Multi-bit quantization networks enable flexible deployment of deep neural networks by supporting multiple precision levels within a single model. However, existing approaches suffer from significant training overhead as full-dataset updates are repeated for each supported bit-width, resulting in a cost that scales linearly with the number of precisions. Additionally, extra fine-tuning stages are often required to support additional or intermediate precision options, further compounding the overall training burden. To address this issue, we propose two techniques that greatly reduce the training overhead without compromising model utility: (i) Weight bias correction enables shared batch normalization and eliminates the need for fine-tuning by neutralizing quantization-induced bias across bit-widths and aligning activation distributions; and (ii) Bit-wise coreset sampling strategy allows each child model to train on a compact, informative subset selected via gradient-based importance scores by exploiting the implicit knowledge transfer phenomenon. Experiments on CIFAR-10/100, TinyImageNet, and ImageNet-1K with both ResNet and ViT architectures demonstrate that our method achieves competitive or superior accuracy while reducing training time up to 7.88 . Our code is released at this link.

In physics, especially in geosciences and climate sciences, the poor performance of neural networks (NN) when applied outside their training distribution or their trained dynamics poses a very strong limitation to their general applicability (Irrgang et al., 2021; Landsberg and Barnes, 2025). In these fields, physical relations such as laws, dependencies or sensitivities are commonly derived (or learned) under well observed conditions and are then applied to less observed conditions to gain knowledge about the latter. For example, results from lab or numerical model experiments are regularly applied to real world problems or observations (e.g., Mehta et al., 2025); knowledge from our Earth and our Solar System are transferred to other planets and other star systems (e.g., Kvorka et al., 2026); learned relations that are derived today are transferred to the distant past or to the future (e.g., Eyring et al., 2016; Wang et al., 2024; Koutsodendris et al., 2014).

We reduce the computational cost of Neural AutoML with transfer learning.

Multi-bit quantization networks enable flexible deployment of deep neural networks by supporting multiple precision levels within a single model. However, existing approaches suffer from significant training overhead as full-dataset updates are repeated for each supported bit-width, resulting in a cost that scales linearly with the number of precisions. Additionally, extra fine-tuning stages are often required to support additional or intermediate precision options, further compounding the overall training burden. To address this issue, we propose two techniques that greatly reduce the training overhead without compromising model utility: (i) Weight bias correction enables shared batch normalization and eliminates the need for fine-tuning by neutralizing quantization-induced bias across bit-widths and aligning activation distributions; and (ii) Bit-wise coreset sampling strategy allows each child model to train on a compact, informative subset selected via gradient-based importance scores by exploiting the implicit knowledge transfer phenomenon. Experiments on CIFAR-10/100, TinyImageNet, and ImageNet-1K with both ResNet and ViT architectures demonstrate that our method achieves competitive or superior accuracy while reducing training time up to 7.88x. Our code is released at https://github.com/a2jinhee/EMQNet_jk.

A prevailing trend in neural network research suggests that model performance improves with increasing depth and capacity - often at the cost of integrability and efficiency. In this paper, we propose a strategy to optimize small, deployable models by enhancing their capabilities through structured knowledge inheritance. We introduce Hereditary Knowledge Transfer (HKT), a biologically inspired framework for modular and selective transfer of task-relevant features from a larger, pretrained parent network to a smaller child model. Unlike standard knowledge distillation, which enforces uniform imitation of teacher outputs, HKT draws inspiration from biological inheritance mechanisms - such as memory RNA transfer in planarians - to guide a multi-stage process of feature transfer. Neural network blocks are treated as functional carriers, and knowledge is transmitted through three biologically motivated components: Extraction, Transfer, and Mixture (ETM). A novel Genetic Attention (GA) mechanism governs the integration of inherited and native representations, ensuring both alignment and selectivity. We evaluate HKT across diverse vision tasks, including optical flow (Sintel, KITTI), image classification (CIFAR-10), and semantic segmentation (LiTS), demonstrating that it significantly improves child model performance while preserving its compactness. The results show that HKT consistently outperforms conventional distillation approaches, offering a general-purpose, interpretable, and scalable solution for deploying high-performance neural networks in resource-constrained environments.

We formalize the notion of a pseudo-ensemble, a (possibly infinite) collection of child models spawned from a parent model by perturbing it according to some noise process. E.g., dropout [9] in a deep neural network trains a pseudo-ensemble of child subnetworks generated by randomly masking nodes in the parent network. We examine the relationship of pseudo-ensembles, which involve perturbation in model-space, to standard ensemble methods and existing notions of robustness, which focus on perturbation in observation-space. We present a novel regularizer based on making the behavior of a pseudo-ensemble robust with respect to the noise process generating it. In the fully-supervised setting, our regularizer matches the performance of dropout. But, unlike dropout, our regularizer naturally extends to the semi-supervised setting, where it produces state-of-the-art results. We provide a case study in which we transform the Recursive Neural Tensor Network of [19] into a pseudo-ensemble, which significantly improves its performance on a real-world sentiment analysis benchmark.

Large language models (LLMs) have demonstrated remarkable capabilities, but their adoption is limited by high computational costs during inference. While increasing parameter counts enhances accuracy, it also widens the gap between state-of-the-art capabilities and practical deployability. We present Puzzle, a framework to accelerate LLM inference on specific hardware while preserving their capabilities. Through an innovative application of neural architecture search (NAS) at an unprecedented scale, Puzzle systematically optimizes models with tens of billions of parameters under hardware constraints. Our approach utilizes blockwise local knowledge distillation (BLD) for parallel architecture exploration and employs mixed-integer programming for precise constraint optimization. We demonstrate the real-world impact of our framework through Llama-3.1-Nemotron-51B-Instruct (Nemotron-51B), a publicly available model derived from Llama-3.1-70B-Instruct. Nemotron-51B achieves a 2.17x inference throughput speedup, fitting on a single NVIDIA H100 GPU while preserving 98.4% of the original model's capabilities. Nemotron-51B currently stands as the most accurate language model capable of inference on a single GPU with large batch sizes. Remarkably, this transformation required just 45B training tokens, compared to over 15T tokens used for the 70B model it was derived from. This establishes a new paradigm where powerful models can be optimized for efficient deployment with only negligible compromise of their capabilities, demonstrating that inference performance, not parameter count alone, should guide model selection. With the release of Nemotron-51B and the presentation of the Puzzle framework, we provide practitioners immediate access to state-of-the-art language modeling capabilities at significantly reduced computational costs.

We formalize the notion of a pseudo-ensemble, a (possibly infinite) collection of child models spawned from a parent model by perturbing it according to some noise process. E.g., dropout [9] in a deep neural network trains a pseudo-ensemble of child subnetworks generated by randomly masking nodes in the parent network. We examine the relationship of pseudo-ensembles, which involve perturbation in model-space, to standard ensemble methods and existing notions of robustness, which focus on perturbation in observation-space. We present a novel regularizer based on making the behavior of a pseudo-ensemble robust with respect to the noise process generating it. In the fully-supervised setting, our regularizer matches the performance of dropout. But, unlike dropout, our regularizer naturally extends to the semi-supervised setting, where it produces state-of-the-art results. We provide a case study in which we transform the Recursive Neural Tensor Network of [19] into a pseudo-ensemble, which significantly improves its performance on a real-world sentiment analysis benchmark.

Transfer learning is a simple and powerful method that can be used to boost model performance of low-resource neural machine translation (NMT). Existing transfer learning methods for NMT are static, which simply transfer knowledge from a parent model to a child model once via parameter initialization. In this paper, we propose a novel transfer learning method for NMT, namely ConsistTL, which can continuously transfer knowledge from the parent model during the training of the child model. Specifically, for each training instance of the child model, ConsistTL constructs the semantically-equivalent instance for the parent model and encourages prediction consistency between the parent and child for this instance, which is equivalent to the child model learning each instance under the guidance of the parent model. Experimental results on five low-resource NMT tasks demonstrate that ConsistTL results in significant improvements over strong transfer learning baselines, with a gain up to 1.7 BLEU over the existing back-translation model on the widely-used WMT17 Turkish-English benchmark. Further analysis reveals that ConsistTL can improve the inference calibration of the child model. Code and scripts are freely available at https://github.com/NLP2CT/ConsistTL.