Related papers: Thinking at the Right Size: Amortized Distillation Across Post-Trained LLMs

Thinking at the Right Size: Amortized Distillation Across Post-Trained LLMs

URL: http://arxiv.org/abs/2608.22854v1
Date: Mon, 24 Aug 2026 06:36:32 GMT
Title: Thinking at the Right Size: Amortized Distillation Across Post-Trained LLMs
Abstract summary: We introduce ADAPT---Amortized Distillation Across PostTrained LLMs---a framework for amortizing distillation across both axes of a model family.<n>First, a two-phase distillation procedure constructs post-trained students through pre-training alignment and supervised fine-tuning distillation.<n>The resulting continuum of models also enables adaptive model-size selection at inference time.
Score: 36.805962131025694
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Abstract: Practical deployment of large language models (LLMs) requires families of post-trained variants---instruction-tuned, reasoning-tuned, and chat-style models---each at multiple sizes to meet diverse latency and memory budgets. Producing each (variant, size) pair independently is prohibitive, so model families typically span only a handful of coarse-grained sizes per post-trained variant. Boomerang distillation (Kangaslahti et al., 2026) reduces this cost along the size axis for base models. Through model size interpolation, it constructs models of intermediate sizes from a single teacher-student pair without additional training. However, it still treats each post-trained variant as a separate object of optimization. We introduce ADAPT---Amortized Distillation Across Post-Trained LLMs---a framework for amortizing distillation across both axes of a model family: size and post-training variant, producing $L \times K$ models for $L$ interpolated sizes across $K$ post-trained variants with a single distillation run. ADAPT combines two components. First, a two-phase distillation procedure constructs post-trained students through pre-training alignment and supervised fine-tuning distillation, enabling smooth size--performance interpolation on generation and reasoning tasks. Second, weight-delta initialization approximates this construction across post-trained variants by transferring the distillation-induced weight change from the base model to students initialized from different post-trained variants. The resulting continuum of interpolated models also enables adaptive model-size selection at inference time, improving the compute--accuracy trade-off for long-form reasoning tasks.

Related papers

Reverse Distillation: Consistently Scaling Protein Language Model Representations [0.6066775336291516]
We introduce Reverse Distillation, a principled framework that decomposes large PLM representations into subspaces guided by smaller models of the same family.<n>A motivating intuition is that smaller models, constrained by capacity, preferentially encode broadly-shared protein features.<n>On ProteinGym benchmarks, reverse-distilled ESM-2 variants outperform their respective baselines at the same embedding dimensionality, with the reverse-distilled 15 billion parameter model achieving the strongest performance.
arXiv Detail & Related papers (2026-03-08T16:24:05Z)
Nemotron Elastic: Towards Efficient Many-in-One Reasoning LLMs [80.72350166388601]
Nemotron Elastic is a framework for building reasoning-oriented LLMs.<n>It embeds nested submodels within a single parent model.<n>Each of these submodels shares weights with the parent model and can be extracted zero-shot during deployment.
arXiv Detail & Related papers (2025-11-20T18:59:21Z)
Boomerang Distillation Enables Zero-Shot Model Size Interpolation [30.74896485280105]
Existing approaches build model families by training each size independently, which is expensive and provides only coarse-grained size options.<n>In this work, we identify a novel phenomenon that we call boomerang distillation.<n>One first distills down to a small student and then progressively reconstructs intermediate-sized models by re-incorporating blocks of teacher layers into the student.
arXiv Detail & Related papers (2025-10-06T17:41:20Z)
FINE: Factorizing Knowledge for Initialization of Variable-sized Diffusion Models [35.40065954148091]
FINE is a method based on the Learngene framework to initializing downstream networks leveraging pre-trained models. It decomposes pre-trained knowledge into the product of matrices (i.e., $U$, $Sigma$, and $V$), where $U$ and $V$ are shared across network blocks as learngenes'' It consistently outperforms direct pre-training, particularly for smaller models, achieving state-of-the-art results across variable model sizes.
arXiv Detail & Related papers (2024-09-28T08:57:17Z)
POA: Pre-training Once for Models of All Sizes [33.72644336390202]
We propose a novel tri-branch self-supervised training framework, termed as POA (Pre-training Once for All) Our approach introduces an innovative elastic student branch into a modern self-distillation paradigm. It achieves state-of-the-art performance using ViT, Swin Transformer and ResNet backbones.
arXiv Detail & Related papers (2024-08-02T06:13:29Z)
MAP: Low-compute Model Merging with Amortized Pareto Fronts via Quadratic Approximation [80.47072100963017]
We introduce a novel and low-compute algorithm, Model Merging with Amortized Pareto Front (MAP)<n>MAP efficiently identifies a set of scaling coefficients for merging multiple models, reflecting the trade-offs involved.<n>We also introduce Bayesian MAP for scenarios with a relatively low number of tasks and Nested MAP for situations with a high number of tasks, further reducing the computational cost of evaluation.
arXiv Detail & Related papers (2024-06-11T17:55:25Z)
Reusing Pretrained Models by Multi-linear Operators for Efficient Training [65.64075958382034]
Training large models from scratch usually costs a substantial amount of resources. Recent studies such as bert2BERT and LiGO have reused small pretrained models to initialize a large model. We propose a method that linearly correlates each weight of the target model to all the weights of the pretrained model.
arXiv Detail & Related papers (2023-10-16T06:16:47Z)
MatFormer: Nested Transformer for Elastic Inference [91.45687988953435]
MatFormer is a novel Transformer architecture designed to provide elastic inference across diverse deployment constraints.<n>MatFormer achieves this by incorporating a nested Feed Forward Network (FFN) block structure within a standard Transformer model.<n>We show that a 850M decoder-only MatFormer language model (MatLM) allows us to extract multiple smaller models spanning from 582M to 850M parameters.
arXiv Detail & Related papers (2023-10-11T17:57:14Z)
Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning [52.29522018586365]
We study structured pruning as an effective means to develop smaller LLMs from pre-trained, larger models. Our approach employs two key techniques: (1) targeted structured pruning, which prunes a larger model to a specified target shape by removing layers, heads, and intermediate and hidden dimensions in an end-to-end manner, and (2) dynamic batch loading, which dynamically updates the composition of sampled data in each training batch based on varying losses across different domains.
arXiv Detail & Related papers (2023-10-10T15:13:30Z)
Training Trajectories of Language Models Across Scales [99.38721327771208]
Scaling up language models has led to unprecedented performance gains. How do language models of different sizes learn during pre-training? Why do larger language models demonstrate more desirable behaviors?
arXiv Detail & Related papers (2022-12-19T19:16:29Z)
bert2BERT: Towards Reusable Pretrained Language Models [51.078081486422896]
We propose bert2BERT, which can effectively transfer the knowledge of an existing smaller pre-trained model to a large model. bert2BERT saves about 45% and 47% computational cost of pre-training BERT_BASE and GPT_BASE by reusing the models of almost their half sizes.
arXiv Detail & Related papers (2021-10-14T04:05:25Z)

This list is automatically generated from the titles and abstracts of the papers in this site.