Branching from checkpointed filesystem and context avoids restarting long coding rollouts from scratch. This yields more verified trajectories for training coding agents.
Key takeaways · AI-distilled
Restoring only the context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → and working directory misses environment state such as installed dependencies, so SPROUT snapshots both the filesystem and AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → context after every assistant turn.
Across several LLMs on long-horizon coding tasks, SPROUT's resume-and-branch rollouts improved on pass@k by up to 23.6% while using 32 to 52% of the tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition →, per the authors.
In RL training of Qwen3.8-27B, the authors report it beat standard GRPO by 13.9 points on three long-horizon coding benchmarks while cutting time per training step by 39.5%. The paper is marked work in progress.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
articleAutoEnvScaling: Automating the Data Flywheel with Terminal AgentsSimon Yu, Baolin Peng, Bo Liu, Zhengyan Shi, Zhaoyang Wang, Hao Cheng, Qianhui Wu, Wenlin Yao, Natasha Jaques, Weiyan Shi, Jianfeng Gao
articleMagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data SynthesisXiming Lu, Aviral Kumar, Jan Kautz, Yejin Choi, David Acuna
articleE2E-SWE: Benchmarking LLMs on Building Working Codebases from ScratchHantian Ding, Chloe Bi, Jiacheng Zhu, John Yang, Matt Deitke, Pengcheng Yin, Zijian Wang, Rui Hou