Research map / Label-efficient and robust learning / Knowledge distillation

SEAD: Competence-Aware On-Policy Distillation via Entropy-Guided Supervision

NeurIPS 2026 · Label-efficient and robust learning › Knowledge distillation · approach: Rethinking distillation objectives · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SEAD, an entropy-guided on-policy distillation method that skips or reweights tokens by joint teacher-student entropy, anneals forward-to-reverse KL, and uses a competence-gated easy-to-hard prompt curriculum, improving math reasoning of a 7B OLMo student.

Problem

On-policy distillation applies supervision uniformly across tokens, training phases and prompts, although teacher supervision quality depends on the student's competence, wasting compute and producing noisy gradients.

Approach

Uses entropy as a unified probe of student competence at three scales: joint teacher-student entropy zones (skip, RKL, FKL), a cosine FKL-to-RKL annealing schedule, and a competence-gated prompt curriculum, which are claimed to be mutually necessary.

Key results

• On OLMo 7B student with 32B teacher, full SEAD reaches 64.0 average vs 59.2 for vanilla OPD (RKL) and 58.2 for the student across six math benchmarks.
• Token Zones combined with KL annealing reaches 63.7 average, a +4.5 jump over either component alone per the paper.
• On Nemotron 8B/49B, SEAD component variants reach at most 69.5 average, below GRPO's 69.3 only marginally and not clearly above vanilla OPD's 68.8 gain margin.

Surprising findings

• On the Nemotron pair, the best SEAD variant (69.5) barely exceeds GRPO (69.3), and the full SEAD method was not reported there.
• Token selection alone (Token Zones) hurts AIME25 relative to vanilla OPD in the ablation, with gains depending on the curriculum.

Limitations

• Curriculum relies on static difficulty scores from the initial student rather than adaptive re-estimation.
• Evaluated only on mathematical reasoning and two model families; long reasoning chains not fully investigated.
• Theoretical assumptions (modulated PL condition, low-entropy agreement) are stated rather than empirically verified, and some proof steps are informal.

Experimental setup

Datasets: DAPO-Math-17K

Benchmarks: MATH-500 Minerva-Math OlympiadBench AMC23 AIME24 AIME25

Metrics: Pass@1 accuracy Pass@32 average pass rate average accuracy

Baselines: GRPO OPD (RKL k=1) OPSD EOPD TIP

Builds on: GKD GRPO EOPD TIP DAPO-Math-17K

Closest papers by meaning

• Your Teacher Can’t Help You Here: Combating Supervision Fidelity Decay in On-Policy Distillation (NeurIPS 2026 · similarity 0.92)
• The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes (NeurIPS 2026 · similarity 0.92)
• When Are Teacher Tokens Reliable? Position-Weighted On-Policy Self-Distillation for Reasoning (NeurIPS 2026 · similarity 0.91)
• Entropy-Aware On-Policy Distillation of Language Models (ICML 2026 · similarity 0.90)
• SimCT: Recovering Lost Supervision for Cross-Tokenizer On-Policy Distillation (NeurIPS 2026 · similarity 0.90)
• Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.89)