Meituan-led SkillRise beats agent RL baselines by up to 8.5 pts

A quiet result out of a Zhejiang University, National University of Singapore, Shanghai Jiao Tong University and Meituan collaboration is worth a slow read, because it points at where the interesting work on agent training is heading. The SkillRise paper on Hugging Face argues that treating each agent task as an independent episode is leaving obvious wins on the table, and reports Pass@1 success rates of 85.9% on ALFWorld, 84.4% on WebShop and 54.6% on ScienceWorld with a Qwen3-4B backbone. Those numbers beat the strongest RL baseline the authors tested, GiGPO, by 2.3, 7.1 and 8.5 percentage points respectively.

The mechanic is worth understanding. Instead of discarding what an agent learned solving one instance, SkillRise groups related instances into a progressively harder sequence and lets a single policy alternate between solving the current task and curating an evolving skill document that gets passed straight to the next task. Credit assignment is split along that seam: the solving step is graded by the current task's outcome, and the curation step is graded by a discounted return over the tasks that follow. No external skill bank, no separate retrieval module, no second teacher model.

That structural choice pays off in ways that matter for anyone thinking about running agent pipelines in production. On ALFWorld the authors report that SkillRise matches an external-memory system called RetroAgent while running at roughly one sixth of RetroAgent's end-to-end time, and outperforms a related SkillRL pipeline by 12.5 percentage points at 4.3x lower runtime. They also see cross-task test-time scaling: Pass@1 climbs from 83.6% with sequences of two related tasks to 87.5% with sequences of six, even though each task is attempted only once.

Take the specifics as reported, not settled. The evaluation covers three text-only interactive benchmarks with 128 held-out instances each, and the backbones tested are Qwen3-1.7B and Qwen3-4B, so whether cross-task curation still helps at frontier scale or in noisy real tool-use is not answered here. What the paper also does not show is how the method degrades when the tasks in a sequence are less obviously related than the curated ALFWorld and WebShop families the authors constructed.

The operator takeaway is the direction of travel. If simply reshaping the RL loop to reward carryover between related tasks can add points on top of GRPO-style baselines without a heavier pipeline, teams shipping agentic products should probably ask whether their own evaluation harness even measures cross-task skill transfer today. The code is on GitHub if you want to poke at it.