Ripple-65B 預覽：美國無限制類別智能體編碼領域的最先進方法

Ripple-65B 预览：美国无限制类别代理编码领域的最先进方法

AI 導讀

Ripple-65B 以 1.4M rollouts 在 TerminalBench 2.1 取得 80.5 分，所用 rollouts 約為此前 SOTA Reflection Beam 的 1/70。團隊自行設計合成數據流程及混合 RL 配方，從具備良好初始 SFT 的檢查點逐步提升模型表現。

正文

Rather than investing in a brute-force RL climb, we carefully iterated on our SFT and RL recipes with a focus on compute efficiency. We opted not to buy any posttraining data, instead devising our own synthetic data pipelines, quality filtering, and coverage metrics.

Our mixed RL recipe uses

truncated importance sampling

score centering

, and

linear length penalty

with high-quality long-horizon synthetic environments. The RL tasks are tuned for our model capabilities, and we continuously added harder tasks as the model improved.

The end result was a steep RL climb from a strong initial SFT checkpoint. We reached a TerminalBench 2.1 score of 80.5 with only 1.4M rollouts, roughly 70x fewer rollouts than Reflection Beam, the prior SOTA.