---
title: "SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.26784v1"
permalink: "https://feed.astralor.com/items/itm_a29e20673f899c7ce8272af1"
published: "2026-07-29T11:26:33.000Z"
score: 81
channel: "ai"
tags: ["ALFWorld","ScienceWorld","SkillRise","WebShop","Cross-task Skill Transfer","Agentic Reinforcement Learning"]
---

# SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution

## AI commentary

> This paper addresses a practical problem: how to enable agents to reuse skills across multiple related tasks rather than learning each from scratch. SkillRise's 'progressive sequence + single policy alternation' design is both simple and effective, with convincing Pass@1 gains and test-time scaling. Highly recommended for teams building multi-task agents or researchers working on skill transfer.

## At a glance

SkillRise is a unified reinforcement learning framework designed to enable LLM agents to learn reusable skills across related but distinct tasks. It organizes task instances into progressively challenging sequences and uses a single policy to alternate between task solving and curating an evolving skill document. Experiments on ALFWorld, WebShop, and ScienceWorld show that SkillRise achieves the strongest Pass@1 performance among compared methods, with gains of 2.3 to 8.5 percentage points over the strongest baseline. It also demonstrates test-time scaling across tasks and substantially reduces runtime overhead of multi-stage skill learning pipelines.

## Key points

- SkillRise organizes related tasks into progressive sequences and uses a single policy for both task solving and skill document curation.
- Decoupled cross-task credit assignment supervises solving with current outcome and curation with discounted future outcomes.
- Outperforms existing methods on ALFWorld, WebShop, and ScienceWorld by 2.3–8.5 percentage points in Pass@1.
- Performance improves with longer sequences of related tasks at test time, indicating cross-task skill transfer rather than repeated sampling.

## Full text

Large language model agents often encounter related yet distinct tasks that share reusable solution patterns. Yet standard agentic reinforcement learning treats tasks as independent episodes, while existing approaches to skill learning either focus on repeated attempts of one task or use pipelines with multiple stages that entangle extraction, retrieval, and execution. We introduce SkillRise, a unified reinforcement learning framework for learning skills across tasks. SkillRise organizes related instances into progressively challenging sequences and uses a single policy to alternate between task solving and curating an evolving skill document passed directly to the next task. Decoupled credit assignment across tasks supervises solving with the current task outcome and curation with discounted downstream outcomes. Experiments on ALFWorld, WebShop, and ScienceWorld show that SkillRise achieves the strongest Pass@1 performance among the compared methods, with gains over the strongest baseline ranging from 2.3 to 8.5 percentage points. Although trained across distinct tasks, its learned curation policy remains effective for repeated attempts on the same task. Further analysis reveals scaling at test time across tasks: performance improves with longer sequences of related tasks even when each task is attempted only once. This trend suggests that SkillRise reuses transferable skills across tasks rather than benefiting from repeated sampling of the same task. SkillRise further retains strong performance while substantially reducing the runtime overhead of skill learning pipelines with multiple stages. Together, these results provide a simple and efficient training paradigm for LLM agents to extract, refine, and reuse transferable skills across tasks.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_a29e20673f899c7ce8272af1