Opus 5, GPT 6-Astra Build Their Own Self-Improvement Loops
TL;DR
- Given Read, Write, Bash and a Gameplay tool with no prescribed loop, Claude Opus-5 improved PuzzleScript-t20 by 20%, GPT 6-Astra by 10%, GLM 5.3 by 6%.
- Older model generations (Opus-4.6, GPT 5.6-Sol, GLM 4.7) show no measurable self-improvement under the same setup, indicating the capability tracks with scale.
- Context tuned on 20-turn PuzzleScript transfers out of domain: +21% on ARC-AGI-L1, +0.15 SPL on MazeBench, and across model families.
Given Read, Write, Bash and a Gameplay tool plus a general self-improvement goal, Claude Opus-5 improves its own PuzzleScript-t20 score by 20%, GPT 6-Astra by 10%, and GLM 5.3 by 6%, all without any externally imposed learning loop. That is the finding of a new paper by Cansu Sancaktar of the University of Tübingen and Phillip Isola of MIT CSAIL. Older generations (Opus-4.6, GPT 5.6-Sol, GLM 4.7) show no measurable improvement under the same setup.
The agents are not just trying harder. The paper describes them A/B testing their own text artifacts, writing procedural level generators with BFS solvers, calibrating difficulty, and keeping validation suites to catch regressions. In one GPT 6-Astra session, the teacher writes a Sokoban level generator seeded at 42 and runs comparisons such as 'v6:swap,soko3,soko2,soko0 → scores 1,0,0,1' against a no-artifact baseline, inspects failure traces via Bash, then validates a revised plan on new seeds. Replay concentration across sessions peaks at a 20–60% observed win rate, an inverted-U the authors align with the Zone of Proximal Development construct from developmental psychology.
Context optimized on 20-turn PuzzleScript levels transfers out of domain: +21% on ARC-AGI-L1 (a different 2D engine with 64×64 representations), +0.15 SPL on MazeBench (a 100-turn 3D navigation benchmark), and across model families, with Opus-5 benefiting from context written by GPT 6-Astra and vice versa. Against GEPA, a prescribed optimization loop running on real in-distribution games under matched rollout budgets, Figure 3 shows Opus-5's self-organized learning posting comparable gains.
'Models are improving at self-improvement,' the authors write. 'The key implication is not necessarily that a single generation of agents effectively self-improves, but rather that this meta-capability improves across generations.' The testbed is gameplay (PuzzleScript, ARC-AGI-3 Level 1, MazeBench), not the production tasks where a self-directed curriculum would most reshape the economics of frontier labs.
Originally reported by huggingface.co
Read the original article →Original headline: Paper: Claude Opus 5 and GPT-6 Astra Can Design Their Own Self-Improvement Loops Without Instructions