it introduced combinatorial optimization and bandit methods to break that pattern. That produced a 5x improvement on GPT pretraining. AIDE2's single loop compounds instead: seven accepted rewrites in eight days, each building on the last. Its guard against fooling itself is a different one. Candidates are scored on evaluations aide2-proposes-changes-to-its-own-code-benchmarks-the-modified-versions-on-ai-rd|the proposing agent cannot see, and the result is checked on aide2s-gains-generalize-to-four-held-out-benchmarks-including-physics-based-weat|four held-out benchmarks, including physics-based weather forecasting, which sits outside the tasks used for selection.
The two loops end up attacking related problems, even though the structures differ. AIDE2's rewrites aide2s-accepted-rewrites-concentrate-on-problems-practitioners-face-building-age|clustered around search plateaus, context budgets, and untrustworthy wins. Getting stuck in search is what bilevel's outer loop was built to break. Context budgets are a known failure elsewhere in the corpus: long-horizon-research-tasks-require-limiting-reasoning-steps-per-turn-not-just-t|unrestricted reasoning in one turn eats the context that later search rounds need. In both designs the useful edits are to the scaffolding around the model, not to the model. autonomous-research-pipelines-discover-ai-architectures-beyond-automls-reach-bec|AUTORESEARCHCLAW's 411% F1 gain shows the same thing. Bug fixes, architecture changes, and prompt engineering each beat all hyperparameter tuning combined, because the agent can read code and reason about how the parts interact.
The evidence is not equally strong on both sides. Bilevel's headline number is a large gain on one task, so we don't yet know how far it travels. AIDE2's gains are smaller per step, but they are tested against hidden and held-out evaluations, which is the stronger check against optimizing for the scoring rather than the goal. One caution on vocabulary: neither design is a looped language model, where 'loop' means rerunning layers inside one network. There the picture is different, with looping-has-a-sweet-spot-not-a-slope-the-second-loop-carries-the-refinement-and|a sweet spot at two loops and regression at three or more. Whether either self-improvement design hits a similar ceiling is not something these notes settle.
Sources 7 notes
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.
An autonomous research agent proposed changes to itself, benchmarked variants on AI R&D tasks, and kept rewrites scoring best on evaluations the proposing agent could not see. Each accepted rewrite became the agent for the next iteration.
The paper reports that AIDE2's improvements transfer to four held-out benchmarks spanning machine learning, algorithm engineering, and physics-based weather forecasting—the last being outside the selection task distribution. This demonstrates transferable gains beyond overfitting to the selection set.
The seven accepted code changes clustered around search plateaus, context budgets, and untrustworthy wins—problems practitioners face building effective agents. This convergence suggests the autonomous loop discovered genuinely useful improvements rather than optimizing for scoring artifacts.
Unrestricted reasoning within single search turns consumes context needed for subsequent retrieval rounds, degrading the agent's ability to incorporate new evidence. Setting per-turn reasoning budgets, not just overall time limits, prevents this context erosion and maintains search quality across iterations.
Show all 7 sources
AUTORESEARCHCLAW achieved 411% F1 improvement on LoCoMo through bug fixes, architectural changes, and prompt engineering—each individually exceeding all hyperparameter tuning combined. This demonstrates a categorical capability gap: autoresearch can read code and reason about system-level interactions; AutoML cannot.
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.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Recursive self-improvement of AI research agents
- Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents
- The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators
- Bilevel Autoresearch: Meta-Autoresearching Itself
- Dream-RSI: Recursive Self-Improvement through Evolving Worlds
- Self-Improvements in Modern Agentic Systems: A Survey
- PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents
- Automated Alignment Researchers: Using large language models to scale scalable oversight