Diversity-Aware Policy Optimization for Large Language Model Reasoning

The reasoning capabilities of large language models (LLMs) have advanced rapidly, particularly following the release of DeepSeek R1, which has inspired a surge of research into data quality and reinforcement learning (RL) algorithms. Despite the pivotal role diversity plays in RL, its influence on LLM reasoning remains largely underexplored. To bridge this gap, this work presents a systematic investigation into the impact of diversity in RL-based training for LLM reasoning, and proposes a novel diversity-aware policy optimization method. Across evaluations on 12 LLMs, we observe a strong positive correlation between the solution diversity and Potential at k (a novel metric quantifying an LLM's reasoning potential) in high-performing models. This finding motivates our method to explicitly promote diversity during RL training. Specifically, we design a token-level diversity and reformulate it into a practical objective, then we selectively apply it to positive samples. Integrated into the R1-zero training framework, our method achieves a 3.5 percent average improvement across four mathematical reasoning benchmarks, while generating more diverse and robust solutions.

Neuromorphic Sequential Arena: A Benchmark for Neuromorphic Temporal Processing

Temporal processing is vital for extracting meaningful information from time-varying signals. Recent advancements in Spiking Neural Networks (SNNs) have shown immense promise in efficiently processing these signals. However, progress in this field has been impeded by the lack of effective and standardized benchmarks, which complicates the consistent measurement of technological advancements and limits the practical applicability of SNNs. To bridge this gap, we introduce the Neuromorphic Sequential Arena (NSA), a comprehensive benchmark that offers an effective, versatile, and application-oriented evaluation framework for neuromorphic temporal processing. The NSA includes seven real-world temporal processing tasks from a diverse range of application scenarios, each capturing rich temporal dynamics across multiple timescales. Utilizing NSA, we conduct extensive comparisons of recently introduced spiking neuron models and neural architectures, presenting comprehensive baselines in terms of task performance, training speed, memory usage, and energy efficiency. Our findings emphasize an urgent need for efficient SNN designs that can consistently deliver high performance across tasks with varying temporal complexities while maintaining low computational costs. NSA enables systematic tracking of advancements in neuromorphic algorithm research and paves the way for developing effective and efficient neuromorphic temporal processing systems.

MSVIT: Improving Spiking Vision Transformer Using Multi-scale Attention Fusion

The combination of Spiking Neural Networks(SNNs) with Vision Transformer architectures has attracted significant attention due to the great potential for energy-efficient and high-performance computing paradigms. However, a substantial performance gap still exists between SNN-based and ANN-based transformer architectures. While existing methods propose spiking self-attention mechanisms that are successfully combined with SNNs, the overall architectures proposed by these methods suffer from a bottleneck in effectively extracting features from different image scales. In this paper, we address this issue and propose MSVIT, a novel spike-driven Transformer architecture, which firstly uses multi-scale spiking attention (MSSA) to enrich the capability of spiking attention blocks. We validate our approach across various main data sets. The experimental results show that MSVIT outperforms existing SNN-based models, positioning itself as a state-of-the-art solution among SNN-transformer architectures. The codes are available at https://github.com/Nanhu-AI-Lab/MSViT.

EventDiff: A Unified and Efficient Diffusion Model Framework for Event-based Video Frame Interpolation

Video Frame Interpolation (VFI) is a fundamental yet challenging task in computer vision, particularly under conditions involving large motion, occlusion, and lighting variation. Recent advancements in event cameras have opened up new opportunities for addressing these challenges. While existing event-based VFI methods have succeeded in recovering large and complex motions by leveraging handcrafted intermediate representations such as optical flow, these designs often compromise high-fidelity image reconstruction under subtle motion scenarios due to their reliance on explicit motion modeling. Meanwhile, diffusion models provide a promising alternative for VFI by reconstructing frames through a denoising process, eliminating the need for explicit motion estimation or warping operations. In this work, we propose EventDiff, a unified and efficient event-based diffusion model framework for VFI. EventDiff features a novel Event-Frame Hybrid AutoEncoder (HAE) equipped with a lightweight Spatial-Temporal Cross Attention (STCA) module that effectively fuses dynamic event streams with static frames. Unlike previous event-based VFI methods, EventDiff performs interpolation directly in the latent space via a denoising diffusion process, making it more robust across diverse and challenging VFI scenarios. Through a two-stage training strategy that first pretrains the HAE and then jointly optimizes it with the diffusion model, our method achieves state-of-the-art performance across multiple synthetic and real-world event VFI datasets. The proposed method outperforms existing state-of-the-art event-based VFI methods by up to 1.98dB in PSNR on Vimeo90K-Triplet and shows superior performance in SNU-FILM tasks with multiple difficulty levels. Compared to the emerging diffusion-based VFI approach, our method achieves up to 5.72dB PSNR gain on Vimeo90K-Triplet and 4.24X faster inference.

Spiking Neural Networks for Temporal Processing: Status Quo and Future Prospects

Temporal processing is fundamental for both biological and artificial intelligence systems, as it enables the comprehension of dynamic environments and facilitates timely responses. Spiking Neural Networks (SNNs) excel in handling such data with high efficiency, owing to their rich neuronal dynamics and sparse activity patterns. Given the recent surge in the development of SNNs, there is an urgent need for a comprehensive evaluation of their temporal processing capabilities. In this paper, we first conduct an in-depth assessment of commonly used neuromorphic benchmarks, revealing critical limitations in their ability to evaluate the temporal processing capabilities of SNNs. To bridge this gap, we further introduce a benchmark suite consisting of three temporal processing tasks characterized by rich temporal dynamics across multiple timescales. Utilizing this benchmark suite, we perform a thorough evaluation of recently introduced SNN approaches to elucidate the current status of SNNs in temporal processing. Our findings indicate significant advancements in recently developed spiking neuron models and neural architectures regarding their temporal processing capabilities, while also highlighting a performance gap in handling long-range dependencies when compared to state-of-the-art non-spiking models. Finally, we discuss the key challenges and outline potential avenues for future research.

MetaLA: Unified Optimal Linear Approximation to Softmax Attention Map

Various linear complexity models, such as Linear Transformer (LinFormer), State Space Model (SSM), and Linear RNN (LinRNN), have been proposed to replace the conventional softmax attention in Transformer structures. However, the optimal design of these linear models is still an open question. In this work, we attempt to answer this question by finding the best linear approximation to softmax attention from a theoretical perspective. We start by unifying existing linear complexity models as the linear attention form and then identify three conditions for the optimal linear attention design: 1) Dynamic memory ability; 2) Static approximation ability; 3) Least parameter approximation. We find that none of the current linear models meet all three conditions, resulting in suboptimal performance. Instead, we propose Meta Linear Attention (MetaLA) as a solution that satisfies these conditions. Our experiments on Multi-Query Associative Recall (MQAR) task, language modeling, image classification, and Long-Range Arena (LRA) benchmark demonstrate that MetaLA is more effective than the existing linear models.

Towards Ultra-Low-Power Neuromorphic Speech Enhancement with Spiking-FullSubNet

Speech enhancement is critical for improving speech intelligibility and quality in various audio devices. In recent years, deep learning-based methods have significantly improved speech enhancement performance, but they often come with a high computational cost, which is prohibitive for a large number of edge devices, such as headsets and hearing aids. This work proposes an ultra-low-power speech enhancement system based on the brain-inspired spiking neural network (SNN) called Spiking-FullSubNet. Spiking-FullSubNet follows a full-band and sub-band fusioned approach to effectively capture both global and local spectral information. To enhance the efficiency of computationally expensive sub-band modeling, we introduce a frequency partitioning method inspired by the sensitivity profile of the human peripheral auditory system. Furthermore, we introduce a novel spiking neuron model that can dynamically control the input information integration and forgetting, enhancing the multi-scale temporal processing capability of SNN, which is critical for speech denoising. Experiments conducted on the recent Intel Neuromorphic Deep Noise Suppression (N-DNS) Challenge dataset show that the Spiking-FullSubNet surpasses state-of-the-art methods by large margins in terms of both speech quality and energy efficiency metrics. Notably, our system won the championship of the Intel N-DNS Challenge (Algorithmic Track), opening up a myriad of opportunities for ultra-low-power speech enhancement at the edge. Our source code and model checkpoints are publicly available at https://github.com/haoxiangsnr/spiking-fullsubnet.

HM3: Hierarchical Multi-Objective Model Merging for Pretrained Models

Model merging is a technique that combines multiple large pretrained models into a single model with enhanced performance and broader task adaptability. It has gained popularity in large pretrained model development due to its ability to bypass the need for original training data and further training processes. However, most existing model merging approaches focus solely on exploring the parameter space, merging models with identical architectures. Merging within the architecture space, despite its potential, remains in its early stages due to the vast search space and the challenges of layer compatibility. This paper marks a significant advance toward more flexible and comprehensive model merging techniques by modeling the architecture-space merging process as a reinforcement learning task. We train policy and value networks using offline sampling of weight vectors, which are then employed for the online optimization of merging strategies. Moreover, a multi-objective optimization paradigm is introduced to accommodate users' diverse task preferences, learning the Pareto front of optimal models to offer customized merging suggestions. Experimental results across multiple tasks, including text translation, mathematical reasoning, and code generation, validate the effectiveness and superiority of the proposed framework in model merging. The code will be made publicly available after the review process.

Advancing Automated Knowledge Transfer in Evolutionary Multitasking via Large Language Models

Evolutionary Multi-task Optimization (EMTO) is a paradigm that leverages knowledge transfer across simultaneously optimized tasks for enhanced search performance. To facilitate EMTO's performance, various knowledge transfer models have been developed for specific optimization tasks. However, designing these models often requires substantial expert knowledge. Recently, large language models (LLMs) have achieved remarkable success in autonomous programming, aiming to produce effective solvers for specific problems. In this work, a LLM-based optimization paradigm is introduced to establish an autonomous model factory for generating knowledge transfer models, ensuring effective and efficient knowledge transfer across various optimization tasks. To evaluate the performance of the proposed method, we conducted comprehensive empirical studies comparing the knowledge transfer model generated by the LLM with existing state-of-the-art knowledge transfer methods. The results demonstrate that the generated model is able to achieve superior or competitive performance against hand-crafted knowledge transfer models in terms of both efficiency and effectiveness.

Distance-Forward Learning: Enhancing the Forward-Forward Algorithm Towards High-Performance On-Chip Learning

The Forward-Forward (FF) algorithm was recently proposed as a local learning method to address the limitations of backpropagation (BP), offering biological plausibility along with memory-efficient and highly parallelized computational benefits. However, it suffers from suboptimal performance and poor generalization, largely due to inadequate theoretical support and a lack of effective learning strategies. In this work, we reformulate FF using distance metric learning and propose a distance-forward algorithm (DF) to improve FF performance in supervised vision tasks while preserving its local computational properties, making it competitive for efficient on-chip learning. To achieve this, we reinterpret FF through the lens of centroid-based metric learning and develop a goodness-based N-pair margin loss to facilitate the learning of discriminative features. Furthermore, we integrate layer-collaboration local update strategies to reduce information loss caused by greedy local parameter updates. Our method surpasses existing FF models and other advanced local learning approaches, with accuracies of 99.7\% on MNIST, 88.2\% on CIFAR-10, 59\% on CIFAR-100, 95.9\% on SVHN, and 82.5\% on ImageNette, respectively. Moreover, it achieves comparable performance with less than 40\% memory cost compared to BP training, while exhibiting stronger robustness to multiple types of hardware-related noise, demonstrating its potential for online learning and energy-efficient computation on neuromorphic chips.