EnvLunch trains agents from environments without written tasks
Original title: EnvLunch: The Environment as a Free Lunch for Learning to Interact
The label-free warm-up beat the baseline by over 9 points on average, and matched it with 8 to 32 times fewer annotations.
EnvLunch is a warm-up for agentic data-analysis training that uses an executable environment instead of human-written or model-generated task text. It samples valid action trajectories, runs them, and trains the agent to recover a functionally equivalent trajectory from only the final observation, before downstream reinforcement learning. The paper says this outperforms the baseline by over 9 points on average across training budgets. With few labels, the baseline needed 8 to 32 times as many annotated examples to match EnvLunch at its smallest budget. Environment-defined tasks beat model-written tasks on the same environments, and the warm-up also helped fine-tuning beyond data analysis.
- No human or model-written task descriptions are used in the warm-up.
- The agent is trained to reconstruct a trajectory from the final observation only.
- Environment-defined tasks outperformed model-written tasks on the same environments.
- The method is positioned as a starting point when annotated agent tasks are scarce.
From the original article
1 min read at huggingface.co
For agentic data analysis, language-model agents often require a large number of annotated tasks to learn how to interact effectively with executable environments (i. e. , databases).
By Wei Liu, Min Gao, Yali Du · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.