Every time a standard AI agent finishes a task, it forgets the experience. An agent that navigated a shopping workflow successfully today starts tomorrow's related task without a trace of what it learned. A research team from Zhejiang University, the National University of Singapore, Shanghai Jiao Tong University, and Chinese tech company Meituan has built a framework called SkillRise that changes this — and on July 29, 2026, they posted the results to arXiv.

SkillRise trains a single reinforcement learning policy to do two things at once: solve the current task, and compose a written skill document that gets handed directly to the next task in a sequence. Across three standard benchmarks — ALFWorld, WebShop, and ScienceWorld — the framework achieves the strongest Pass@1 results of any compared method, with gains over the best prior baselines ranging from 2.3 to 8.5 percentage points. In a field where single-point benchmark gains are considered meaningful, that margin is notable.

The paper appeared on HuggingFace Papers the following day, drawing immediate attention from the agentic AI community.

The Problem: Every Episode Begins at Zero

Standard approaches to training LLM-based agents treat each task as a fresh, isolated episode. The practical consequence is that an agent which has successfully solved a shopping task — learned which search strategies work, which product comparisons matter, which steps tend to lead to wrong purchases — carries none of that strategic knowledge into the next shopping task. Each encounter starts from zero.

Prior work has addressed this through two broad approaches. The first focuses on repeated attempts at a single task (multi-sample methods), using extra inference compute to improve performance through sampling. The second builds multi-stage skill-learning pipelines with separate components for extracting skills from experience, retrieving relevant skills for new tasks, and executing based on retrieved skills. Both approaches have documented weaknesses. Multi-sample methods scale compute cost with the number of attempts. Multi-stage pipelines introduce compounding errors: the component that extracts skills doesn't know what the retrieval component will look for; the retrieval component doesn't know how the execution component will apply what it finds. The stages are tuned independently, and their hand-offs are brittle, as the SkillRise paper's framing of existing approaches makes clear.

SkillRise takes a different path: collapse all three functions — extraction, refinement, and reuse — into a single policy trained end-to-end with a single reinforcement learning objective.

One Policy, Two Roles: How SkillRise Works

The core architecture is deliberately simple. SkillRise organizes related task instances into progressively challenging sequences. For each task in the sequence, the agent receives two inputs: the current task description, and the skill document that was curated during all prior tasks in the sequence. The agent then alternates between two roles:

In its solver role, the agent reads the task plus the inherited skill document and generates actions to complete the task. In its curator role, the agent reads the completed task's trajectory and produces an updated skill document — effectively writing a memo to its future self about what worked, what to avoid, and which strategies seem broadly applicable.

The updated skill document is passed directly to the next task in the sequence. The agent that wrote it is also the agent that will read it; there is no hand-off between separately trained components.

How the Training Signal Works Across Episodes

The most technically important element of SkillRise is not its architecture but its training signal: a mechanism the team calls decoupled credit assignment.

In standard reinforcement learning, the credit assignment problem asks: given a sequence of actions and a reward at the end, which actions "caused" the success? This is the intra-episode version of the problem — distributing credit across actions within a single task trajectory.

SkillRise has to solve a harder version: the inter-episode credit assignment problem. When an agent curates a skill document during task 3 in a sequence, and that curation decision helps the agent succeed at task 7, the training signal needs to recognize the connection across multiple distinct episodes. Standard RL reward signals cannot do this, because each episode is treated as independent.

Decoupled credit assignment handles this by training the two roles with different signals. The solver role is supervised by the outcome of the current task — the standard binary success signal. The curator role is supervised by discounted downstream outcomes: whether the skill document it produces actually helps the agent succeed at later tasks in the sequence. The discount applied to downstream outcomes is the mechanism that propagates the credit signal backward through the task chain, making curation decisions accountable to their long-term effects rather than their immediate plausibility.

The practical consequence: the agent is not rewarded for writing long or comprehensive skill summaries. It is rewarded for writing skill summaries that demonstrably improve its own future performance. This is a stricter and more meaningful training signal than any approach that evaluates skill quality in isolation.

Benchmark Results: Three Environments, Consistent Gains

The team evaluated SkillRise across three standard agentic benchmarks, each testing a different dimension of long-horizon reasoning:

ALFWorld is a text-based household environment where agents complete multi-step goals — placing objects in specified locations, cleaning or heating items, navigating between rooms — under partial observability. It is one of the most widely used evaluation environments for text-based household agents.

WebShop is a simulated e-commerce environment populated with over 1.18 million real-world products and crowd-sourced shopping instructions, where agents must search, compare, customize, and purchase products according to natural-language user goals.

ScienceWorld is an interactive text environment spanning topics at the level of an elementary-school science curriculum: thermodynamics, electrical circuits, chemistry, and biological processes.

Across all three, SkillRise achieved the best Pass@1 performance among compared methods, with gains over the strongest prior baseline ranging from 2.3 to 8.5 percentage points. In a research landscape where multi-percentage-point improvements on these benchmarks represent meaningful progress, the consistency of the gains across three structurally different environments is the result's most notable feature.

What Happens When Task Sequences Get Longer

One of the paper's more striking findings concerns what happens at inference time. The team found that performance improves with longer sequences of related tasks, even when each task is attempted only once. The authors call this property "scaling at test time across tasks."

This is a different kind of scaling than the more familiar compute-based variety, where models improve by generating more candidate solutions or running more steps per task. In SkillRise, the improvement comes from accumulated experience — each new related task enriches the skill document, making it more precisely targeted to the task distribution. Later tasks in the sequence benefit from a more developed and specific skill base, without any additional compute per task.

The finding matters because it suggests a new design principle for agent deployment: the more related tasks a SkillRise-trained agent encounters in a sequence, the better it performs — and no additional training is required to get this benefit. An agent deployed in a structured, related-task environment (a legal workflow, a data analysis pipeline, a standardized customer service queue) should improve its own performance simply by working through more tasks.

Is the Skill-Learning Capability General?

A natural concern about any cross-task framework: does the learned capability transfer, or is it narrowly fitted to sequential task chains?

The team directly tested whether the curation policy remained useful when applied to a single task attempted multiple times — the standard single-task RL setting. Despite being trained across distinct tasks in a sequence, the learned curation policy remained effective in this repeated single-task setting. The agent had learned a genuinely general skill-curation capability, not merely a strategy tuned to inter-task transfer.

Where SkillRise Sits in a Dense Research Landscape

The 2026 research landscape around agentic skill learning is dense. A succession of frameworks — SkillRL, Skill1, SkillOpt, SkillWeaver, SkillOS, SkillPyramid, SKILLC, and others — have each attacked pieces of the same problem: how to make LLM agents acquire and reuse transferable strategies from experience. Most of these approaches address one or two stages of what can be called the skill lifecycle — selecting skills, using them, or distilling new ones — while leaving the others to separate components.

SkillRise's claim is not a marginal improvement on the existing pipeline paradigm. It is a different paradigm: one policy, two roles, one training objective, one credit signal that spans the full sequence. The elimination of inter-stage hand-offs removes the brittleness that multi-stage pipelines inherit from their architecture. Whether that architectural simplicity holds at larger scale, in less structured task sequences, or in environments with noisier feedback signals is an open question — but the empirical case for the unified approach is now benchmarked against the alternatives.

What the Framework Does Not Yet Solve

The paper is direct about its limitations. All three benchmarks are text-based simulation environments. SkillRise has not been tested in visual agent settings (where agents must read a live screen), in real-world deployment environments (where feedback signals are noisier and less structured), or in domains where transferable strategies are harder to express as text.

The framework also assumes task sequences that are related and progressively challenging. It is not clear how the system behaves when task sequences are poorly ordered, or when the skill document grows long enough to crowd out task-relevant context in the agent's limited context window. The paper does not report performance on very long sequences where this pressure becomes significant.

These are constraints that matter for deployment. An engineering team considering SkillRise for a real workflow should note that the demonstrated gains are within text-based simulation; the path from simulation to production is the open research question.

Institutional Context

The research was conducted by a 16-person team with affiliations at Zhejiang University, the National University of Singapore, Shanghai Jiao Tong University, and Meituan. As with any research involving Chinese institutions, Meituan and the two Chinese universities are subject to China's National Intelligence Law (2017), which in Article 7 requires all organizations to "support, assist, and cooperate with national intelligence work." Legal scholars note that Article 7's scope is contested and may not require proactive data sharing beyond what other security laws already mandate — but the structural legal obligation exists. The SkillRise project itself involves no consumer data and no product; the paper is published as an open-access arXiv preprint under a Creative Commons license. The affiliation is a factual matter of record.

The corresponding author is Yongliang Shen of Zhejiang University. The full author list is available at arXiv:2607.26784.

Frequently Asked Questions

What makes SkillRise different from other agent skill-learning approaches?

Most prior approaches use multi-stage pipelines: one component extracts skills from experience, another retrieves relevant skills for new tasks, and a third executes using retrieved skills. These components are trained separately, and errors compound at the hand-offs between them. SkillRise trains a single policy to handle all three functions in one reinforcement learning objective, using decoupled credit assignment to ensure that the skill curation step is rewarded based on whether it actually helps future tasks — not just whether it looks reasonable in the moment.

What is "test-time scaling across tasks," and why does it matter?

It means SkillRise's performance improves the more related tasks an agent works through, even when each task is attempted only once. Unlike compute-based test-time scaling (which improves performance by running more steps per task), this is experience-based: a richer skill document accumulated over more tasks makes each subsequent task easier. For practitioners, the implication is that deploying a SkillRise-trained agent in a structured, related-task environment should produce self-improving performance over time — without additional training.

What are the real constraints on SkillRise's current results?

All three evaluated benchmarks — ALFWorld, WebShop, and ScienceWorld — are text-based simulation environments. SkillRise has not been tested in visual agent settings, in real-world deployments with noisy feedback, or in domains where transferable strategies are difficult to express as text. The framework also assumes progressively challenging, related task sequences; performance on arbitrarily ordered or unrelated task sequences is not reported. These are the constraints that matter most for engineering teams considering the framework for production use.

How does the inter-episode credit assignment problem differ from standard RL credit assignment?

Standard RL credit assignment asks which actions within a single task trajectory deserved credit for a delayed reward. Inter-episode credit assignment — the harder problem SkillRise addresses — asks which curation decisions, made several tasks earlier, contributed to success on the current task. SkillRise's decoupled credit assignment solves this by rewarding curation based on discounted downstream outcomes across the task sequence, not just the current task's result.