Data science agents can automate end‑to‑end machine learning pipelines, yet they still lag behind human rankings. Typical process errors include insufficient exploration, unreliable validation, coding bugs, and wasteful reuse of the same exploratory analysis and baseline models under a fixed budget. AdaPRA introduces a persistent coding‑agent harness that accumulates knowledge across multiple attempts on the same task, coupled with rubric‑guided supervision that spots and fixes these errors. A process reward agent audits the agent's trajectory and workspace by running checks, and it adapts its rubrics to the specifics of each task. Experiments show that rubrics seeded from easier tasks transfer to harder ones, and further adaptation yields additional gains. With five attempts per task, AdaPRA lifts an 80B Qwen model’s leaderboard percentile by 24.5 points on DSPredict‑Easy and 14.8 points on DSPredict‑Hard while markedly cutting process errors. Moreover, rubrics mined from one base model’s runs improve another model (DeepSeek‑V3.2), and AdaPRA outperforms the competitive DataPRM system. The code is released for reproducibility.
Review