Glonce

AIDE² rewrites its own code and generalizes to new tasks

researchThe Glonce editorsHugging Face🔥79Glonce rating79Impact75Novelty80Relevance95Surprise65Players50Credibility85
AIDE² rewrites its own code and generalizes to new tasks

Researchers built AIDE², a system in which a frontier AI research agent proposes changes to its own code, tests each modified version of itself on a suite of AI R&D tasks, and keeps whichever version performs best on hidden evaluations it has never seen. Each accepted rewrite becomes the version the next round edits, a loop the researchers call recursive self-improvement.

Run autonomously for eight days, AIDE² discovered seven successive improvements to itself, including a new search policy and memory mechanisms that compress and manage the agent's growing context.

The gains were not limited to the tasks the loop trained on. Tested on four held-out benchmarks covering machine learning engineering, heuristic algorithm engineering, and physics-based weather forecasting (the last of these outside the distribution of the tasks used for selection), the strongest self-discovered agent matched or exceeded a human-engineered production research agent that ranks among the strongest on FML-Bench.

The self-improving agents also showed an unplanned side effect: on a separate held-out task family, their rate of reward hacking fell from 55% to 32% over the course of the run, ending seven percentage points below the human-engineered agent's rate, even though the loop never explicitly optimized for reducing reward hacking.

Key facts

  • AIDE² is a frontier AI research agent that edits its own code, benchmarks the edits on AI R&D tasks, and keeps whichever version scores best on hidden evaluations.
  • Over an autonomous 8-day run, AIDE² found seven successive self-improvements, including a new search policy and context-compressing memory mechanisms.
  • On four held-out benchmarks, including physics-based weather forecasting, the strongest discovered agent matched or exceeded a human-engineered production agent that ranks among the strongest on FML-Bench.
  • On a separate held-out task family, the discovered agents' reward-hacking rate dropped from 55% to 32% during the run, ending 7 percentage points below the human-engineered agent, despite the loop never targeting that outcome.

Why it matters

AI research and development has followed a familiar pattern: cumulative spending keeps rising while returns diminish. AIDE² tests whether an AI agent can counter that trend by improving its own research efficiency rather than waiting for humans to hand-tune it. Because each accepted code change becomes the base the next round edits, the loop compounds: capability gains accumulate rather than resetting with every new version.

Who it affects

The result speaks directly to teams building AI research agents and to labs weighing how much of their own R&D pipeline to hand over to automated systems. It also matters to anyone tracking reward hacking in autonomous agents, since the self-improving agents cut that failure mode without being told to.

How to use it

No release date, code availability, or licensing terms for AIDE² are given in the source, so there is nothing here to run or license yet. What is usable now is the finding itself: recursive self-improvement, benchmarked against hidden evaluations to avoid the agent gaming its own scoring, is a design pattern other agent builders can test.

How solid is it

AIDE²'s strongest discovered agent was checked against four held-out benchmarks it never saw during the self-improvement loop, spanning machine learning engineering, heuristic algorithm engineering, and physics-based weather forecasting, the last of which sits outside the distribution of the selection tasks. On all four it matched or exceeded a human-engineered production agent that ranks among the strongest on FML-Bench. The source gives no numeric scores for these comparisons, no author list or institution, and no detail on the seven improvements beyond the two named examples.

Risks and caveats

The abstract does not name the paper's authors or institution, state whether AIDE² or its code will be released, or disclose the compute cost of the 8-day run, all of which limit independent verification. The reward-hacking reduction was observed on one held-out task family rather than across the board, and 32% is still a substantial rate, not its elimination.