Meta$^n$: Recursive Self-Improvement through Emergent Depth

One fixed improver rewrites the agent layer by layer, each time seeing everything below it.

0.331score

ARC-AGI-2 score, GPT-5.2 backbone

Benchmark built to resist memorized skills; baselines stay near zero.

How it works

Improvers stay shallow
Self-improving agents fix answers, and self-editing ones freeze part of their machinery, so real meta-depth stops near two.
Recurse on the input
A fixed prompt Ω reads all task traces plus earlier layers' code, then writes a new layer of guidance and helpers.
Wins on eight families
Beats OpenEvolve and Gödel Agent on every benchmark family. Layers take on distinct roles, like rollback, without being told.

Which comes out on top on CO-Bench validation score?

try it

Drag the stack depth and watch the average score per layer.

Stack depth2 layers

Average score peaks at depth 3, yet deeper layers still win rare tasks the archive keeps.Illustrative model, not the paper's data

6 terms

Ω (Omega) One fixed prompt that reads traces and code, then writes the next layer.Pause

Like a manager reviewing both the team's results and the memos earlier managers wrote, then adding a new memo, rather than redoing the work.

Where it breaks

Every 'manager' here is the same fixed prompt, not a new person.

From the paper