Sparsity-Controllable Dynamic Top-p MoE for Large Foundation Model Pre-training

Sparse Mixture-of-Experts (MoE) architectures effectively scale model capacity by activating only a subset of experts for each input token. However, the standard Top-k routing strategy imposes a uniform sparsity pattern that ignores the varying difficulty of tokens. While Top-p routing offers a flexible alternative, existing implementations typically rely on a fixed global probability threshold, which results in uncontrolled computational costs and sensitivity to hyperparameter selection. In this paper, we propose DTop-p MoE, a sparsity-controllable dynamic Top-p routing mechanism. To resolve the challenge of optimizing a non-differentiable threshold, we utilize a Proportional-Integral (PI) Controller that dynamically adjusts the probability threshold to align the running activated-expert sparsity with a specified target. Furthermore, we introduce a dynamic routing normalization mechanism that adapts layer-wise routing logits, allowing different layers to learn distinct expert-selection patterns while utilizing a global probability threshold. Extensive experiments on Large Language Models and Diffusion Transformers demonstrate that DTop-p consistently outperforms both Top-k and fixed-threshold Top-p baselines. Our analysis confirms that DTop-p maintains precise control over the number of activated experts while adaptively allocating resources across different tokens and layers. Furthermore, DTop-p exhibits strong scaling properties with respect to expert granularity, expert capacity, model size, and dataset size, offering a robust framework for large-scale MoE pre-training.

Sparse by Rule: Probability-Based N:M Pruning for Spiking Neural Networks

Brain-inspired Spiking neural networks (SNNs) promise energy-efficient intelligence via event-driven, sparse computation, but deeper architectures inflate parameters and computational cost, hindering their edge deployment. Recent progress in SNN pruning helps alleviate this burden, yet existing efforts fall into only two families: \emph{unstructured} pruning, which attains high sparsity but is difficult to accelerate on general hardware, and \emph{structured} pruning, which eases deployment but lack flexibility and often degrades accuracy at matched sparsity. In this work, we introduce \textbf{SpikeNM}, the first SNN-oriented \emph{semi-structured} \(N{:}M\) pruning framework that learns sparse SNNs \emph{from scratch}, enforcing \emph{at most \(N\)} non-zeros per \(M\)-weight block. To avoid the combinatorial space complexity \(\sum_{k=1}^{N}\binom{M}{k}\) growing exponentially with \(M\), SpikeNM adopts an \(M\)-way basis-logit parameterization with a differentiable top-\(k\) sampler, \emph{linearizing} per-block complexity to \(\mathcal O(M)\) and enabling more aggressive sparsification. Further inspired by neuroscience, we propose \emph{eligibility-inspired distillation} (EID), which converts temporally accumulated credits into block-wise soft targets to align mask probabilities with spiking dynamics, reducing sampling variance and stabilizing search under high sparsity. Experiments show that at \(2{:}4\) sparsity, SpikeNM maintains and even with gains across main-stream datasets, while yielding hardware-amenable patterns that complement intrinsic spike sparsity.

Breaking the Modality Wall: Time-step Mixup for Efficient Spiking Knowledge Transfer from Static to Event Domain

The integration of event cameras and spiking neural networks (SNNs) promises energy-efficient visual intelligence, yet scarce event data and the sparsity of DVS outputs hinder effective training. Prior knowledge transfers from RGB to DVS often underperform because the distribution gap between modalities is substantial. In this work, we present Time-step Mixup Knowledge Transfer (TMKT), a cross-modal training framework with a probabilistic Time-step Mixup (TSM) strategy. TSM exploits the asynchronous nature of SNNs by interpolating RGB and DVS inputs at various time steps to produce a smooth curriculum within each sequence, which reduces gradient variance and stabilizes optimization with theoretical analysis. To employ auxiliary supervision from TSM, TMKT introduces two lightweight modality-aware objectives, Modality Aware Guidance (MAG) for per-frame source supervision and Mixup Ratio Perception (MRP) for sequence-level mix ratio estimation, which explicitly align temporal features with the mixing schedule. TMKT enables smoother knowledge transfer, helps mitigate modality mismatch during training, and achieves superior performance in spiking image classification tasks. Extensive experiments across diverse benchmarks and multiple SNN backbones, together with ablations, demonstrate the effectiveness of our method.

Learning from Dense Events: Towards Fast Spiking Neural Networks Training via Event Dataset Distillatio

Event cameras sense brightness changes and output binary asynchronous event streams, attracting increasing attention. Their bio-inspired dynamics align well with spiking neural networks (SNNs), offering a promising energy-efficient alternative to conventional vision systems. However, SNNs remain costly to train due to temporal coding, which limits their practical deployment. To alleviate the high training cost of SNNs, we introduce \textbf{PACE} (Phase-Aligned Condensation for Events), the first dataset distillation framework to SNNs and event-based vision. PACE distills a large training dataset into a compact synthetic one that enables fast SNN training, which is achieved by two core modules: \textbf{ST-DSM} and \textbf{PEQ-N}. ST-DSM uses residual membrane potentials to densify spike-based features (SDR) and to perform fine-grained spatiotemporal matching of amplitude and phase (ST-SM), while PEQ-N provides a plug-and-play straight through probabilistic integer quantizer compatible with standard event-frame pipelines. Across DVS-Gesture, CIFAR10-DVS, and N-MNIST datasets, PACE outperforms existing coreset selection and dataset distillation baselines, with particularly strong gains on dynamic event streams and at low or moderate IPC. Specifically, on N-MNIST, it achieves \(84.4\%\) accuracy, about \(85\%\) of the full training set performance, while reducing training time by more than \(50\times\) and storage cost by \(6000\times\), yielding compact surrogates that enable minute-scale SNN training and efficient edge deployment.

Effective Policy Learning for Multi-Agent Online Coordination Beyond Submodular Objectives

In this paper, we present two effective policy learning algorithms for multi-agent online coordination(MA-OC) problem. The first one, \texttt{MA-SPL}, not only can achieve the optimal $(1-\frac{c}{e})$-approximation guarantee for the MA-OC problem with submodular objectives but also can handle the unexplored $\alpha$-weakly DR-submodular and $(\gamma,\beta)$-weakly submodular scenarios, where $c$ is the curvature of the investigated submodular functions, $\alpha$ denotes the diminishing-return(DR) ratio and the tuple $(\gamma,\beta)$ represents the submodularity ratios. Subsequently, in order to reduce the reliance on the unknown parameters $\alpha,\gamma,\beta$ inherent in the \texttt{MA-SPL} algorithm, we further introduce the second online algorithm named \texttt{MA-MPL}. This \texttt{MA-MPL} algorithm is entirely \emph{parameter-free} and simultaneously can maintain the same approximation ratio as the first \texttt{MA-SPL} algorithm. The core of our \texttt{MA-SPL} and \texttt{MA-MPL} algorithms is a novel continuous-relaxation technique termed as \emph{policy-based continuous extension}. Compared with the well-established \emph{multi-linear extension}, a notable advantage of this new \emph{policy-based continuous extension} is its ability to provide a lossless rounding scheme for any set function, thereby enabling us to tackle the challenging weakly submodular objectives. Finally, extensive simulations are conducted to validate the effectiveness of our proposed algorithms.

Your Reward Function for RL is Your Best PRM for Search: Unifying RL and Search-Based TTS

Test-time scaling (TTS) for large language models (LLMs) has thus far fallen into two largely separate paradigms: (1) reinforcement learning (RL) methods that optimize sparse outcome-based rewards, yet suffer from instability and low sample efficiency; and (2) search-based techniques guided by independently trained, static process reward models (PRMs), which require expensive human- or LLM-generated labels and often degrade under distribution shifts. In this paper, we introduce AIRL-S, the first natural unification of RL-based and search-based TTS. Central to AIRL-S is the insight that the reward function learned during RL training inherently represents the ideal PRM for guiding downstream search. Specifically, we leverage adversarial inverse reinforcement learning (AIRL) combined with group relative policy optimization (GRPO) to learn a dense, dynamic PRM directly from correct reasoning traces, entirely eliminating the need for labeled intermediate process data. At inference, the resulting PRM simultaneously serves as the critic for RL rollouts and as a heuristic to effectively guide search procedures, facilitating robust reasoning chain extension, mitigating reward hacking, and enhancing cross-task generalization. Experimental results across eight benchmarks, including mathematics, scientific reasoning, and code generation, demonstrate that our unified approach improves performance by 9 % on average over the base model, matching GPT-4o. Furthermore, when integrated into multiple search algorithms, our PRM consistently outperforms all baseline PRMs trained with labeled data. These results underscore that, indeed, your reward function for RL is your best PRM for search, providing a robust and cost-effective solution to complex reasoning tasks in LLMs.

Data-Efficient Biomedical In-Context Learning: A Diversity-Enhanced Submodular Perspective

Recent progress in large language models (LLMs) has leveraged their in-context learning (ICL) abilities to enable quick adaptation to unseen biomedical NLP tasks. By incorporating only a few input-output examples into prompts, LLMs can rapidly perform these new tasks. While the impact of these demonstrations on LLM performance has been extensively studied, most existing approaches prioritize representativeness over diversity when selecting examples from large corpora. To address this gap, we propose Dual-Div, a diversity-enhanced data-efficient framework for demonstration selection in biomedical ICL. Dual-Div employs a two-stage retrieval and ranking process: First, it identifies a limited set of candidate examples from a corpus by optimizing both representativeness and diversity (with optional annotation for unlabeled data). Second, it ranks these candidates against test queries to select the most relevant and non-redundant demonstrations. Evaluated on three biomedical NLP tasks (named entity recognition (NER), relation extraction (RE), and text classification (TC)) using LLaMA 3.1 and Qwen 2.5 for inference, along with three retrievers (BGE-Large, BMRetriever, MedCPT), Dual-Div consistently outperforms baselines-achieving up to 5% higher macro-F1 scores-while demonstrating robustness to prompt permutations and class imbalance. Our findings establish that diversity in initial retrieval is more critical than ranking-stage optimization, and limiting demonstrations to 3-5 examples maximizes performance efficiency.

Fractional Spike Differential Equations Neural Network with Efficient Adjoint Parameters Training

Spiking Neural Networks (SNNs) draw inspiration from biological neurons to create realistic models for brain-like computation, demonstrating effectiveness in processing temporal information with energy efficiency and biological realism. Most existing SNNs assume a single time constant for neuronal membrane voltage dynamics, modeled by first-order ordinary differential equations (ODEs) with Markovian characteristics. Consequently, the voltage state at any time depends solely on its immediate past value, potentially limiting network expressiveness. Real neurons, however, exhibit complex dynamics influenced by long-term correlations and fractal dendritic structures, suggesting non-Markovian behavior. Motivated by this, we propose the Fractional SPIKE Differential Equation neural network (fspikeDE), which captures long-term dependencies in membrane voltage and spike trains through fractional-order dynamics. These fractional dynamics enable more expressive temporal patterns beyond the capability of integer-order models. For efficient training of fspikeDE, we introduce a gradient descent algorithm that optimizes parameters by solving an augmented fractional-order ODE (FDE) backward in time using adjoint sensitivity methods. Extensive experiments on diverse image and graph datasets demonstrate that fspikeDE consistently outperforms traditional SNNs, achieving superior accuracy, comparable energy efficiency, reduced training memory usage, and enhanced robustness against noise. Our approach provides a novel open-sourced computational toolbox for fractional-order SNNs, widely applicable to various real-world tasks.

MaskPro: Linear-Space Probabilistic Learning for Strict (N:M)-Sparsity on Large Language Models

The rapid scaling of large language models (LLMs) has made inference efficiency a primary bottleneck in the practical deployment. To address this, semi-structured sparsity offers a promising solution by strategically retaining $N$ elements out of every $M$ weights, thereby enabling hardware-friendly acceleration and reduced memory. However, existing (N:M)-compatible approaches typically fall into two categories: rule-based layerwise greedy search, which suffers from considerable errors, and gradient-driven combinatorial learning, which incurs prohibitive training costs. To tackle these challenges, we propose a novel linear-space probabilistic framework named MaskPro, which aims to learn a prior categorical distribution for every $M$ consecutive weights and subsequently leverages this distribution to generate the (N:M)-sparsity throughout an $N$-way sampling without replacement. Furthermore, to mitigate the training instability induced by the high variance of policy gradients in the super large combinatorial space, we propose a novel update method by introducing a moving average tracker of loss residuals instead of vanilla loss. Finally, we conduct comprehensive theoretical analysis and extensive experiments to validate the superior performance of MaskPro, as well as its excellent scalability in memory efficiency and exceptional robustness to data samples. Our code is available at https://github.com/woodenchild95/Maskpro.git.

Dynamic Text Bundling Supervision for Zero-Shot Inference on Text-Attributed Graphs

Large language models (LLMs) have been used in many zero-shot learning problems, with their strong generalization ability. Recently, adopting LLMs in text-attributed graphs (TAGs) has drawn increasing attention. However, the adoption of LLMs faces two major challenges: limited information on graph structure and unreliable responses. LLMs struggle with text attributes isolated from the graph topology. Worse still, they yield unreliable predictions due to both information insufficiency and the inherent weakness of LLMs (e.g., hallucination). Towards this end, this paper proposes a novel method named Dynamic Text Bundling Supervision (DENSE) that queries LLMs with bundles of texts to obtain bundle-level labels and uses these labels to supervise graph neural networks. Specifically, we sample a set of bundles, each containing a set of nodes with corresponding texts of close proximity. We then query LLMs with the bundled texts to obtain the label of each bundle. Subsequently, the bundle labels are used to supervise the optimization of graph neural networks, and the bundles are further refined to exclude noisy items. To justify our design, we also provide theoretical analysis of the proposed method. Extensive experiments across ten datasets validate the effectiveness of the proposed method.