NeurIPS 2025 Papers

Modern neural networks exhibit state-of-the-art performance on many existing benchmarks, but their high computational requirements and energy usage cause researchers to explore more efficient solutions for real-world deployment. Differentiable logic gate networks (DLGNs) learns a large network of logic gates for efficient image classification. However, learning a network that can solve simple problems like CIFAR-10 or CIFAR-100 can take days to weeks to train. Even then, almost half of the neurons remains unused, causing a \emph{discretization gap}. This discretization gap hinders real-world deployment of DLGNs, as the performance drop between training and inference negatively impacts accuracy. We inject Gumbel noise with a straight-through estimator during training to significantly speed up training, improve neuron utilization, and decrease the discretization gap. We theoretically show that this results from implicit Hessian regularization, which improves the convergence properties of DLGNs. We train networks $4.5 \times$ faster in wall-clock time, reduce the discretization gap by 98\%, and reduce the number of unused gates by 100\%.

Instruction-Tuning (IT) was recently found the impressive data efficiency in post-training large language models (LLMs). While the pursuit of efficiency predominantly focuses on sequence-level curation, often overlooking the nuanced impact of critical tokens and the inherent risks of token noise and biases. Drawing inspiration from bi-level coreset selection, our work provides the principled view of the motivation behind selecting instructions' responses. It leads to our approach Quadratic Coreset Selection (QCS) that reconciles sequence-level and token-level influence contributions, deriving more expressive LLMs with established theoretical result. Despite the original QCS framework challenged by prohibitive computation from inverted LLM-scale Hessian matrices, we overcome this barrier by proposing a novel QCS probabilistic variant, which relaxes the original formulation through re-parameterized densities. This innovative solver is efficiently learned using hierarchical policy gradients without requiring back-propagation, achieving provable convergence and certified asymptotic equivalence to the original objective. Our experiments demonstrate QCS's superior sequence-level data efficiency and reveal how strategically leveraging token-level influence elevates the performance ceiling of data-efficient IT. Furthermore, QCS's adaptability is showcased through its successes in regular IT and challenging targeted IT scenarios, particularly in the cases of free-form complex instruction-following and CoT reasoning. They underscore QCS's potential for a wide array of versatile post-training applications.

Efficient LLM pre-training requires well-tuned hyperparameters (HPs), including learning rate η and weight decay λ. We study scaling laws for HPs: formulas for how to scale HPs as we scale model size N, dataset size D, and batch size B. Recent work suggests the AdamW timescale, τ = B/(ηλD), should remain constant across training settings, and we verify the implication that optimal λ scales linearly with B, for a fixed N and D. However, as N and D scale, we show optimal τ obeys a precise power law in the tokens-per-parameter ratio, D/N. This law thus provides a method to accurately predict λopt in advance of large-scale training. We also study scaling laws for optimal batch size Bopt (the B enabling lowest loss at a given N,D) and critical batch size Bcrit (the B beyond which further data parallelism becomes ineffective). In contrast to prior work, we find both Bopt and Bcrit scale as power laws in D, independent of model size, N. Finally, we analyze how these findings inform the real-world selection of Pareto-optimal N and D under dual training time and compute objectives.

As large language models (LLMs) increasingly underpin technological advancements, the privacy of their training data emerges as a critical concern. Differential Privacy (DP) serves as a rigorous mechanism to protect this data, yet its integration via Differentially Private Stochastic Gradient Descent (DP-SGD) introduces substantial challenges, primarily due to the complexities of per-sample gradient clipping. Current explicit methods, such as Opacus, necessitate extensive storage for per-sample gradients, significantly inflating memory requirements. Conversely, implicit methods like GhostClip reduce storage needs by recalculating gradients multiple times, which leads to inefficiencies due to redundant computations. This paper introduces FlashDP, an innovative cache-friendly per-layer DP-SGD that consolidates necessary operations into a single task, calculating gradients only once in a fused manner. This approach not only diminishes memory movement by up to 50\% but also cuts down redundant computations by 20\%, compared to previous methods. Consequently, FlashDP does not increase memory demands and achieves a 90\% throughput compared to the Non-DP method on a four-A100 system during the pre-training of the Llama-13B model, while maintaining parity with standard per-layer clipped DP-SGD in terms of accuracy. These advancements establish FlashDP as a pivotal development for efficient and privacy-preserving training of LLMs. FlashDP's code has been open-sourced in https://github.com/kaustpradalab/flashdp.

Linear RNNs with gating recently demonstrated competitive performance compared to Transformers in language modeling. Although their linear compute scaling in sequence length offers theoretical runtime advantages over Transformers, realizing these benefits in practice requires optimized custom kernels, as Transformers rely on the highly efficient Flash Attention kernels (Dao, 2024). Leveraging the chunkwise-parallel formulation of linear RNNs, Flash Linear Attention (FLA) (Yang & Zhang, 2024) shows that linear RNN kernels are faster than Flash Attention, by parallelizing over chunks of the input sequence. However, since the chunk size of FLA is limited, many intermediate states must be materialized in GPU memory. This leads to low arithmetic intensity and causes high memory consumption and IO cost, especially for long-context pre-training. In this work, we present Tiled Flash Linear Attention (TFLA), a novel kernel algorithm for linear RNNs, that enables arbitrary large chunk sizes and high arithmetic intensity by introducing an additional level of sequence parallelization within each chunk. First, we apply TFLA to the xLSTM with matrix memory, the mLSTM (Beck et al., 2024). Second, we propose an mLSTM variant with sigmoid input gate and reduced computation for even faster kernel runtimes at equal language modeling performance. In our speed benchmarks, we show that our new mLSTM kernels based on TFLA outperform highly optimized Flash Attention, Linear Attention and Mamba kernels, setting a new state of the art for efficient long-context sequence modeling primitives. Our code is available at: https://github.com/NX-AI/mlstm_kernels

Vision transformers are ever larger, more accurate, and more expensive to compute. At high resolution, the expense is even more extreme as the number of tokens grows quadratically in the image size. We turn to adaptive computation to cope with this cost by learning to predict where to compute. Our LookWhere method divides the computation between a low-resolution selector and a high-resolution extractor without ever processing the full high-resolution input. We jointly pretrain the selector and extractor without task supervision by distillation from a self-supervised teacher, in effect learning where and what to compute at the same time. Unlike prior token reduction methods, which pay to save by pruning already-computed tokens, and prior token selection methods, which require complex and expensive per-task optimization, LookWhere economically and accurately selects and extracts transferrable representations of images. We show that LookWhere excels at sparse recognition on high-resolution inputs (Traffic Signs), maintaining accuracy while reducing FLOPs by 17x and time by 4x, and standard recognition tasks that are global (ImageNet classification) and local (ADE20K segmentation), improving accuracy while reducing time by 1.36x.

Graph condensation (GC) is an emerging technique designed to learn a significantly smaller graph that retains the essential information of the original graph. This condensed graph has shown promise in accelerating graph neural networks while preserving performance comparable to those achieved with the original, larger graphs. Additionally, this technique facilitates downstream applications like neural architecture search and deepens our understanding of redundancies in large graphs. Despite the rapid development of GC methods, particularly for node classification, a unified evaluation framework is still lacking to systematically compare different GC methods or clarify key design choices for improving their effectiveness. To bridge these gaps, we introduce GC4NC, a comprehensive framework for evaluating diverse GC methods on node classification across multiple dimensions including performance, efficiency, privacy preservation, denoising ability, NAS effectiveness, and transferability. Our systematic evaluation offers novel insights into how condensed graphs behave and the critical design choices that drive their success. These findings pave the way for future advancements in GC methods, enhancing both performance and expanding their real-world applications. The code is available at https://github.com/Emory-Melody/GraphSlim/tree/main/benchmark.

In this paper, we address the problem of cost-sensitive hyperparameter optimization (HPO) built upon freeze-thaw Bayesian optimization (BO). Specifically, we assume a scenario where users want to early-stop the HPO process when the expected performance improvement is not satisfactory with respect to the additional computational cost. Motivated by this scenario, we introduce \emph{utility} in the freeze-thaw framework, a function describing the trade-off between the cost and performance that can be estimated from the user's preference data. This utility function, combined with our novel acquisition function and stopping criterion, allows us to dynamically continue training the configuration that we expect to maximally improve the utility in the future, and also automatically stop the HPO process around the maximum utility. Further, we improve the sample efficiency of existing freeze-thaw methods with transfer learning to develop a specialized surrogate model for the cost-sensitive HPO problem. We validate our algorithm on established multi-fidelity HPO benchmarks and show that it outperforms all the previous freeze-thaw BO and transfer-BO baselines we consider, while achieving a significantly better trade-off between the cost and performance.

Parameter-efficient fine-tuning (PEFT) methods have shown promise in adapting large language models, yet existing approaches exhibit counter-intuitive phenomena: integrating either matrix decomposition or mixture-of-experts (MoE) individually decreases performance across tasks, though decomposition improves results on specific domains despite reducing parameters, while MoE increases parameter count without corresponding decrease in training efficiency. Motivated by these observations and the modular nature of PT, we propose PT-MoE, a novel framework that integrates matrix decomposition with MoE routing for efficient PT. Evaluation results across 17 datasets demonstrate that PT-MoE achieves state-of-the-art performance in both question answering (QA) and mathematical problem solving tasks, improving F1 score by 1.49 points over PT and 2.13 points over LoRA in QA tasks, while improving mathematical accuracy by 10.75 points over PT and 0.44 points over LoRA, all while using 25% fewer parameters than LoRA. Our analysis reveals that while PT methods generally excel in QA tasks and LoRA-based methods in math datasets, the integration of matrix decomposition and MoE in PT-MoE yields complementary benefits: decomposition enables efficient parameter sharing across experts while MoE provides dynamic adaptation, collectively enabling PT-MoE to demonstrate cross-task consistency and generalization abilities. These findings, along with ablation studies on routing mechanisms and architectural components, provide insights for future PEFT methods.

Current parameter-efficient fine-tuning methods for adapting pre-trained language models to downstream tasks are susceptible to interference from noisy data. Conventional noise-handling approaches either rely on laborious data pre-processing or employ model architecture modifications prone to error accumulation. In contrast to existing noise-process paradigms, we propose a noise-robust adaptation method via asymmetric LoRA poisoning experts (LoPE), a novel framework that enhances model robustness to noise only with generated noisy data. Drawing inspiration from the mixture-of-experts architecture, LoPE strategically integrates a dedicated poisoning expert in an asymmetric LoRA configuration. Through a two-stage paradigm, LoPE performs noise injection on the poisoning expert during fine-tuning to enhance its noise discrimination and processing ability. During inference, we selectively mask the dedicated poisoning expert to leverage purified knowledge acquired by normal experts for noise-robust output. Extensive experiments demonstrate that LoPE achieves strong performance and robustness purely through the low-cost noise injection, which completely eliminates the requirement of data cleaning.

Finetuning large pretrained neural networks is known to be resource-intensive, both in terms of memory and computational cost. To mitigate this, a common approach is to restrict training to a subset of the model parameters. By analyzing the relationship between gradients and weights during finetuning, we observe a notable pattern: large gradients are often associated with small-magnitude weights. This correlation is more pronounced in fine-tuning settings than in training from scratch. Motivated by this observation, we propose \textsc{NanoAdam}, which dynamically updates only the small-magnitude weights during fine-tuning and offers several practical advantages: first, the criterion is \emph{gradient-free}—the parameter subset can be determined without gradient computation; second, it preserves large-magnitude weights, which are likely to encode critical features learned during pre-training, thereby reducing the risk of catastrophic forgetting; thirdly, it permits the use of larger learning rates and consistently leads to better generalization performance in experiments. We demonstrate this for both NLP and vision tasks.

As we scale to more massive machine learning models, the frequent synchronization demands inherent in data-parallel approaches create significant slowdowns, posing a critical challenge to further scaling. Recent work develops an approach (DiLoCo) that relaxes synchronization demands without compromising model quality. However, these works do not carefully analyze how DiLoCo's behavior changes with model size. In this work, we study the scaling law behavior of DiLoCo when training LLMs under a fixed compute budget. We focus on how algorithmic factors, including number of model replicas, hyperparameters, and token budget affect training in ways that can be accurately predicted via scaling laws. We find that DiLoCo scales both predictably and robustly with model size. When well-tuned, DiLoCo scales better than data-parallel training with model size, and can outperform data-parallel training even at small model sizes. Our results showcase a more general set of benefits of DiLoCo than previously documented, including increased optimal batch sizes, improved downstream generalization with scale, and improved evaluation loss for a fixed token budget.

Prompt-based learning has emerged as a parameter-efficient finetuning (PEFT) approach to facilitate Large Language Model (LLM) adaptation to downstream tasks by conditioning generation with task-aware guidance. Despite its successes, current prompt-based learning methods heavily rely on laborious grid searching for optimal prompt length and typically require considerable number of prompts, introducing additional computational burden. Worse yet, our pioneer findings indicate that the task-aware prompt design is inherently limited by its absence of instance-aware information, leading to a subtle attention interplay with the input sequence. In contrast, simply incorporating instance-aware information as a part of the guidance can enhance the prompt-tuned model performance without additional fine-tuning. Moreover, we find an interesting phenomenon, namely "attention anchor", that incorporating instance-aware tokens at the earliest position of the sequence can successfully preserve strong attention to critical structural information and exhibit more active attention interaction with all input tokens. In light of our observation, we introduce Capsule Prompt-Tuning (CaPT), an efficient and effective solution that leverages off-the-shelf, informative instance semantics into prompt-based learning. Our approach innovatively integrates both instance-aware and task-aware information in a nearly parameter-free manner (i.e., one single capsule prompt). Empirical results demonstrate that our method can exhibit superior performance across various language tasks (e.g., 84.03\% average accuracy on T5-Large), serving as an "attention anchor," while enjoying high parameter efficiency (e.g., 0.003\% of model parameters on Llama3.2-1B).

We present Jet-Nemotron, a new family of hybrid-architecture language models, which matches or exceeds the accuracy of leading full-attention models while significantly improving generation throughput. Jet-Nemotron is developed using Post Neural Architecture Search (PostNAS), a novel neural architecture exploration pipeline that enables efficient model design. Unlike prior approaches, PostNAS begins with a pre-trained full-attention model and freezes its MLP weights, allowing efficient exploration of attention block designs. The pipeline includes four key components: (1) learning optimal full-attention layer placement and elimination, (2) linear attention block selection, (3) designing new attention blocks, and (4) performing hardware-aware hyperparameter search. Our Jet-Nemotron-2B model achieves comparable or superior accuracy to Qwen3, Qwen2.5, Gemma3, and Llama3.2 across a comprehensive suite of benchmarks while delivering up to 53.6× generation throughput speedup and 6.1× prefilling speedup. It also achieves higher accuracy on MMLU and MMLU-Pro than recent advanced MoE full-attention models, such as DeepSeek-V3-Small and Moonlight, despite their larger scale with 15B total and 2.2B activated parameters.

Prefix caching is a key technique for reducing Large Language Model (LLM) inference costs. However, the prevalent least-recently-used (LRU) eviction algorithm has a large gap to the optimal algorithm. This paper introduces LPC, the first learned method to perform LLM prefix cache eviction. LPC leverages conversational content analysis to provide predictive guidance for eviction, determining which conversations are likely to continue. These insights, combined with last access timestamps, inform more effective cache management. Extensive evaluations across three real-world datasets demonstrate that LPC achieves 18-47% reductions in required cache sizes for equivalent hit ratios and has an 11% improvement in LLM prefilling throughput in an emulated environment.

Spiking neural networks (SNNs) are emerging as a promising alternative to traditional artificial neural networks (ANNs), offering biological plausibility and energy efficiency. Despite these merits, SNNs are frequently hampered by limited capacity and insufficient representation power, yet remain underexplored in remote sensing image (RSI) super-resolution (SR) tasks. In this paper, we first observe that spiking signals exhibit drastic intensity variations across diverse textures, highlighting an active learning state of the neurons. This observation motivates us to apply SNNs for efficient SR of RSIs. Inspired by the success of attention mechanisms in representing salient information, we devise the spiking attention block (SAB), a concise yet effective component that optimizes membrane potentials through inferred attention weights, which, in turn, regulates spiking activity for superior feature representation. Our key contributions include: 1) we bridge the independent modulation between temporal and channel dimensions, facilitating joint feature correlation learning, and 2) we access the global self-similar patterns in large-scale remote sensing imagery to infer spatial attention weights, incorporating effective priors for realistic and faithful reconstruction. Building upon SAB, we proposed SpikeSR, which achieves state-of-the-art performance across various remote sensing benchmarks such as AID, DOTA, and DIOR, while maintaining high computational efficiency. Code of SpikeSR will be available at https://github.com/XY-boy/SpikeSR.

We find that the cross-entropy loss curves of neural language models empirically adhere to a scaling law with learning rate (LR) annealing over training steps: $$L(s) = L_0 + A\cdot S_1^{-\alpha} - C\cdot S_2,$$ where $L(s)$ is the validation loss at step $s$, $S_1$ is the area under the LR curve, $S_2$ is the LR annealing area, and $L_0$, $A$, $C$, $\alpha$ are constant parameters. This formulation accounts for two main effects: (1) power-law scaling over data size, and (2) the additional loss reduction during LR annealing. Unlike previous studies that only fit losses at final steps, our formulation captures the entire training curve, allowing for parameter fitting using losses from any training step. Applying the scaling law with LR annealing and fitting only one or two training curves, we can accurately predict the loss at any given step under any learning rate scheduler (LRS). This approach significantly reduces computational cost in formulating scaling laws while providing more accuracy and expressiveness. Extensive experiments demonstrate that our findings hold across a range of hyper-parameters and model architectures and can extend to scaling effect of model sizes. Moreover, our formulation provides accurate theoretical insights into empirical results observed in numerous previous studies, particularly those focusing on LR schedule and annealing. We believe that this work is promising to enhance the understanding of LLM training dynamics while democratizing scaling laws, and it is helpful to guide both research and industrial participants in refining training strategies for further LLMs.

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×.

Data selection improves computational efficiency by choosing informative subsets of training samples. However, existing methods ignore the compute budget, treating data selection and importance evaluation independently of compute budget constraints. Yet empirical studies show no algorithm can consistently outperform others (or even random selection) across varying budgets. We therefore argue that compute budget must be integral to data-selection strategies, since different budgets impose distinct requirements on data quantity, quality, and distribution for effective training. To this end, we propose a novel Computational budget-Aware Data Selection (CADS) method and naturally formulate it into a bilevel optimization framework, where the inner loop trains the model within the constraints of the computational budget on some selected subset of training data, while the outer loop optimizes data selection based on model evaluation. Our technical contributions lie in addressing two main challenges in solving this bilevel optimization problem: the expensive Hessian matrix estimation for outer-loop gradients and the computational burden of achieving inner-loop optimality during iterations. To solve the first issue, we propose a probabilistic reparameterization strategy and compute the gradient using a Hessian-free policy gradient estimator. To address the second challenge, we transform the inner optimization problem into a penalty term in the outer objective, further discovering that we only need to estimate the minimum of a one-dimensional loss to calculate the gradient, significantly improving efficiency. To accommodate different data selection granularities, we present two complementary CADS variants: an example-level version (CADS-E) offering fine-grained control and a source-level version (CADS-S) aggregating samples into source groups for scalable, efficient selection without sacrificing effectiveness. Extensive experiments show that our method achieves performance gains of up to 14.42\% over baselines in vision and language benchmarks. Additionally, CADS achieves a 3-20× speedup compared to conventional bilevel implementations, with acceleration correlating positively with compute budget size.