A new paper proposes a smarter way to teach smaller AI models by having them learn from a larger "teacher" model's mistakes, not just its confidence.
Researchers describe SPOT (Sparse Probing and Outcome-calibrated Targets), a technique for on-policy distillation - the process of training a smaller student model on trajectories it generates itself while a larger teacher model supervises. Standard versions of this approach lean on the teacher's raw probability estimates, but the paper argues that misses cases where the teacher is uncertain across many plausible next steps, or where the student already handles a given step fine. SPOT instead spends a limited "probing budget" on the moments that matter most, tests teacher-suggested alternatives using a verifier that checks whether they actually lead to correct answers, and builds training targets that reward good outcomes while staying anchored to the teacher's overall distribution. The authors tested it across multiple student models and several reasoning benchmarks.
Distillation is how companies turn expensive frontier models into cheaper, faster ones without losing much capability, and small inefficiencies in that process compound at scale. By spending compute on genuinely ambiguous reasoning steps and checking outcomes rather than trusting teacher confidence blindly, SPOT points at a subtler failure mode in current training pipelines: models that sound confident but haven't actually learned to solve the problem.
The paper reports that SPOT improves reasoning performance overall, but it does not disclose specific benchmark score deltas, so how large the real-world gain is will depend on independent replication.