Skip to main content
Back

Alibaba's Qwen2.5-Math-PRM-7B Catches Exactly Where Your Math AI Goes Wrong

Alibaba's Qwen team released a 7B process reward model that grades each step of a math solution and beats larger open-source PRMs.

  • Qwen released Qwen2.5-Math-PRM-7B, a 7B Process Reward Model scoring each reasoning step from 0 to 1.
  • Fine-tuned from Qwen2.5-Math-7B-Instruct with a scalar head replacing the LM head.
  • Beats open-source PRMs and GPT-4o on ProcessBench style error localization; still trails o1-mini.
  • Trained with consensus filtering on Monte Carlo step labels, detailed in arXiv:2501.07301.
  • Use via Transformers with \n\n step breaks and <extra_0> sentinel tokens for per-step rewards.
  • Ideal for Best-of-N verification, PRM-guided search, and reasoning RL in English and Chinese math.

Alibaba’s Qwen team has released Qwen2.5-Math-PRM-7B, a 7-billion-parameter process reward model that assigns a score to each step in a candidate math solution. Developers can use those scores to rank sampled answers, guide search, or provide step-level rewards during training.

The release joins a larger 72B process reward model and Qwen’s existing outcome reward model. The outcome model assigns one score to a complete solution. The process model exposes local signals that can identify where a derivation deteriorates, while a separate policy model handles generation.

A 7B verifier in one view

Property Details
Primary role Score individual steps in mathematical reasoning
Input A question and a marked, multi-step candidate solution
Output One positive-class score from 0 to 1 per marked step
Base model Qwen2.5-Math-7B-Instruct
Related models A 72B PRM and Qwen’s outcome reward model
License Downloadable weights under the license terms
Adoption The Hub showed 166,575 downloads in the preceding month.

One score for every step

A process reward model receives a problem and a completed reasoning chain, then emits a score at each marked boundary. Each value estimates the model’s confidence that the preceding step is acceptable in context. In Qwen’s worked example, scores of [1.00, 0.19, 0.98, 1.00] identify the second step as the likely failure point.

  • Best-of-N ranking: Generate N candidate solutions, score each chain, and select the highest-ranked answer.
  • Tree search: Score partial derivations so weak branches can be pruned before further decoding.
  • Reinforcement learning: Use intermediate scores as denser rewards than a single end-of-response grade.
  • Feedback systems: Flag a likely faulty step for a tutor, grader, or review interface.

Sentinels make the API work

Qwen’s interface depends on two delimiters: double newlines separate the written steps, and <extra_0> marks every position that requires a score. The custom model returns token-level logits, and the positive-class softmax value at each sentinel becomes the step reward.