Towards Effective Process Supervision in Mathematical Reasoning
GitHub Pages
Qwen released new AI models that check math reasoning step-by-step, not just final answers. They also built a benchmark to test how well models catch errors mid-solution.
Based on reporting by GitHub Pages — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Qwen just dropped two new tools aimed at a problem that's been quietly nagging at anyone who trusts LLMs with math: models often get the right answer while getting there through nonsense. The company released Qwen2.5-Math-PRM-7B and Qwen2.5-Math-PRM-72B, a pair of process reward models trained to grade each individual step of a solution rather than just the final number.
Alongside the models, Qwen open-sourced ProcessBench, a benchmark of 3,400 test cases built mostly from competition and Olympiad-level math problems. Each case comes with a human-annotated solution where experts marked exactly where things went wrong, if they went wrong at all. The task for a model is blunt: find the earliest broken step, or say the whole thing checks out. It works for both PRMs and general critic models that get prompted to review solutions line by line.
The numbers back up the release. In Best-of-N testing across seven benchmarks — GSM8K, MATH, Minerva Math, GaoKao 2023 En, OlympiadBench, College Math, and MMLU STEM — Qwen2.5-Math-PRM-7B beat the majority-voting baseline on every single one, averaging a 1.4% lift. On ProcessBench itself, the 7B model outperformed every other open-source model its size, and even edged past GPT-4o-0806, though OpenAI's o1-mini still holds an edge on error-spotting.
One detail buried in the results is more interesting than the headline scores: Qwen2.5-Math-RM-72B, which is technically an outcome reward model that only judges final answers, turned out to be surprisingly good at catching intermediate errors too — good enough to beat some dedicated PRMs. That's a small hint that reward signals trained on outcomes alone are picking up more structural understanding than expected, which matters for anyone designing the next generation of supervision techniques.
Qwen's own paper is candid about the limits here. The team flags that current approaches to building training data for PRMs have real gaps, and that leaning purely on response-level Best-of-N evaluation can mask biases in how well a PRM actually understands reasoning quality. ProcessBench exists specifically to expose that blind spot, forcing evaluation down to the step level where sloppy reasoning can't hide behind a lucky correct answer.
My take — AI-written commentary, not fact-checked reporting
This is the kind of release that doesn't trend on social media but actually matters — process supervision is the difference between an AI that reasons and one that guesses well. The finding that an outcome-only reward model can rival dedicated step-checkers is the real story here, and I'd bet it says more about what these models learn implicitly than anything in the leaderboard. Also worth noting: this is a fully open-source drop from a Chinese lab outperforming GPT-4o on error detection, which should put to rest any lingering assumption that frontier reasoning tools are a closed-model monopoly.
Read more about this at: GitHub Pages