A More Reliable Data Analysis Agent

Process Reward Models(PRMs) worked well in static domains like mathematics, but proved weak in tasks like data analysis agents where the environment keeps changing. Problems emerged where they missed silent errors that produce wrong results without interpreter exceptions, and even misjudged necessary exploratory actions as grounding failures.

DataPRM, which addresses this, operates as an active verifier that directly interacts with the environment. It's designed to actively inspect intermediate execution states to find hidden errors, and to properly evaluate segments that require trial and error.

• A reflection-aware ternary reward distinguishes between correctable grounding errors and unrecoverable mistakes.
• Step-level annotation combining diversity-based trajectory generation with knowledge augmentation built 8K+ training instances.

The results were clear. In Best-of-N inference, it boosted performance by 7.21% on ScienceAgentBench and 11.28% on DABStep, and even at a 4B model size, it outperformed strong baselines.

It also maintained generalization across various test-time scaling strategies, and when applied to RL, it produced better results than outcome-reward-based approaches. Ultimately, it achieved 78.73% on DABench and 64.84% on TableBench, and the code has been released in the DataMind repository.