ZJU-REAL

ZJU-REAL / SkillZero

Public

Official code for "SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization"

agent curriculum-learning in-context-reinforcement-learning rl skill
48
1
100% credibility
Found Apr 03, 2026 at 48 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

SkillZero is a research codebase for training AI agents to internalize skills through in-context reinforcement learning in simulated household and search environments.

How It Works

1
📖 Discover SkillZero

You stumble upon an exciting research paper about teaching AI helpers to master everyday skills like a human would, and find the free code to try it yourself.

2
🛠️ Prepare your setup

You quickly set up your computer with the simple tools needed to run the training, feeling ready to experiment.

3
🌍 Load game worlds

You download fun virtual worlds where your AI can practice tasks like picking up objects or searching for info.

4
🚀 Start training

With one command, you launch the training and watch your AI helper learn skills step by step, getting smarter right before your eyes.

5
📈 Track improvements

You check charts showing your agent nailing more challenges, like a student acing tests after practice.

🏆 AI masters skills

Your trained helper now tackles tough tasks effortlessly, proving it internalized real skills for the real world.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 48 to 48 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is SkillZero?

SkillZero is a Python framework for in-context agentic reinforcement learning that helps LLMs internalize skills from demonstrations, boosting performance on benchmarks like ALFWorld household tasks and Search-QA retrieval. Users get ready-to-run training scripts for 3B-parameter models, integrating vLLM for fast inference and environments like ALFWorld via simple bash commands such as `bash scripts/train_alfworld_skillzero_3b.sh`. As the official code from the SKILL0 paper, it solves the challenge of getting agents to reuse learned skills without fine-tuning.

Why is it gaining traction?

It stands out by delivering big wins over vanilla RL baselines—up to substantial metric gains on ALFWorld and Search—using agentic in-context prompts that let models plan and execute like experts. Developers dig the plug-and-play setup: conda env, pip installs for flash-attn and alfworld, then train or merge checkpoints effortlessly. Low stars (48) but official GitHub repository status draws RL folks chasing reproducible agentic code.

Who should use this?

RL researchers replicating embodied AI papers or tweaking ALFWorld agents for pick/place/heat tasks. LLM devs building retrieval-augmented agents for Search-like envs, especially those needing curriculum skill management without rebuilding from scratch. Avoid if you're new to RL setups requiring WandB login and custom retriever servers.

Verdict

Grab it if you're in agentic RL and want the official code baseline—solid README guides install-to-train, but 1.0% credibility and 48 stars signal early-stage maturity with room for more tests/docs. Worth forking for ALFWorld experiments.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.