TEAL evolves reusable tools for data-lake coding agents without labels
Original title: Unsupervised Tool Evolution for Coding Agents in Data-Lake Analytics
At a 15-step budget, Sonnet-5 rose 9.3 points on KramaBench and 21.7 on LakeQA.
Coding agents redo data discovery and joins for every data-lake question. TEAL builds a tool and skill library from an Explorer and a Critic, keeping a change only if the Worker adopts it on a rerun and changed answers are supported by trajectory evidence. TEAL-Q supplies synthetic questions when real ones are missing. Libraries evolved on synthetic questions performed comparably to those evolved on real questions.
Unconfirmed: Scores are the authors' reported benchmark results; independent replication is not stated.
- TEAL improved every tested Worker on both benchmarks at budgets of 10 or 15 steps.
- A Sonnet-5 LakeQA library evolved at 20 steps scored 68.3 when deployed at 30.
- That beat the 59.6 baseline and a library evolved at 30 steps, which scored 64.2.
- A tighter step budget also yielded better feedback for evolution.
- Products
- TEALTEAL-QKramaBenchLakeQA
- AI models
- Sonnet-5
From the original article
1 min read at huggingface.co
Coding agents can answer analytical questions over a data lake by discovering, cleaning, and joining the relevant data. However, they discard this work after each question and redo it from scratch for the next one. Tool evolution, which automatically builds and refines reusable tools across tasks, is a natural fit for this repetition.
By Yuqing Yang, Yurong Liu, Deqing Fu · 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.