ResearchGym - Improving Replay Buffers: leaderboard
ResearchGym task on memory designs for experience replay in reinforcement learning on DeepMind Control Suite and Gym tasks. A run's score is divided by the source paper's reported result (100 = matches it) and the best of three runs counts; higher is better.
Metric: Best-of-3 score normalised to the task's reference (SOTA) result, SOTA = 100. Source: anikethh.github.io. Saturation forecast: Around August 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 Codex (xHigh) | 49.7 | #185 (GPT-5.2 Codex) |
| 2 | Claude Opus 4.5 | 21.7 | #152 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=researchgym-improving-replay-buffers · How It Works · Data refreshed daily, snapshot 2026-11-12.