Self-improving agents move past code: DGM-Hyperagents makes the improvement mechanism itself editable
Synopsis
Self-improving systems have relied on fixed, handcrafted meta-level mechanisms, which bound how fast they can improve. DGM-Hyperagents merges the task agent and the meta agent that modifies it into one editable program, so the improvement procedure itself can be rewritten. On paper review and robotics reward design it raises held-out test scores from the initial agent's 0.0 and 0.060 to 0.710 and 0.372, and it shows cross-domain transfer on Olympiad-level math grading.
Interpretation
A self-improving agent's improvement mechanism can now itself be rewritten, not just the task solutions it improves. Earlier recursive self-improvement systems, including the Darwin Gödel Machine (DGM), relied on a fixed, handcrafted meta-level mechanism to generate self-improvement instructions, so the base system could only improve within the boundaries set by the meta agent's design. On held-out paper-review test tasks, DGM-H raises the score from the initial agent's 0.0 to 0.710 (CI 0.590–0.750), above the open-sourced static baseline AI-Scientist-v2 at 0.630; on robotics reward design it rises from 0.060 to 0.372 (CI 0.355–0.436), above the default reward function that directly optimizes the evaluation metric at 0.348.
The approach writes the task agent and the meta agent into a single editable program, so the code that produces improvements falls inside the scope of modification. This process of making the improvement procedure itself modifiable is called metacognitive self-modification; DGM-H keeps the DGM's open-ended exploration structure, an archive of candidate agents from which parents are sampled probabilistically by performance, modified, evaluated and returned to the archive. Experiments span four domains, coding on Polyglot, paper review, robotics reward design and Olympiad-level math grading, with each method run 5 times and medians reported with 95% bootstrap confidence intervals; on Polyglot, DGM-H raises the training-subset score from 0.140 to 0.340 and the full benchmark from 0.084 to 0.267, comparable to the original DGM's reported 0.140 to 0.380 and 0.142 to 0.307.
Improvement ability learned in one domain transfers to another, and the benefit comes mainly from the transferred meta agent. Transfer agents from the original DGM can barely generate better task agents in a new domain, with an imp@50 of 0.0 (CI 0.0–0.010), close to the initial agent. On the held-out Olympiad-level math grading task, transfer hyperagents selected from the paper-review and robotics runs reach an imp@50 of 0.630 (CI 0.540–0.630) after 50 iterations; transferring only the meta agent gives 0.620 (CI 0.550–0.660), while transferring only the task agent gives 0.100 (CI 0.0–0.560).
Perspective
The results come from finite benchmark settings with fixed task distributions, frozen closed foundation models and fixed evaluation protocols, and parent selection stays fixed in the main experiments; all experiments ran under sandboxing, resource limits and human oversight, with test tasks hidden from the task and meta agents.
The authors state that the results show task-level and meta-level improvements and their transfer across domains, but do not yet establish unbounded or self-accelerating improvement; on math grading, the difference between DGM-H starting from transfer hyperagents and from the initial agent is not statistically significant (p>0.05).
