SEAD boosts AI teacher-student training by 4.8% accuracy with entropy

On-policy distillation (OPD) has long suffered from a core inefficiency: the quality of teacher supervision depends on student competence. Incoherent rollouts produce noisy gradients, while already-mastered tokens generate redundant ones. Existing methods apply uniform supervision regardless, wasting compute at three scales—tokens, training phases, and prompts. A new paper from researchers at University of Rochester, Amazon, and MIT-IBM Watson introduces SEAD (Competence-Aware On-Policy Distillation via Entropy-Guided Supervision), which uses entropy as a unified probe to dynamically adjust supervision. SEAD partitions tokens into zones using joint teacher-student entropy, skipping about 50% of tokens with zero gradient. It also employs a cosine schedule that anneals from forward KL divergence to reverse KL as student competence grows, and a competence-gated curriculum that introduces prompts from easy to hard. These components are symbiotically necessary: token selection requires coherent rollouts (curriculum), and annealing requires monotonic improvement (also from curriculum).

On OLMo-3 models ranging from 7B to 32B parameters, SEAD achieves +4.8 average accuracy improvement over vanilla OPD across six diverse math benchmarks (e.g., GSM8K, MATH, etc.). Ablation studies confirm that the three components interact super-additively—removing any one degrades performance significantly. The method is particularly impactful for large-scale distillation where compute budgets are tight. By skipping nearly half the tokens and tailoring the loss function to student proficiency, SEAD reduces training time and improves final accuracy. This aligns with a growing trend in AI training: moving from static, one-size-fits-all pipelines to adaptive, competence-aware systems. The code and models are not yet released, but the paper provides clear mathematical formulation and empirical validation.