Early Experience skill
A skill for the "Early Experience" method (ICML 2026) that helps an agent prepare reward-free training data for language agents.
- 103 ★ on the repository
- Last commit
- MIT
Minimal access
Why this level
- The skill holds a text guide only; running training and environments is up to the user.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/early-experienceDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skill holds a text guide only; running training and environments is up to the user.
This is third-party code. Review the repository files before installing.
What it does
Early Experience is the research repository from OSU NLP and NeoCognition for "Agent Learning via Early Experience". The method sits between imitation learning and RL: the agent collects its own data by proposing non-expert actions at expert states, and the resulting next states serve as a reward-free training signal. Two implementations: Implicit World Modeling (the model predicts the next state) and Self-Reflection (training on reflections comparing expert and alternative actions). For agents, `skill/` holds a guide with the methodology, method mapping and common pitfalls that a code agent uses as a workflow for any new environment. Ten environments (ALFWorld, AppWorld, BFCL v4, ScienceWorld, SearchQA, tau-bench, TextCraft, TravelPlanner, WebArena, WebShop) have notes in `envs/`, and ready SFT data is on Hugging Face.
Who it is for. Researchers and engineers who train language agents and want to port the Early Experience method to their own environment.
Good fit when
- Implicit World Modeling data must be prepared for a new environment.
- Self-Reflection reasoning must be gathered for reward-free training.
- The paper's numbers on WebShop or ALFWorld should be reproduced.
- An agent should draft a method integration plan from an environment description.
Not a fit when
- A ready software development tool is wanted rather than a research recipe.
- There is no environment in which actions can be executed.
- Reward-based RL already works and no replacement is needed.
Example request
Read skill/METHOD.md and propose how to generate Implicit World Modeling data for our environment, starting with the number of alternative actions.Limitations
This is research code and a guide, not an install package: porting the method to your own environment takes work that the agent does together with a human, and parameter values (in particular the number of alternative actions) depend on the environment. The results listed in the README were obtained in specific environments for specific models. Reproduction needs Hugging Face and compute for fine-tuning; details are in `envs/<env>/README.md`. The skill installs by copying the `skill/` folder into a project.
How to disable. Delete the copied `skill` folder from the project. Datasets and checkpoints downloaded from Hugging Face are removed separately.
FAQ
How do I add the skill?
Run `git clone https://github.com/OSU-NLP-Group/EarlyExperience.git && cp -r EarlyExperience/skill /path/to/your/project/` and point the agent at `SKILL.md`.
What are IWM and SR?
Implicit World Modeling trains the model to predict the next state; Self-Reflection adds reasoning that compares the expert with an alternative.
Which environments are covered?
Ten: ALFWorld, AppWorld, BFCL v4, ScienceWorld, SearchQA, tau-bench, TextCraft, TravelPlanner, WebArena and WebShop.
Where is the ready data?
On Hugging Face: the `osunlp/early-experience` dataset with expert, IWM and reflection data.
Is an environment reward needed?
No, the method targets environments without verifiable rewards or with long horizons.
Discussion
No questions or reviews yet.
Related
SkillsAI Research Skills
AI Research Skills Library
98 skills for ML research, from ideation and model training to inference, evaluation and paper writing
The agents-meet-rl skill for Claude Code: a handbook on RL training of AI agents, failure diagnosis and framework choice.
Self-learning
self-learning-skills
A meta-skill that saves a hard-won working procedure as a new skill or rule, so the next session starts already knowing it
A CLI for training agents: your agent works on a task, and a mentor model reviews the result and sends back remarks
A CLI installer and package manager for coding agent skills: one command installs skills into many runtimes and helps build your own
A meta-skill that splits a task into stages, picks a profile and runs multi-agent work in Codex and Claude Code.
README in short
The ICML 2026 paper repository with a guide for code agents, notes for ten environments and data on Hugging Face.