Alibaba's Waypoint-OPD Looks at the Middle of Training to Lift Math and Code Scores
When transferring the skills of a well-trained AI to another model, the method compares not just the finished result of training but the teacher's state partway through as well. In the researchers' math and code benchmarks, it averaged higher scores than the existing Direct-OPD approach.
Researchers at Alibaba Cloud released a new AI training method called Waypoint-OPD on October 15, publishing both a paper and code. The technique transfers the gains from reinforcement learning, where a model builds skill by being rewarded for good answers, into a different model. The model doing the teaching is called the “teacher model,” and the one learning is the “student model.” Waypoint-OPD looks at how the teacher changed over the course of its own training, and uses that to adjust how much weight to give the information passed along to the student.
Why look at the middle of the teacher’s training?
The existing approach, Direct-OPD, compares only the teacher before and after reinforcement learning. It takes the answers the student produced and examines which phrasings the teacher came to prefer. But looking only at the start and the end makes it impossible to tell whether the preference flipped somewhere in between. The researchers’ concern is that it becomes hard to distinguish whether changes from early training and changes from late training reinforced each other, or whether some of them canceled each other out.
Waypoint-OPD adds one more point of comparison: the teacher’s state partway through training. For the answers the student produced, it compares all three teacher states together, before, midway, and after. If the early and late preferences shifted in the same direction, more weight goes to the final change; if they point in opposite directions, the weight is reduced. The tradeoff is that it requires computing the teacher model’s scores one extra time compared with the existing method.
How much did math and code scores improve?
The researchers measured math and code problem-solving ability using three Qwen3 student models of different sizes. Averaged across seven math benchmarks, Waypoint-OPD scored higher than Direct-OPD. The gains were 2.18 percentage points on Qwen3-8B, 1.59 percentage points on Qwen3-14B, and 1.89 percentage points on Qwen3-30B-A3B.
Averaged across five code benchmarks, the margins in the same order were 2.66, 1.52, and 1.19 percentage points. According to the researchers, Waypoint-OPD posted the highest average score among the training methods compared in all six combinations of the two domains and three models. On individual benchmarks, though, other training methods came out ahead in some cases. The experiments were limited to math and code problem solving with models in the Qwen3 family.
Who can try it right now?
This is a paper and codebase aimed at researchers and developers who train models themselves. Anyone can read it, but no already-trained model files are included. Actually running it requires preparing your own models and training environment.
For the runtime environment, the project points to a Linux setup with NVIDIA GPUs. The list of software needed for installation is tailored to the Ubuntu and CUDA environment the researchers were using. The researchers noted that they have not verified installation on a freshly built environment.
Waypoint-OPD is a method that feeds not just the final outcome of training but also the direction of mid-training changes into student model training. The researchers report that this approach raised average scores on math and code benchmarks. The open question is whether the same effect holds for models outside the Qwen3 family. Performance on Korean-language tasks, which matters to users in Korea, has also yet to be confirmed.