AI Papers Reader

Personalized digests of latest AI research

View on GitHub

From Manuals to Muscle Memory: How SKILL0 Teaches AI to Internalize Expertise

In the current landscape of Artificial Intelligence, “agentic” models—AI designed to complete multi-step tasks like booking travel or cleaning a virtual house—often operate like a novice chef who must read a recipe line-by-line every single time they boil an egg. This approach, known as “skill augmentation,” involves retrieving instructions from a database and pasting them into the AI’s active memory (the “context window”) at runtime.

While effective, this method is fundamentally “rented” intelligence. It is slow, expensive due to the high “token tax” of processing long instructions, and prone to “retrieval noise”—where the AI accidentally pulls up a recipe for pancakes while trying to make an omelet.

Now, a team of researchers from Zhejiang University, Meituan, and Tsinghua University has unveiled SKILL0, a reinforcement learning framework that allows AI agents to move beyond reading manuals. Instead, SKILL0 helps agents “internalize” these skills directly into their neural parameters, turning external instructions into intuitive, autonomous behavior.

The Training Wheels Approach

The core innovation of SKILL0 is a process the researchers call In-Context Reinforcement Learning (ICRL) paired with a Dynamic Curriculum.

Think of it like teaching a child to ride a bike with training wheels that slowly retract as the child gains balance. During the initial phase of training, the AI is provided with full “skill files”—structured packages of knowledge that explain how to handle specific scenarios. For example, in a household simulator (ALFWorld), a skill might be “Systematic Exploration”: Search every drawer and cabinet once before revisiting any location.

As the AI practices, the SKILL0 framework monitors “helpfulness.” If the AI’s performance on a task remains high even when the instructions are partially hidden, the system concludes that the model has begun to “learn” the pattern. The Dynamic Curriculum then progressively withdraws the external text until the agent can complete the task in a “zero-shot” setting—relying entirely on its own internal weights rather than a prompt.

Efficiency Through Vision

One of the primary hurdles in training these agents is “context bloat.” As an agent interacts with its environment over dozens of turns, the history of its actions and the retrieved skills can become thousands of words long.

To solve this, SKILL0 introduces a “visual context rendering” mechanism. Instead of feeding the model a massive wall of text, it compresses the interaction history and skill descriptions into a compact visual representation—essentially a digital “snapshot” of the task’s state. This allows the model to “see” the relevant information at a fraction of the computational cost.

Real-World Results

The researchers tested SKILL0 across two demanding benchmarks: ALFWorld (robotic house-task simulation) and Search-QA (complex web searching).

The results were stark. In ALFWorld, SKILL0 achieved a success rate nearly 10% higher than standard reinforcement learning baselines. In Search-QA, it boosted performance by 6.6%. Perhaps most importantly, it did this while maintaining an incredibly lean “inference cost.” While traditional methods might require 2,000 tokens of text to guide a single step, SKILL0 operates on fewer than 500 tokens.

By shifting expertise from the prompt to the parameters, SKILL0 paves the way for a new generation of AI agents that are not just faster and cheaper, but truly more capable—possessing the “muscle memory” needed to navigate the digital and physical worlds without a handbook in hand.