SKILL0: In-Context Agentic RL for Skill Internalization (April 2026)

Free

Progressively withdraws skill documentation during training until agents operate zero-shot — +9.7% on ALFWorld, +6.6% on Search-QA with <0.5k tokens per step; 133 HF likes

FreeFree tier

Inputs: text, codeOutputs: text

Type

Open Source

Visit Website GitHub (40 stars)

About SKILL0: In-Context Agentic RL for Skill Internalization (April 2026)

SKILL0 is a novel in-context reinforcement learning framework designed to internalize agent skills directly into model parameters, eliminating the need for runtime skill retrieval. It introduces a training-time curriculum that starts with full skill context and progressively withdraws it, using a Dynamic Curriculum that evaluates each skill file's on-policy helpfulness and retains only useful skills within a linearly decaying budget. Skills are grouped offline by category and rendered with interaction history into a compact visual context, teaching the model tool invocation and multi-turn task completion. Extensive experiments on ALFWorld, Search-QA, and WebShop demonstrate substantial improvements over standard RL baselines (+9.7%, +6.6%, +10.1% respectively) while maintaining highly efficient context (<0.5k tokens per step). The code is publicly available.

Key Features

Training-time curriculum with progressive withdrawal of skill documentation until zero-shot operation

Skill grouping by category and rendering with interaction history into compact visual context

Dynamic Curriculum that evaluates each skill file's on-policy helpfulness

Linearly decaying budget for retaining useful skills during training

Zero-shot autonomous behavior without runtime skill retrieval

Efficient context of fewer than 0.5k tokens per inference step

Publicly available code and implementation

Pros & Cons

Pros

Significant improvements over standard RL baselines: +9.7% on ALFWorld, +6.6% on Search-QA, +10.1% on WebShop
Highly efficient token usage (<0.5k tokens per step) reduces computational cost
Eliminates runtime retrieval noise and token overhead from skill augmentation
Enables true zero-shot autonomous behavior without external skill loading
Open source code available for reproduction and extension

Cons

Requires computationally expensive reinforcement learning training phase
Currently validated only on specific benchmarks (ALFWorld, Search-QA, WebShop); broader generalization unverified
Still a research framework; not yet a polished product for production use
Dependency on high-quality initial skill documentation and interaction history

Best For

Agentic tasks requiring tool invocation and multi-turn completionWeb-based tasks such as ALFWorld, Search-QA, and WebShopEnabling LLM agents to internalize procedural knowledge for zero-shot executionReducing token overhead and retrieval noise in skill-augmented agents

FAQ

What is SKILL0?

SKILL0 is an in-context reinforcement learning framework that internalizes agent skills into model parameters, enabling zero-shot autonomous behavior without runtime skill retrieval.

How does SKILL0 achieve zero-shot performance?

It uses a training-time curriculum that starts with full skill context and progressively withdraws it, using a Dynamic Curriculum that retains only helpful skills until the agent operates without any external skill documentation.

What benchmarks does SKILL0 improve upon?

SKILL0 achieves +9.7% on ALFWorld, +6.6% on Search-QA, and +10.1% on WebShop over standard RL baselines.

Is the code for SKILL0 available?

Yes, the code is publicly available at the URL provided in the paper.