cs.AINov 11, 2026

Effective Teacher Support for Reasoning Distillation

Authors: Lavinia Ghita, Dhruv Desai

Organizations: NVIDIA

Abstract

Forward-KL language-model distillation normally computes dense teacher and student scores over the full vocabulary at every supervised position, making training and stored teacher targets expensive. We instead distill from the teacher's highest-probability vocabulary fraction, renormalizing the retained mass as we sweep support from 100% to 0.01%. Our main study distills Qwen3 4B -> 1.7B. We confirm the broader support-reduction finding on larger Qwen3 8B -> 4B and Nemotron Nano 12B -> 9B pairs and under answer-only supervision. We evaluate generation primarily on MMLU-Pro and secondarily on GPQA-Diamond. Across the Qwen3 settings, retaining 0.01% of the vocabulary causes no observed MMLU-Pro degradation relative to full-vocabulary distillation, although disagreement with the complete teacher distribution increases. Nemotron Nano matches full-vocabulary MMLU-Pro at 0.1%, but smaller fractions fail to emit the closing reasoning tag, showing that the reliable cutoff depends on the model family. At 0.01% support, caching the retained next-token scores from the teacher forward pass cuts training step time by 38%, while restricting student output computation to those tokens cuts peak memory by 24%; neither causes observed MMLU-Pro degradation.

Explore similar work

Dec 24, 2025

Distilling the Essence: Efficient Reasoning Distillation via Sequence Truncation

Distilling the capabilities from a large reasoning model (LRM) to a smaller student model often involves training on substantial amounts of reasoning data. However, knowledge distillation (KD) over lengthy sequences with prompt (P), chain-of-thought (CoT), and answer (A) sections makes the process computationally expensive. In this work, we investigate how the allocation of supervision across different sections (P, CoT, A) affects student performance. Our analysis shows that selective KD over only the CoT tokens can be effective when the prompt and answer information is encompassed by it. Building on this insight, we establish a truncation protocol to quantify computation-quality tradeoffs as a function of sequence length. We observe that beyond a specific length, longer training sequences provide marginal returns for downstream performance but require substantially higher memory and FLOPs. To this end, training on only the first 50%50\% of tokens of every training sequence can retain, on average, ≈91%\approx91\% of full-sequence performance on math benchmarks while reducing training time, memory usage, and FLOPs by about 50%50\% each. Codes are available at https://github.com/weiruichen01/distilling-the-essence.
Feb 18, 2026

Task-Specific Knowledge Distillation via Intermediate Probes

Knowledge distillation from large language models (LLMs) assumes that the teacher's output distribution is a high-quality training signal. On reasoning tasks, this assumption is frequently violated. A model's intermediate representations may encode the correct answer, yet this information is lost or distorted through the vocabulary projection, where prompt formatting and answer-token choices creates brittle, noisy outputs. We introduce \method{}, a distillation framework that bypasses this bottleneck by training lightweight probes on frozen teacher hidden states and using the probe's predictions, rather than output logits, as supervision for student training. This simple change yields consistent improvements across four reasoning benchmarks (AQuA-RAT, ARC Easy/Challenge, and MMLU), with gains most pronounced under limited data. Probes trained on intermediate representations provide cleaner labels than the teacher's own outputs, effectively denoising the distillation signal. \method{} requires no architectural changes to student or teacher, is architecture-agnostic, and adds minimal compute since probe training is cheap and teacher representations can be cached. By exploiting internal representations, \method{} enables practitioners to extract more value from large teacher models without additional training data or architectural complexity.
Nov 13, 2026

Folding Language Models into Themselves

Knowledge distillation usually trains a compact student from scratch, which needs a large amount of data even though the teacher's weights already encode most of what the student has to learn. In this work, we explore building the student directly from the teacher's weights. We propose Concertina, which folds every group of N adjacent teacher layers into one shared block that is executed N times. Each loop step carries a low-rank correction initialized from the layer it replaces, and a per-token gate mixes these corrections, to let every token recover the behavior of the original layers from a single shared block. We further design a distillation data recipe that balances general, mathematical, academic and targeted multiple-choice data, to recover all capabilities of the teacher together rather than one at the expense of another. On Qwen3-4B, a 2.27B Concertina model distilled on 22B tokens retains about 90% of the teacher on commonsense and mathematical reasoning and 84% on average across eight benchmarks. It outperforms same-size dense and layer-sharing baselines, as well as Qwen3-1.7B, which is trained on over a thousand times more data. These results suggest that a compact model need not be retrained at scale, since folding the teacher into itself and distilling it on the right data recovers most of its capabilities. Code is available at https://github.com/myboker/concertina.