Pu Li

Hongyi Yao, Pu Li, Jian Cao et al.

We introduce a Power-of-Two low-bit post-training quantization(PTQ) method for deep neural network that meets hardware requirements and does not call for long-time retraining. Power-of-Two quantization can convert the multiplication introduced by quantization and dequantization to bit-shift that is adopted by many efficient accelerators. However, the Power-of-Two scale factors have fewer candidate values, which leads to more rounding or clipping errors. We propose a novel Power-of-Two PTQ framework, dubbed RAPQ, which dynamically adjusts the Power-of-Two scales of the whole network instead of statically determining them layer by layer. It can theoretically trade off the rounding error and clipping error of the whole network. Meanwhile, the reconstruction method in RAPQ is based on the BN information of every unit. Extensive experiments on ImageNet prove the excellent performance of our proposed method. Without bells and whistles, RAPQ can reach accuracy of 65% and 48% on ResNet-18 and MobileNetV2 respectively with weight INT2 activation INT4. We are the first to propose the more constrained but hardware-friendly Power-of-Two quantization scheme for low-bit PTQ specially and prove that it can achieve nearly the same accuracy as SOTA PTQ method. The code was released.

Zijun Liu, Peiyi Wang, Runxin Xu et al. · tsinghua

Reinforcement learning (RL) has been widely adopted in post-training for large language models (LLMs) at scale. Recently, the incentivization of reasoning capabilities in LLMs from RL indicates that $\textit{proper learning methods could enable effective inference-time scalability}$. A key challenge of RL is to obtain accurate reward signals for LLMs in various domains beyond verifiable questions or artificial rules. In this work, we investigate how to improve reward modeling (RM) with more inference compute for general queries, i.e. the $\textbf{inference-time scalability of generalist RM}$, and further, how to improve the effectiveness of performance-compute scaling with proper learning methods. For the RM approach, we adopt pointwise generative reward modeling (GRM) to enable flexibility for different input types and potential for inference-time scaling. For the learning method, we propose Self-Principled Critique Tuning (SPCT) to foster scalable reward generation behaviors in GRMs through online RL, to generate principles adaptively and critiques accurately, resulting in $\textbf{DeepSeek-GRM}$ models. Furthermore, for effective inference-time scaling, we use parallel sampling to expand compute usage, and introduce a meta RM to guide voting process for better scaling performance. Empirically, we show that SPCT significantly improves the quality and scalability of GRMs, outperforming existing methods and models in various RM benchmarks without severe biases, and could achieve better performance compared to training-time scaling. DeepSeek-GRM still meets challenges in some tasks, which we believe can be addressed by future efforts in generalist reward systems. The models are released at Hugging Face and ModelScope.

Ziyue Wang, Yurui Dong, Fuwen Luo et al.

The rapid advancing of Multimodal Large Language Models (MLLMs) has spurred interest in complex multimodal reasoning tasks in the real-world and virtual environment, which require coordinating multiple abilities, including visual perception, visual reasoning, spatial awareness, and target deduction. However, existing evaluations primarily assess the final task completion, often degrading assessments to isolated abilities such as visual grounding and visual question answering. Less attention is given to comprehensively and quantitatively analyzing reasoning process in multimodal environments, which is crucial for understanding model behaviors and underlying reasoning mechanisms beyond merely task success. To address this, we introduce MM-Escape, an extensible benchmark for investigating multimodal reasoning, inspired by real-world escape games. MM-Escape emphasizes intermediate model behaviors alongside final task completion. To achieve this, we develop EscapeCraft, a customizable and open environment that enables models to engage in free-form exploration for assessing multimodal reasoning. Extensive experiments show that MLLMs, regardless of scale, can successfully complete the simplest room escape tasks, with some exhibiting human-like exploration strategies. Yet, performance dramatically drops as task difficulty increases. Moreover, we observe that performance bottlenecks vary across models, revealing distinct failure modes and limitations in their multimodal reasoning abilities, such as repetitive trajectories without adaptive exploration, getting stuck in corners due to poor visual spatial awareness, and ineffective use of acquired props, such as the key. We hope our work sheds light on new challenges in multimodal reasoning, and uncovers potential improvements in MLLMs capabilities.

Fangwei Hao, Ji Du, Desheng Kong et al.

In recent years, the performance of lightweight Single-Image Super-Resolution (SISR) has been improved significantly with the application of Convolutional Neural Networks (CNNs) and Large Kernel Attention (LKA). However, existing information distillation modules for lightweight SISR struggle to map inputs into High-Dimensional Non-Linear (HDNL) feature spaces, limiting their representation learning. And their LKA modules possess restricted ability to capture the multi-shape multi-scale information for long-range dependencies while encountering a quadratic increase in the computational burden with increasing convolutional kernel size of its depth-wise convolutional layer. To address these issues, we firstly propose a Star Distillation Module (SDM) to enhance the discriminative representation learning via information distillation in the HDNL feature spaces. Besides, we present a Multi-shape Multi-scale Large Kernel Attention (MM-LKA) module to learn representative long-range dependencies while incurring low computational and memory footprints, leading to improving the performance of CNN-based self-attention significantly. Integrating SDM and MM-LKA, we develop a Residual Star Distillation Attention Module (RSDAM) and take it as the building block of the proposed efficient Star Distillation Attention Network (SDAN) which possesses high reconstruction efficiency to recover a higher-quality image from the corresponding low-resolution (LR) counterpart. When compared with other lightweight state-of-the-art SISR methods, extensive experiments show that our SDAN with low model complexity yields superior performance quantitatively and visually.

Chuanyu Luo, Nuo Cheng, Ren Zhong et al.

With the rapid advancement of hardware and software technologies, research in autonomous driving has seen significant growth. The prevailing framework for multi-sensor autonomous driving encompasses sensor installation, perception, path planning, decision-making, and motion control. At the perception phase, a common approach involves utilizing neural networks to infer 3D bounding box (Bbox) attributes from raw sensor data, including classification, size, and orientation. In this paper, we present a novel attribute and its corresponding algorithm: 3D object visibility. By incorporating multi-task learning, the introduction of this attribute, visibility, negligibly affects the model's effectiveness and efficiency. Our proposal of this attribute and its computational strategy aims to expand the capabilities for downstream tasks, thereby enhancing the safety and reliability of real-time autonomous driving in real-world scenarios.

Pengtao Xu, Jian Cao, Fanhua Shang et al.

In order to deploy deep convolutional neural networks (CNNs) on resource-limited devices, many model pruning methods for filters and weights have been developed, while only a few to layer pruning. However, compared with filter pruning and weight pruning, the compact model obtained by layer pruning has less inference time and run-time memory usage when the same FLOPs and number of parameters are pruned because of less data moving in memory. In this paper, we propose a simple layer pruning method using fusible residual convolutional block (ResConv), which is implemented by inserting shortcut connection with a trainable information control parameter into a single convolutional layer. Using ResConv structures in training can improve network accuracy and train deep plain networks, and adds no additional computation during inference process because ResConv is fused to be an ordinary convolutional layer after training. For layer pruning, we convert convolutional layers of network into ResConv with a layer scaling factor. In the training process, the L1 regularization is adopted to make the scaling factors sparse, so that unimportant layers are automatically identified and then removed, resulting in a model of layer reduction. Our pruning method achieves excellent performance of compression and acceleration over the state-of-the-arts on different datasets, and needs no retraining in the case of low pruning rate. For example, with ResNet-110, we achieve a 65.5%-FLOPs reduction by removing 55.5% of the parameters, with only a small loss of 0.13% in top-1 accuracy on CIFAR-10.