A fixed training set assumes that an example has roughly stable value while the learner changes. Reasoning distillation makes that assumption particularly fragile. An advanced solution may initially be incomprehensible, become useful once basic techniques are learned and later add little because the student already solves that kind of problem.
BoundaryEvo follows this moving relationship. It first gives the student enough reasoning capability to make its performance informative, then selects questions near the student’s current boundary. The teacher must reliably solve them while the student has not mastered them. Difficulty is therefore defined relative to two models and a stage of training, rather than by a permanent label attached to the question.
Priming makes the boundary measurable
BoundaryEvo, posted October 9, first primes a small model with teacher trajectories, then repeatedly selects questions the student has not mastered but the teacher can reliably solve. The authors report gains across their reasoning evaluations with a limited training set. Read the research paper.
In Capability Priming, the method directly distils a small set of long teacher-generated reasoning trajectories without a separate model filtering them. This is intended to establish an initial competence before capability-aware selection begins. A nearly unprepared student may otherwise produce noisy evidence about which examples it can benefit from.
The second stage repeatedly updates the training questions as the student evolves. It is not simply a static collection of hard questions or a rule to choose every current failure. The selected problem also needs a reliable expert solution, making supervision quality part of the boundary definition.
Teacher correctness and student learnability are separate
A question the expert can solve and the student cannot is a sensible candidate, but the gap can have different causes. The student may lack a small technique, misunderstand the notation or be unable to follow the long reasoning trajectory. A correct teacher answer does not by itself establish that the student can absorb it efficiently.
BoundaryEvo’s priming stage is relevant because it gives later demonstrations a more capable recipient. The adaptive stage then follows the student’s observed mastery rather than relying on static difficulty labels. This should be understood as an empirical curriculum design, not proof that every selected question lies at an exactly optimal learning distance. For a practitioner, useful diagnostics include whether failures shrink after exposure and whether the curriculum keeps returning to the same unlearnable cases. Those checks connect the selection boundary to actual learning rather than simply to the current list of wrong answers.
The small final set does not measure the full search
With 1,000 training examples, the paper reports improvements over static and dynamic selection baselines across eleven reasoning benchmarks. For Qwen3.5-4B, reported accuracy rises from 62.9 to 75.3 on IMO-AnswerBench and from 26.8 to 46.0 on AMO-Bench.
These figures support data efficiency in the tested setup, while the cost of finding suitable questions and producing teacher trajectories remains relevant. A thousand retained examples can require many rejected candidates. Comparisons should therefore include generation, scoring and repeated student assessment, not merely the size of the final training file.
The frontier should move without chasing noise
An adaptive curriculum can overreact to a temporary error or misread a lucky answer as mastery. Repeated or sufficiently robust capability estimates are important if selection is to follow genuine learning rather than fluctuations. This is a deployment concern implied by the approach, not a claim that the paper demonstrated failure on it.
The wider contribution is a model-relative view of data value. A lesson is useful when the learner can absorb something new from trustworthy supervision. That view encourages revisiting the curriculum as the model changes, while keeping the teacher’s reliability and the cost of measuring progress in the same account.
Questions, news tips or corrections? Contact [email protected].
