@kazani

Hyperagents: an AI system that not only improves at solving tasks, but also improves how it improves itself. Hyperagents are self-referential agents that combine a task agent and a meta agent in one editable program, allowing them to modify both how they solve tasks and how they generate future improvements. The authors call this metacognitive self-modification: learning not just to perform better, but to improve at improving. The main idea is to go beyond earlier recursive-improvement systems (Darwin Gödel Machine), which could improve task performance but still relied on a largely fixed, handcrafted self-improvement procedure. In DGM-H, the paper’s implementation, that meta-level procedure becomes editable too. The most interesting result is evidence that the system can discover general self-improvement strategies. One example is the autonomous invention of persistent memory: instead of merely logging scores, the hyperagent stores synthesized insights, causal hypotheses, and forward-looking plans, then consults them during later self-modification steps. This lets later generations build on earlier discoveries and avoid repeating mistakes. The paper also reports suggestive evidence of compounding self-improvements: improvements discovered in one run can be transferred to a new setting and continue accumulating. Paper: https://arxiv.org/abs/2603.19461 Code: https://github.com/facebookresearch/Hyperagents

0 reply

0 recast

2 reactions