Chang Gao

Haoran Sun, Lixia Wang, Yin-Chen Hsu et al.

Geometric distortion in prostate diffusion-weighted imaging (DWI) can impair lesion localization and reduce the reliability of MRI-based clinical assessment. We propose AutoIQ, an ensemble machine learning framework for automatic quantification and classification of DWI geometric distortion severity. A total of 140 retrospective prostate biparametric MRI examinations were analyzed, including 33 scans with severe distortion requiring repeat acquisition and 107 scans with acceptable distortion based on expert radiologist assessment. AutoIQ combines two complementary distortion quantification strategies: a segmentation-based method measuring prostate boundary mismatch between T2-weighted imaging (T2WI) and DWI, and a registration-based method estimating deformation magnitude after DWI-to-T2WI alignment. The resulting distortion scores were used to train individual classifiers and a logistic-regression ensemble model. Both computational methods significantly differentiated severe from acceptable distortion cases (p < 0.001). On an independent test set, the ensemble model achieved an accuracy of 0.95, F1-score of 0.93, and AUC of 0.98, outperforming individual models. These results suggest that AutoIQ can provide automated, quantitative quality assessment for prostate DWI and may help identify scans that require repeat acquisition.

Wenxuan Zhang, Sharifah Mahani Aljunied, Chang Gao et al.

Despite the existence of various benchmarks for evaluating natural language processing models, we argue that human exams are a more suitable means of evaluating general intelligence for large language models (LLMs), as they inherently demand a much wider range of abilities such as language understanding, domain knowledge, and problem-solving skills. To this end, we introduce M3Exam, a novel benchmark sourced from real and official human exam questions for evaluating LLMs in a multilingual, multimodal, and multilevel context. M3Exam exhibits three unique characteristics: (1) multilingualism, encompassing questions from multiple countries that require strong multilingual proficiency and cultural knowledge; (2) multimodality, accounting for the multimodal nature of many exam questions to test the model's multimodal understanding capability; and (3) multilevel structure, featuring exams from three critical educational periods to comprehensively assess a model's proficiency at different levels. In total, M3Exam contains 12,317 questions in 9 diverse languages with three educational levels, where about 23\% of the questions require processing images for successful solving. We assess the performance of top-performing LLMs on M3Exam and find that current models, including GPT-4, still struggle with multilingual text, particularly in low-resource and non-Latin script languages. Multimodal LLMs also perform poorly with complex multimodal questions. We believe that M3Exam can be a valuable resource for comprehensively evaluating LLMs by examining their multilingual and multimodal abilities and tracking their development. Data and evaluation code is available at \url{https://github.com/DAMO-NLP-SG/M3Exam}.

Qinyu Chen, Zuowen Wang, Shih-Chii Liu et al.

This paper presents a sparse Change-Based Convolutional Long Short-Term Memory (CB-ConvLSTM) model for event-based eye tracking, key for next-generation wearable healthcare technology such as AR/VR headsets. We leverage the benefits of retina-inspired event cameras, namely their low-latency response and sparse output event stream, over traditional frame-based cameras. Our CB-ConvLSTM architecture efficiently extracts spatio-temporal features for pupil tracking from the event stream, outperforming conventional CNN structures. Utilizing a delta-encoded recurrent path enhancing activation sparsity, CB-ConvLSTM reduces arithmetic operations by approximately 4.7$\times$ without losing accuracy when tested on a \texttt{v2e}-generated event dataset of labeled pupils. This increase in efficiency makes it ideal for real-time eye tracking in resource-constrained devices. The project code and dataset are openly available at \url{https://github.com/qinche106/cb-convlstm-eyetracking}.

Chang Gao, Bowen Li, Wenxuan Zhang et al.

Text-to-SQL parsing tackles the problem of mapping natural language questions to executable SQL queries. In practice, text-to-SQL parsers often encounter various challenging scenarios, requiring them to be generalizable and robust. While most existing work addresses a particular generalization or robustness challenge, we aim to study it in a more comprehensive manner. In specific, we believe that text-to-SQL parsers should be (1) generalizable at three levels of generalization, namely i.i.d., zero-shot, and compositional, and (2) robust against input perturbations. To enhance these capabilities of the parser, we propose a novel TKK framework consisting of Task decomposition, Knowledge acquisition, and Knowledge composition to learn text-to-SQL parsing in stages. By dividing the learning process into multiple stages, our framework improves the parser's ability to acquire general SQL knowledge instead of capturing spurious patterns, making it more generalizable and robust. Experimental results under various generalization and robustness settings show that our framework is effective in all scenarios and achieves state-of-the-art performance on the Spider, SParC, and CoSQL datasets. Code can be found at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/tkk.

Marcin Paluch, Florian Bolli, Xiang Deng et al.

Nonlinear model predictive control (NMPC) has proven to be an effective control method, but it is expensive to compute. This work demonstrates the use of hardware FPGA neural network controllers trained to imitate NMPC with supervised learning. We use these Neural Controllers (NCs) implemented on inexpensive embedded FPGA hardware for high frequency control on physical cartpole and F1TENTH race car. Our results show that the NCs match the control performance of the NMPCs in simulation and outperform it in reality, due to the faster control rate that is afforded by the quick FPGA NC inference. We demonstrate kHz control rates for a physical cartpole and offloading control to the FPGA hardware on the F1TENTH car. Code and hardware implementation for this paper are available at https:// github.com/SensorsINI/Neural-Control-Tools.

Fabrizio Ottati, Chang Gao, Qinyu Chen et al.

As deep learning models scale, they become increasingly competitive from domains spanning from computer vision to natural language processing; however, this happens at the expense of efficiency since they require increasingly more memory and computing power. The power efficiency of the biological brain outperforms any large-scale deep learning ( DL ) model; thus, neuromorphic computing tries to mimic the brain operations, such as spike-based information processing, to improve the efficiency of DL models. Despite the benefits of the brain, such as efficient information transmission, dense neuronal interconnects, and the co-location of computation and memory, the available biological substrate has severely constrained the evolution of biological brains. Electronic hardware does not have the same constraints; therefore, while modeling spiking neural networks ( SNNs) might uncover one piece of the puzzle, the design of efficient hardware backends for SNN s needs further investigation, potentially taking inspiration from the available work done on the artificial neural networks ( ANNs) side. As such, when is it wise to look at the brain while designing new hardware, and when should it be ignored? To answer this question, we quantitatively compare the digital hardware acceleration techniques and platforms of ANNs and SNN s. As a result, we provide the following insights: (i) ANNs currently process static data more efficiently, (ii) applications targeting data produced by neuromorphic sensors, such as event-based cameras and silicon cochleas, need more investigation since the behavior of these sensors might naturally fit the SNN paradigm, and (iii) hybrid approaches combining SNN s and ANNs might lead to the best solutions and should be investigated further at the hardware level, accounting for both efficiency and loss optimization.

Junyuan Ding, Ziteng Wang, Chang Gao et al.

Eye tracking is a key technology for gaze-based interactions in Extended Reality (XR), but traditional frame-based systems struggle to meet XR's demands for high accuracy, low latency, and power efficiency. Event cameras offer a promising alternative due to their high temporal resolution and low power consumption. In this paper, we present FACET (Fast and Accurate Event-based Eye Tracking), an end-to-end neural network that directly outputs pupil ellipse parameters from event data, optimized for real-time XR applications. The ellipse output can be directly used in subsequent ellipse-based pupil trackers. We enhance the EV-Eye dataset by expanding annotated data and converting original mask labels to ellipse-based annotations to train the model. Besides, a novel trigonometric loss is adopted to address angle discontinuities and a fast causal event volume event representation method is put forward. On the enhanced EV-Eye test set, FACET achieves an average pupil center error of 0.20 pixels and an inference time of 0.53 ms, reducing pixel error and inference time by 1.6$\times$ and 1.8$\times$ compared to the prior art, EV-Eye, with 4.4$\times$ and 11.7$\times$ less parameters and arithmetic operations. The code is available at https://github.com/DeanJY/FACET.

Chang Gao, Shu-Fu Shih, J. Paul Finn et al.

The recent development of deep learning combined with compressed sensing enables fast reconstruction of undersampled MR images and has achieved state-of-the-art performance for Cartesian k-space trajectories. However, non-Cartesian trajectories such as the radial trajectory need to be transformed onto a Cartesian grid in each iteration of the network training, slowing down the training process and posing inconvenience and delay during training. Multiple iterations of nonuniform Fourier transform in the networks offset the deep learning advantage of fast inference. Current approaches typically either work on image-to-image networks or grid the non-Cartesian trajectories before the network training to avoid the repeated gridding process. However, the image-to-image networks cannot ensure the k-space data consistency in the reconstructed images and the pre-processing of non-Cartesian k-space leads to gridding errors which cannot be compensated by the network training. Inspired by the Transformer network to handle long-range dependencies in sequence transduction tasks, we propose to rearrange the radial spokes to sequential data based on the chronological order of acquisition and use the Transformer to predict unacquired radial spokes from acquired ones. We propose novel data augmentation methods to generate a large amount of training data from a limited number of subjects. The network can be generated to different anatomical structures. Experimental results show superior performance of the proposed framework compared to state-of-the-art deep neural networks.

Yingxiu Zhao, Yinhe Zheng, Zhiliang Tian et al.

Lifelong learning (LL) is vital for advanced task-oriented dialogue (ToD) systems. To address the catastrophic forgetting issue of LL, generative replay methods are widely employed to consolidate past knowledge with generated pseudo samples. However, most existing generative replay methods use only a single task-specific token to control their models. This scheme is usually not strong enough to constrain the generative model due to insufficient information involved. In this paper, we propose a novel method, prompt conditioned VAE for lifelong learning (PCLL), to enhance generative replay by incorporating tasks' statistics. PCLL captures task-specific distributions with a conditional variational autoencoder, conditioned on natural language prompts to guide the pseudo-sample generation. Moreover, it leverages a distillation process to further consolidate past knowledge by alleviating the noise in pseudo samples. Experiments on natural language understanding tasks of ToD systems demonstrate that PCLL significantly outperforms competitive baselines in building LL models.

Qinyu Chen, Chang Gao, Xinyuan Fang et al.

Spiking Neural Networks (SNNs) are developed as a promising alternative to Artificial Neural networks (ANNs) due to their more realistic brain-inspired computing models. SNNs have sparse neuron firing over time, i.e., spatio-temporal sparsity; thus, they are useful to enable energy-efficient hardware inference. However, exploiting spatio-temporal sparsity of SNNs in hardware leads to unpredictable and unbalanced workloads, degrading the energy efficiency. In this work, we propose an FPGA-based convolutional SNN accelerator called Skydiver that exploits spatio-temporal workload balance. We propose the Approximate Proportional Relation Construction (APRC) method that can predict the relative workload channel-wisely and a Channel-Balanced Workload Schedule (CBWS) method to increase the hardware workload balance ratio to over 90%. Skydiver was implemented on a Xilinx XC7Z045 FPGA and verified on image segmentation and MNIST classification tasks. Results show improved throughput by 1.4X and 1.2X for the two tasks. Skydiver achieved 22.6 KFPS throughput, and 42.4 uJ/Image prediction energy on the classification task with 98.5% accuracy.

Chang Gao, Wenxuan Zhang, Wai Lam

The goal-oriented document-grounded dialogue aims at responding to the user query based on the dialogue context and supporting document. Existing studies tackle this problem by decomposing it into two sub-tasks: knowledge identification and response generation. However, such pipeline methods would unavoidably suffer from the error propagation issue. This paper proposes to unify these two sub-tasks via sequentially generating the grounding knowledge and the response. We further develop a prompt-connected multi-task learning strategy to model the characteristics and connections of different tasks and introduce linear temperature scheduling to reduce the negative effect of irrelevant document information. Experimental results demonstrate the effectiveness of our framework.

Binghai Wang, Yantao Liu, Yuxuan Liu et al.

Generative Reward Models (GenRMs) and LLM-as-a-Judge exhibit deceptive alignment by producing correct judgments for incorrect reasons, as they are trained and evaluated to prioritize Outcome Accuracy, which undermines their ability to generalize during RLHF. We introduce Rationale Consistency, a fine-grained metric that quantifies the alignment between the model's reasoning process and human judgment. Our evaluation of frontier models reveals that rationale consistency effectively discriminates among state-of-the-art models and detects deceptive alignment, while outcome accuracy falls short in both respects. To mitigate this gap, we introduce a hybrid signal that combines rationale consistency with outcome accuracy for GenRM training. Our training method achieves state-of-the-art performance on RM-Bench (87.1%) and JudgeBench (82%), surpassing outcome-only baselines by an average of 5%. Using RM during RLHF, our method effectively improves performance as demonstrated on Arena Hard v2, notably yielding a 7% improvement in creative writing tasks. Further analysis confirms that our method escapes the deceptive alignment trap, effectively reversing the decline in rationale consistency observed in outcome-only training.

Shenzhi Wang, Shixuan Liu, Jing Zhou et al.

Vision-language models (VLMs) show strong multimodal capabilities but still struggle with fine-grained vision-language reasoning. We find that long chain-of-thought (CoT) reasoning exposes diverse failure modes, including perception, reasoning, knowledge, and hallucination errors, which can compound across intermediate steps. However, most existing vision-language data used for reinforcement learning with verifiable rewards (RLVR) does not involve complex reasoning chains that rely on visual evidence throughout, leaving these weaknesses largely unexposed. We therefore propose HopChain, a scalable framework for synthesizing multi-hop vision-language reasoning data for RLVR training of VLMs. Each synthesized multi-hop query forms a logically dependent chain of instance-grounded hops, where earlier hops establish the instances, sets, or conditions needed for later hops, while the final answer remains a specific, unambiguous number suitable for verifiable rewards. We train Qwen3.5-35B-A3B and Qwen3.5-397B-A17B under two RLVR settings: the original data alone, and the original data plus HopChain's multi-hop data, and compare them across 24 benchmarks spanning STEM and Puzzle, General VQA, Text Recognition and Document Understanding, and Video Understanding. Although this multi-hop data is not synthesized for any specific benchmark, it improves 20 of 24 benchmarks on both models, indicating broad and generalizable gains. Consistently, replacing full chained queries with half-multi-hop or single-hop variants reduces the average score across five representative benchmarks from 70.4 to 66.7 and 64.3, respectively. Notably, multi-hop gains peak in long-CoT vision-language reasoning, exceeding 50 points in the ultra-long-CoT regime. These experiments establish HopChain as an effective, scalable framework for synthesizing multi-hop data that improves generalizable vision-language reasoning.

Manno Versluis, Yizhuo Wu, Chang Gao

Digital predistortion (DPD) is crucial for linearizing radio frequency (RF) power amplifiers (PAs), improving signal integrity and efficiency in wireless systems. Neural network (NN)-based DPD methods surpass traditional polynomial models but face computational challenges limiting their practical deployment. This paper introduces SparseDPD, an FPGA accelerator employing a spatially sparse phase-normalized time-delay neural network (PNTDNN), optimized through unstructured pruning to reduce computational load without accuracy loss. Implemented on a Xilinx Zynq-7Z010 FPGA, SparseDPD operates at 170 MHz, achieving exceptional linearization performance (ACPR: -59.4 dBc, EVM: -54.0 dBc, NMSE: -48.2 dB) with only 241 mW dynamic power, using 64 parameters with 74% sparsity. This work demonstrates FPGA-based acceleration, making NN-based DPD practical and efficient for real-time wireless communication applications. Code is publicly available at https://github.com/MannoVersluis/SparseDPD.

Chang Gao, Haiyun Jiang, Deng Cai et al.

Most existing prompting methods suffer from the issues of generalizability and consistency, as they often rely on instance-specific solutions that may not be applicable to other instances and lack task-level consistency across the selected few-shot examples. To address these limitations, we propose a comprehensive framework, StrategyLLM, allowing LLMs to perform inductive reasoning, deriving general strategies from specific task instances, and deductive reasoning, applying these general strategies to particular task examples, for constructing generalizable and consistent few-shot prompts. It employs four LLM-based agents: strategy generator, executor, optimizer, and evaluator, working together to generate, evaluate, and select promising strategies for a given task. Experimental results demonstrate that StrategyLLM outperforms the competitive baseline CoT-SC that requires human-annotated solutions on 13 datasets across 4 challenging tasks without human involvement, including math reasoning (34.2\% $\rightarrow$ 38.8\%), commonsense reasoning (70.3\% $\rightarrow$ 72.5\%), algorithmic reasoning (73.7\% $\rightarrow$ 85.0\%), and symbolic reasoning (30.0\% $\rightarrow$ 79.2\%). Further analysis reveals that StrategyLLM is applicable to various LLMs and demonstrates advantages across numerous scenarios.

Chang Gao, Jianfei Chen, Kang Zhao et al. · tencent-ai

Fully quantized training (FQT) accelerates the training of deep neural networks by quantizing the activations, weights, and gradients into lower precision. To explore the ultimate limit of FQT (the lowest achievable precision), we make a first attempt to 1-bit FQT. We provide a theoretical analysis of FQT based on Adam and SGD, revealing that the gradient variance influences the convergence of FQT. Building on these theoretical results, we introduce an Activation Gradient Pruning (AGP) strategy. The strategy leverages the heterogeneity of gradients by pruning less informative gradients and enhancing the numerical precision of remaining gradients to mitigate gradient variance. Additionally, we propose Sample Channel joint Quantization (SCQ), which utilizes different quantization strategies in the computation of weight gradients and activation gradients to ensure that the method is friendly to low-bitwidth hardware. Finally, we present a framework to deploy our algorithm. For fine-tuning VGGNet-16 and ResNet-18 on multiple datasets, our algorithm achieves an average accuracy improvement of approximately 6%, compared to per-sample quantization. Moreover, our training speedup can reach a maximum of 5.13x compared to full precision training.

Chengxing Xie, Bowen Li, Chang Gao et al.

Large Language Models (LLMs) have demonstrated remarkable proficiency across a variety of complex tasks. One significant application of LLMs is in tackling software engineering challenges, particularly in resolving real-world tasks on GitHub by fixing code based on the issues reported by the users. However, many current approaches rely on proprietary LLMs, which limits reproducibility, accessibility, and transparency. The critical components of LLMs for addressing software engineering issues and how their capabilities can be effectively enhanced remain unclear. To address these challenges, we introduce SWE-Fixer, a novel open-source framework designed to effectively and efficiently resolve GitHub issues. SWE-Fixer comprises two essential modules: a code file retrieval module and a code editing module. The retrieval module employs BM25 along with a lightweight model to achieve coarse-to-fine file retrieval. Subsequently, the code editing module utilizes the other model to generate patches for the identified files. To mitigate the lack of publicly available datasets, we compile an extensive dataset that includes 110K GitHub issues along with their corresponding patches and train the two models of SWE-Fixer separately. We assess our approach on the SWE-Bench Lite and Verified benchmarks, achieving competitive performance among open-source models with scores of 22.0% and 30.2%. Furthermore, SWE-Fixer reaches state-of-the-art performance (24.7% on Lite and 32.8% on Verified) with PASS_TO_PASS (P2P) filtering. Additionally, our approach requires only two model calls per instance, making it significantly more efficient than existing methods. These results highlight the effectiveness of SWE-Fixer in real-world code-fixing scenarios. We will make our model, dataset, and code publicly available at https://github.com/InternLM/SWE-Fixer.

Chang Gao, Wenxuan Zhang, Guizhen Chen et al.

Instruction tuning is vital for enhancing the performance of large language models (LLMs), but existing text-to-text methods, referred to as TextTuning, struggle with issues such as generalization, robustness, and controllability due to their lack of explicit task structures. We introduce JsonTuning, a structure-to-structure approach that uses JSON structures to represent tasks. This method improves generalization by clarifying task elements and their relations, boosts robustness by minimizing ambiguity, and enhances controllability by allowing precise control over outputs. We conduct an extensive comparative analysis between JsonTuning and TextTuning using various language models and benchmarks. Our findings reveal that JsonTuning consistently surpasses TextTuning in terms of performance, robustness, and controllability across different scenarios. By overcoming the limitations of TextTuning, JsonTuning demonstrates significant potential for developing more effective and reliable LLMs capable of handling diverse scenarios.

Jingren Hou, Hong Wang, Pengyu Xu et al.

Real-world scientific applications frequently encounter incomplete observational data due to sensor limitations, geographic constraints, or measurement costs. Although neural operators significantly advanced PDE solving in terms of computational efficiency and accuracy, their underlying assumption of fully-observed spatial inputs severely restricts applicability in real-world applications. We introduce the first systematic framework for learning neural operators from partial observation. We identify and formalize two fundamental obstacles: (i) the supervision gap in unobserved regions that prevents effective learning of physical correlations, and (ii) the dynamic spatial mismatch between incomplete inputs and complete solution fields. Specifically, our proposed Latent Autoregressive Neural Operator(LANO) introduces two novel components designed explicitly to address the core difficulties of partial observations: (i) a mask-to-predict training strategy that creates artificial supervision by strategically masking observed regions, and (ii) a Physics-Aware Latent Propagator that reconstructs solutions through boundary-first autoregressive generation in latent space. Additionally, we develop POBench-PDE, a dedicated and comprehensive benchmark designed specifically for evaluating neural operators under partial observation conditions across three PDE-governed tasks. LANO achieves state-of-the-art performance with 18--69$\%$ relative L2 error reduction across all benchmarks under patch-wise missingness with less than 50$\%$ missing rate, including real-world climate prediction. Our approach effectively addresses practical scenarios involving up to 75$\%$ missing rate, to some extent bridging the existing gap between idealized research settings and the complexities of real-world scientific computing.

An Yang, Anfeng Li, Baosong Yang et al. · tsinghua

In this work, we present Qwen3, the latest version of the Qwen model family. Qwen3 comprises a series of large language models (LLMs) designed to advance performance, efficiency, and multilingual capabilities. The Qwen3 series includes models of both dense and Mixture-of-Expert (MoE) architectures, with parameter scales ranging from 0.6 to 235 billion. A key innovation in Qwen3 is the integration of thinking mode (for complex, multi-step reasoning) and non-thinking mode (for rapid, context-driven responses) into a unified framework. This eliminates the need to switch between different models--such as chat-optimized models (e.g., GPT-4o) and dedicated reasoning models (e.g., QwQ-32B)--and enables dynamic mode switching based on user queries or chat templates. Meanwhile, Qwen3 introduces a thinking budget mechanism, allowing users to allocate computational resources adaptively during inference, thereby balancing latency and performance based on task complexity. Moreover, by leveraging the knowledge from the flagship models, we significantly reduce the computational resources required to build smaller-scale models, while ensuring their highly competitive performance. Empirical evaluations demonstrate that Qwen3 achieves state-of-the-art results across diverse benchmarks, including tasks in code generation, mathematical reasoning, agent tasks, etc., competitive against larger MoE models and proprietary models. Compared to its predecessor Qwen2.5, Qwen3 expands multilingual support from 29 to 119 languages and dialects, enhancing global accessibility through improved cross-lingual understanding and generation capabilities. To facilitate reproducibility and community-driven research and development, all Qwen3 models are publicly accessible under Apache 2.0.

Yizhuo Wu, Ang Li, Mohammadreza Beikmirza et al.

Digital Pre-Distortion (DPD) enhances signal quality in wideband RF power amplifiers (PAs). As signal bandwidths expand in modern radio systems, DPD's energy consumption increasingly impacts overall system efficiency. Deep Neural Networks (DNNs) offer promising advancements in DPD, yet their high complexity hinders their practical deployment. This paper introduces open-source mixed-precision (MP) neural networks that employ quantized low-precision fixed-point parameters for energy-efficient DPD. This approach reduces computational complexity and memory footprint, thereby lowering power consumption without compromising linearization efficacy. Applied to a 160MHz-BW 1024-QAM OFDM signal from a digital RF PA, MP-DPD gives no performance loss against 32-bit floating-point precision DPDs, while achieving -43.75 (L)/-45.27 (R) dBc in Adjacent Channel Power Ratio (ACPR) and -38.72 dB in Error Vector Magnitude (EVM). A 16-bit fixed-point-precision MP-DPD enables a 2.8X reduction in estimated inference power. The PyTorch learning and testing code is publicly available at \url{https://github.com/lab-emi/OpenDPD}.

Zhuohao Yu, Chang Gao, Wenjin Yao et al. · pku

The rapid development of large language model (LLM) evaluation methodologies and datasets has led to a profound challenge: integrating state-of-the-art evaluation techniques cost-effectively while ensuring reliability, reproducibility, and efficiency. Currently, there is a notable absence of a unified and adaptable framework that seamlessly integrates various evaluation approaches. Moreover, the reliability of evaluation findings is often questionable due to potential data contamination, with the evaluation efficiency commonly overlooked when facing the substantial costs associated with LLM inference. In response to these challenges, we introduce FreeEval, a modular and scalable framework crafted to enable trustworthy and efficient automatic evaluations of LLMs. Firstly, FreeEval's unified abstractions simplify the integration and improve the transparency of diverse evaluation methodologies, encompassing dynamic evaluation that demand sophisticated LLM interactions. Secondly, the framework integrates meta-evaluation techniques like human evaluation and data contamination detection, which, along with dynamic evaluation modules in the platform, enhance the fairness of the evaluation outcomes. Lastly, FreeEval is designed with a high-performance infrastructure, including distributed computation and caching strategies, enabling extensive evaluations across multi-node, multi-GPU clusters for open-source and proprietary LLMs.

Chang Gao, Xi Lin, Fang He et al.

This study proposes an innovative model-based modular approach (MMA) to dynamically optimize order matching and vehicle relocation in a ride-hailing platform. MMA utilizes a two-layer and modular modeling structure. The upper layer determines the spatial transfer patterns of vehicle flow within the system to maximize the total revenue of the current and future stages. With the guidance provided by the upper layer, the lower layer performs rapid vehicle-to-order matching and vehicle relocation. MMA is interpretable, and equipped with the customized and polynomial-time algorithm, which, as an online order-matching and vehicle-relocation algorithm, can scale past thousands of vehicles. We theoretically prove that the proposed algorithm can achieve the global optimum in stylized networks, while the numerical experiments based on both the toy network and realistic dataset demonstrate that MMA is capable of achieving superior systematic performance compared to batch matching and reinforcement-learning based methods. Moreover, its modular and lightweight modeling structure further enables it to achieve a high level of robustness against demand variation while maintaining a relatively low computational cost.

Yanghao Li, Rui Qian, Bowen Pan et al.

Unified multimodal Large Language Models (LLMs) that can both understand and generate visual content hold immense potential. However, existing open-source models often suffer from a performance trade-off between these capabilities. We present Manzano, a simple and scalable unified framework that substantially reduces this tension by coupling a hybrid image tokenizer with a well-curated training recipe. A single shared vision encoder feeds two lightweight adapters that produce continuous embeddings for image-to-text understanding and discrete tokens for text-to-image generation within a common semantic space. A unified autoregressive LLM predicts high-level semantics in the form of text and image tokens, with an auxiliary diffusion decoder subsequently translating the image tokens into pixels. The architecture, together with a unified training recipe over understanding and generation data, enables scalable joint learning of both capabilities. Manzano achieves state-of-the-art results among unified models, and is competitive with specialist models, particularly on text-rich evaluation. Our studies show minimal task conflicts and consistent gains from scaling model size, validating our design choice of a hybrid tokenizer.

Sjoerd Groot, Qinyu Chen, Jan C. van Gemert et al.

This paper presents CleanUMamba, a time-domain neural network architecture designed for real-time causal audio denoising directly applied to raw waveforms. CleanUMamba leverages a U-Net encoder-decoder structure, incorporating the Mamba state-space model in the bottleneck layer. By replacing conventional self-attention and LSTM mechanisms with Mamba, our architecture offers superior denoising performance while maintaining a constant memory footprint, enabling streaming operation. To enhance efficiency, we applied structured channel pruning, achieving an 8X reduction in model size without compromising audio quality. Our model demonstrates strong results in the Interspeech 2020 Deep Noise Suppression challenge. Specifically, CleanUMamba achieves a PESQ score of 2.42 and STOI of 95.1% with only 442K parameters and 468M MACs, matching or outperforming larger models in real-time performance. Code will be available at: https://github.com/lab-emi/CleanUMamba

Guorui Lu, Jing Peng, Bingyuan Huang et al.

Epileptic seizures cause abnormal brain activity, and their unpredictability can lead to accidents, underscoring the need for long-term seizure prediction. Although seizures can be predicted by analyzing electroencephalogram (EEG) signals, existing methods often require too many electrode channels or larger models, limiting mobile usability. This paper introduces a SlimSeiz framework that utilizes adaptive channel selection with a lightweight neural network model. SlimSeiz operates in two states: the first stage selects the optimal channel set for seizure prediction using machine learning algorithms, and the second stage employs a lightweight neural network based on convolution and Mamba for prediction. On the Children's Hospital Boston-MIT (CHB-MIT) EEG dataset, SlimSeiz can reduce channels from 22 to 8 while achieving a satisfactory result of 94.8% accuracy, 95.5% sensitivity, and 94.0% specificity with only 21.2K model parameters, matching or outperforming larger models' performance. We also validate SlimSeiz on a new EEG dataset, SRH-LEI, collected from Shanghai Renji Hospital, demonstrating its effectiveness across different patients. The code and SRH-LEI dataset are available at https://github.com/guoruilu/SlimSeiz.

Wenxin Dong, Chang Gao, Guanghui Yu et al.

Large language models (LLMs) have shown remarkable potential in advertising scenarios such as ad creative generation and targeted advertising. However, deploying LLMs in real-time advertising systems poses significant challenges due to their high inference latency and computational cost. In this paper, we propose an Efficient Generative Targeting framework that integrates adaptive group quantization, layer-adaptive hierarchical sparsification, and prefix-tree parallel verification to accelerate LLM inference while preserving generation quality. Extensive experiments on two real-world advertising scenarios demonstrate that our framework achieves significant speedup with acceptable quality degradation, making it operationally viable for practical deployments.

Shenzhi Wang, Le Yu, Chang Gao et al. · tsinghua

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful approach to enhancing the reasoning capabilities of Large Language Models (LLMs), while its mechanisms are not yet well understood. In this work, we undertake a pioneering exploration of RLVR through the novel perspective of token entropy patterns, comprehensively analyzing how different tokens influence reasoning performance. By examining token entropy patterns in Chain-of-Thought (CoT) reasoning, we observe that only a small fraction of tokens exhibit high entropy, and these tokens act as critical forks that steer the model toward diverse reasoning pathways. Furthermore, studying how entropy patterns evolve during RLVR training reveals that RLVR largely adheres to the base model's entropy patterns, primarily adjusting the entropy of high-entropy tokens. These findings highlight the significance of high-entropy tokens (i.e., forking tokens) to RLVR. We ultimately improve RLVR by restricting policy gradient updates to forking tokens and uncover a finding even beyond the 80/20 rule: utilizing only 20% of the tokens while maintaining performance comparable to full-gradient updates on the Qwen3-8B base model and significantly surpassing full-gradient updates on the Qwen3-32B (+11.04 on AIME'25 and +7.71 on AIME'24) and Qwen3-14B (+4.79 on AIME'25 and +5.21 on AIME'24) base models, highlighting a strong scaling trend. In contrast, training exclusively on the 80% lowest-entropy tokens leads to a marked decline in performance. These findings indicate that the efficacy of RLVR primarily arises from optimizing the high-entropy tokens that decide reasoning directions. Collectively, our results highlight the potential to understand RLVR through a token-entropy perspective and optimize RLVR by leveraging high-entropy minority tokens to further improve LLM reasoning.

Chujie Zheng, Shixuan Liu, Mingze Li et al. · tsinghua

This paper introduces Group Sequence Policy Optimization (GSPO), our stable, efficient, and performant reinforcement learning algorithm for training large language models. Unlike previous algorithms that adopt token-level importance ratios, GSPO defines the importance ratio based on sequence likelihood and performs sequence-level clipping, rewarding, and optimization. We demonstrate that GSPO achieves superior training efficiency and performance compared to the GRPO algorithm, notably stabilizes Mixture-of-Experts (MoE) RL training, and has the potential for simplifying the design of RL infrastructure. These merits of GSPO have contributed to the remarkable improvements in the latest Qwen3 models.

Yizhuo Wu, Francesco Fioranelli, Chang Gao

Radar-based human activity recognition (HAR) is attractive for unobtrusive and privacy-preserving monitoring, yet many CNN/RNN solutions remain too heavy for edge deployment, and even lightweight ViT/SSM variants often exceed practical compute and memory budgets. We introduce Neural-HAR, a dimension-gated CNN accelerator tailored for real-time radar HAR on resource-constrained platforms. At its core is GateCNN, a parameter-efficient Doppler-temporal network that (i) embeds Doppler vectors to emphasize frequency evolution over time and (ii) applies dual-path gated convolutions that modulate Doppler-aware content features with temporal gates, complemented by a residual path for stable training. On the University of Glasgow UoG2020 continuous radar dataset, GateCNN attains 86.4% accuracy with only 2.7k parameters and 0.28M FLOPs per inference, comparable to CNN-BiGRU at a fraction of the complexity. Our FPGA prototype on Xilinx Zynq-7000 Z-7007S reaches 107.5 $μ$s latency and 15 mW dynamic power using LUT-based ROM and distributed RAM only (zero DSP/BRAM), demonstrating real-time, energy-efficient edge inference. Code and HLS conversion scripts are available at https://github.com/lab-emi/AIRHAR.

Zhen Xu, Guorui Lu, Chang Gao et al.

Hand tracking holds great promise for intuitive interaction paradigms, but frame-based methods often struggle to meet the requirements of accuracy, low latency, and energy efficiency, especially in resource-constrained settings such as Extended Reality (XR) devices. Event cameras provide $μ$s-level temporal resolution at mW-level power by asynchronously sensing brightness changes. In this work, we present EvHand-FPV, a lightweight framework for egocentric First-Person-View 3D hand tracking from a single event camera. We construct an event-based FPV dataset that couples synthetic training data with 3D labels and real event data with 2D labels for evaluation to address the scarcity of egocentric benchmarks. EvHand-FPV also introduces a wrist-based region of interest (ROI) that localizes the hand region via geometric cues, combined with an end-to-end mapping strategy that embeds ROI offsets into the network to reduce computation without explicit reconstruction, and a multi-task learning strategy with an auxiliary geometric feature head that improves representations without test-time overhead. On our real FPV test set, EvHand-FPV improves 2D-AUCp from 0.77 to 0.85 while reducing parameters from 11.2M to 1.2M by 89% and FLOPs per inference from 1.648G to 0.185G by 89%. It also maintains a competitive 3D-AUCp of 0.84 on synthetic data. These results demonstrate accurate and efficient egocentric event-based hand tracking suitable for on-device XR applications. The dataset and code are available at https://github.com/zen5x5/EvHand-FPV.

Yizhuo Wu, Ang Li, Chang Gao

Neural network (NN)-based Digital Predistortion (DPD) stands out in improving signal quality in wideband radio frequency (RF) power amplifiers (PAs) employing complex modulation. However, NN DPDs usually rely on a large number of parameters for effective linearization and can significantly contribute to the energy consumption of the digital back-end in RF systems. This paper presents OpenDPDv2, a unified framework for PA modeling, DPD learning, and model optimization to reduce power consumption while maintaining high linearization performance. The optimization techniques feature a novel DPD algorithm, TRes-DeltaGRU, alongside two energy-efficient methods. The top-performing 32-bit floating-point (FP32) TRes-DeltaGRU-DPD model achieves an Adjacent Channel Power Ratio (ACPR) of -59.4 dBc and Error Vector Magnitude (EVM) of -42.1 dBc. By exploiting fixed-point quantization and dynamic temporal sparsity of input signals and hidden neurons, the inference energy of our model can be reduced by 4.5X while still maintaining -50.3 dBc ACPR and -35.2 dB EVM with 56% temporal sparsity. This was evaluated using a TM3.1a 200 MHz bandwidth 256-QAM OFDM signal applied to a 3.5 GHz GaN Doherty RF PA. OpenDPDv2 code, datasets, and documentation are publicly accessible at: https://github.com/lab-emi/OpenDPD.

Yizhuo Wu, Francesco Fioranelli, Chang Gao

Radar-based HAR has emerged as a promising alternative to conventional monitoring approaches, such as wearable devices and camera-based systems, due to its unique privacy preservation and robustness advantages. However, existing solutions based on convolutional and recurrent neural networks, although effective, are computationally demanding during deployment. This limits their applicability in scenarios with constrained resources or those requiring multiple sensors. Advanced architectures, such as Vision Transformer (ViT) and State-Space Model (SSM) architectures, offer improved modeling capabilities and have made efforts toward lightweight designs. However, their computational complexity remains relatively high. To leverage the strengths of transformer architectures while simultaneously enhancing accuracy and reducing computational complexity, this paper introduces RadMamba, a parameter-efficient, radar micro-Doppler-oriented Mamba SSM specifically tailored for radar-based HAR. Across three diverse datasets, RadMamba matches the top-performing previous model's 99.8% classification accuracy on Dataset DIAT with only 1/400 of its parameters and equals the leading models' 92.0% accuracy on Dataset CI4R with merely 1/10 of their parameters. In scenarios with continuous sequences of actions evaluated on Dataset UoG2020, RadMamba surpasses other models with significantly higher parameter counts by at least 3%, achieving this with only 6.7k parameters. Our code is available at: https://github.com/lab-emi/AIRHAR.

Yizhuo Wu, Gagan Deep Singh, Mohammadreza Beikmirza et al.

With the rise in communication capacity, deep neural networks (DNN) for digital pre-distortion (DPD) to correct non-linearity in wideband power amplifiers (PAs) have become prominent. Yet, there is a void in open-source and measurement-setup-independent platforms for fast DPD exploration and objective DPD model comparison. This paper presents an open-source framework, OpenDPD, crafted in PyTorch, with an associated dataset for PA modeling and DPD learning. We introduce a Dense Gated Recurrent Unit (DGRU)-DPD, trained via a novel end-to-end learning architecture, outperforming previous DPD models on a digital PA (DPA) in the new digital transmitter (DTX) architecture with unconventional transfer characteristics compared to analog PAs. Measurements show our DGRU-DPD achieves an ACPR of -44.69/-44.47 dBc and an EVM of -35.22 dB for 200 MHz OFDM signals. OpenDPD code, datasets, and documentation are publicly available at https://github.com/lab-emi/OpenDPD.

Chang Gao, Wenxuan Zhang, Wai Lam et al.

Information extraction (IE) systems aim to automatically extract structured information, such as named entities, relations between entities, and events, from unstructured texts. While most existing work addresses a particular IE task, universally modeling various IE tasks with one model has achieved great success recently. Despite their success, they employ a one-stage learning strategy, i.e., directly learning to extract the target structure given the input text, which contradicts the human learning process. In this paper, we propose a unified easy-to-hard learning framework consisting of three stages, i.e., the easy stage, the hard stage, and the main stage, for IE by mimicking the human learning process. By breaking down the learning process into multiple stages, our framework facilitates the model to acquire general IE task knowledge and improve its generalization ability. Extensive experiments across four IE tasks demonstrate the effectiveness of our framework. We achieve new state-of-the-art results on 13 out of 17 datasets. Our code is available at \url{https://github.com/DAMO-NLP-SG/IE-E2H}.

Qibing Ren, Chang Gao, Jing Shao et al.

The rapid advancement of Large Language Models (LLMs) has brought about remarkable generative capabilities but also raised concerns about their potential misuse. While strategies like supervised fine-tuning and reinforcement learning from human feedback have enhanced their safety, these methods primarily focus on natural languages, which may not generalize to other domains. This paper introduces CodeAttack, a framework that transforms natural language inputs into code inputs, presenting a novel environment for testing the safety generalization of LLMs. Our comprehensive studies on state-of-the-art LLMs including GPT-4, Claude-2, and Llama-2 series reveal a new and universal safety vulnerability of these models against code input: CodeAttack bypasses the safety guardrails of all models more than 80\% of the time. We find that a larger distribution gap between CodeAttack and natural language leads to weaker safety generalization, such as encoding natural language input with data structures. Furthermore, we give our hypotheses about the success of CodeAttack: the misaligned bias acquired by LLMs during code training, prioritizing code completion over avoiding the potential safety risk. Finally, we analyze potential mitigation measures. These findings highlight new safety risks in the code domain and the need for more robust safety alignment algorithms to match the code capabilities of LLMs.

Zhuohao Yu, Chang Gao, Wenjin Yao et al. · pku

Automatic evaluation methods for large language models (LLMs) are hindered by data contamination, leading to inflated assessments of their effectiveness. Existing strategies, which aim to detect contaminated texts, focus on quantifying contamination status instead of accurately gauging model performance. In this paper, we introduce KIEval, a Knowledge-grounded Interactive Evaluation framework, which incorporates an LLM-powered "interactor" role for the first time to accomplish a dynamic contamination-resilient evaluation. Starting with a question in a conventional LLM benchmark involving domain-specific knowledge, KIEval utilizes dynamically generated, multi-round, and knowledge-focused dialogues to determine whether a model's response is merely a recall of benchmark answers or demonstrates a deep comprehension to apply knowledge in more complex conversations. Extensive experiments on seven leading LLMs across five datasets validate KIEval's effectiveness and generalization. We also reveal that data contamination brings no contribution or even negative effect to models' real-world applicability and understanding, and existing contamination detection methods for LLMs can only identify contamination in pre-training but not during supervised fine-tuning.

Zuowen Wang, Chang Gao, Zongwei Wu et al.

This survey reviews the AIS 2024 Event-Based Eye Tracking (EET) Challenge. The task of the challenge focuses on processing eye movement recorded with event cameras and predicting the pupil center of the eye. The challenge emphasizes efficient eye tracking with event cameras to achieve good task accuracy and efficiency trade-off. During the challenge period, 38 participants registered for the Kaggle competition, and 8 teams submitted a challenge factsheet. The novel and diverse methods from the submitted factsheets are reviewed and analyzed in this survey to advance future event-based eye tracking research.

Qinyu Chen, Chang Gao, Min Liu et al.

This survey serves as a review for the 2025 Event-Based Eye Tracking Challenge organized as part of the 2025 CVPR event-based vision workshop. This challenge focuses on the task of predicting the pupil center by processing event camera recorded eye movement. We review and summarize the innovative methods from teams rank the top in the challenge to advance future event-based eye tracking research. In each method, accuracy, model size, and number of operations are reported. In this survey, we also discuss event-based eye tracking from the perspective of hardware design.

Rui Xie, Zhengran Zeng, Zhuohao Yu et al. · pku

Code large language models mark a pivotal breakthrough in artificial intelligence. They are specifically crafted to understand and generate programming languages, significantly boosting the efficiency of coding development workflows. In this technical report, we present CodeShell-Base, a seven billion-parameter foundation model with 8K context length, showcasing exceptional proficiency in code comprehension. By incorporating Grouped-Query Attention and Rotary Positional Embedding into GPT-2, CodeShell-Base integrates the structural merits of StarCoder and CodeLlama and forms its unique architectural design. We then carefully built a comprehensive data pre-processing process, including similar data deduplication, perplexity-based data filtering, and model-based data filtering. Through this process, We have curated 100 billion high-quality pre-training data from GitHub. Benefiting from the high-quality data, CodeShell-Base outperforms CodeLlama in Humaneval after training on just 500 billion tokens (5 epochs). We have conducted extensive experiments across multiple language datasets, including Python, Java, and C++, and the results indicate that our model possesses robust foundational capabilities in code comprehension and generation.

Ethan Li, Anders Boesen Lindbo Larsen, Chen Zhang et al. · apple-ml, cmu

We introduce two multilingual, multimodal foundation language models that power Apple Intelligence features across Apple devices and services: i a 3B-parameter on-device model optimized for Apple silicon through architectural innovations such as KV-cache sharing and 2-bit quantization-aware training; and ii a scalable server model built on a novel Parallel-Track Mixture-of-Experts PT-MoE transformer that combines track parallelism, mixture-of-experts sparse computation, and interleaved global-local attention to deliver high quality with competitive cost on Apple's Private Cloud Compute platform. Both models are trained on large-scale multilingual and multimodal datasets sourced via responsible web crawling, licensed corpora, and high-quality synthetic data, then further refined with supervised fine-tuning and reinforcement learning on a new asynchronous platform. The resulting models support several additional languages while understanding images and executing tool calls. In public benchmarks and human evaluations, both the server model and the on-device model match or surpass comparably sized open baselines. A new Swift-centric Foundation Models framework exposes guided generation, constrained tool calling, and LoRA adapter fine-tuning, allowing developers to integrate these capabilities with a few lines of code. The latest advancements in Apple Intelligence models are grounded in our Responsible AI approach with safeguards like content filtering and locale-specific evaluation, as well as our commitment to protecting our users' privacy with innovations like Private Cloud Compute.

Ang Li, Haolin Wu, Yizhuo Wu et al.

The increasing adoption of Deep Neural Network (DNN)-based Digital Pre-distortion (DPD) in modern communication systems necessitates efficient hardware implementations. This paper presents DPD-NeuralEngine, an ultra-fast, tiny-area, and power-efficient DPD accelerator based on a Gated Recurrent Unit (GRU) neural network (NN). Leveraging a co-designed software and hardware approach, our 22 nm CMOS implementation operates at 2 GHz, capable of processing I/Q signals up to 250 MSps. Experimental results demonstrate a throughput of 256.5 GOPS and power efficiency of 1.32 TOPS/W with DPD linearization performance measured in Adjacent Channel Power Ratio (ACPR) of -45.3 dBc and Error Vector Magnitude (EVM) of -39.8 dB. To our knowledge, this work represents the first AI-based DPD application-specific integrated circuit (ASIC) accelerator, achieving a power-area efficiency (PAE) of 6.6 TOPS/W/mm$^2$.

Xi Chen, Chang Gao, Zuowen Wang et al.

Recurrent Neural Networks (RNNs) are useful in temporal sequence tasks. However, training RNNs involves dense matrix multiplications which require hardware that can support a large number of arithmetic operations and memory accesses. Implementing online training of RNNs on the edge calls for optimized algorithms for an efficient deployment on hardware. Inspired by the spiking neuron model, the Delta RNN exploits temporal sparsity during inference by skipping over the update of hidden states from those inactivated neurons whose change of activation across two timesteps is below a defined threshold. This work describes a training algorithm for Delta RNNs that exploits temporal sparsity in the backward propagation phase to reduce computational requirements for training on the edge. Due to the symmetric computation graphs of forward and backward propagation during training, the gradient computation of inactivated neurons can be skipped. Results show a reduction of $\sim$80% in matrix operations for training a 56k parameter Delta LSTM on the Fluent Speech Commands dataset with negligible accuracy loss. Logic simulations of a hardware accelerator designed for the training algorithm show 2-10X speedup in matrix computations for an activation sparsity range of 50%-90%. Additionally, we show that the proposed Delta RNN training will be useful for online incremental learning on edge devices with limited computing resources.

Yizhuo Wu, Yi Zhu, Kun Qian et al.

Digital Predistortion (DPD) is a popular technique to enhance signal quality in wideband RF power amplifiers (PAs). With increasing bandwidth and data rates, DPD faces significant energy consumption challenges during deployment, contrasting with its efficiency goals. State-of-the-art DPD models rely on recurrent neural networks (RNN), whose computational complexity hinders system efficiency. This paper introduces DeltaDPD, exploring the dynamic temporal sparsity of input signals and neuronal hidden states in RNNs for energy-efficient DPD, reducing arithmetic operations and memory accesses while preserving satisfactory linearization performance. Applying a TM3.1a 200MHz-BW 256-QAM OFDM signal to a 3.5 GHz GaN Doherty RF PA, DeltaDPD achieves -50.03 dBc in Adjacent Channel Power Ratio (ACPR), -37.22 dB in Normalized Mean Square Error (NMSE) and -38.52 dBc in Error Vector Magnitude (EVM) with 52% temporal sparsity, leading to a 1.8X reduction in estimated inference power. The DeltaDPD code will be released after formal publication at https://www.opendpd.com.

Kang Zhao, Tao Yuan, Han Bao et al.

To date, 2:4 sparsity has stood as the only sparse pattern that can be accelerated using sparse tensor cores on GPUs. In practice, 2:4 sparsity often possesses low actual speedups ($\leq 1.3$) and requires fixed sparse ratios, meaning that other ratios, such as 4:8, 8:16, or those exceeding 50% sparsity, do not incur any speedups on GPUs. Recent studies suggest that V:N:M sparsity is promising in addressing these limitations of 2:4 sparsity. However, regarding accuracy, the effects of V:N:M sparsity on broader Transformer models, such as vision Transformers and large language models (LLMs), are largely unexamined. Moreover, Some specific issues related to V:N:M sparsity, such as how to select appropriate V and M values, remain unresolved. In this study, we thoroughly investigate the application of V:N:M sparsity in vision models and LLMs across multiple tasks, from pertaining to downstream tasks. We propose three key approaches to enhance the applicability and accuracy of V:N:M-sparse Transformers, including heuristic V and M selection, V:N:M-specific channel permutation, and three-staged LoRA training techniques. Experimental results show that, with our methods, the DeiT-small achieves lossless accuracy at 64:2:5 sparsity, while the DeiT-base maintains accuracy even at 64:2:8 sparsity. In addition, the fine-tuned LLama2-7B at 64:2:5 sparsity performs comparably or better than training-free 2:4 sparse alternatives on downstream tasks. More importantly, V:N:M-sparse Transformers offer a wider range of speedup-accuracy trade-offs compared to 2:4 sparsity. Overall, our exploration largely facilitates the V:N:M sparsity to act as a truly effective acceleration solution for Transformers in cost-sensitive inference scenarios.

Jiawen Qi, Chang Gao, Zhaochun Ren et al.

Deploying Large Language Models (LLMs) on edge devices remains challenging due to their quadratically increasing computations with the sequence length. Existing studies for dynamic attention pruning are designed for hardware with massively parallel computation capabilities, such as GPUs or TPUs, and aim at long context lengths (e.g., 64K), making them unsuitable for edge scenarios. We present DeltaLLM, a training-free framework that exploits temporal sparsity in attention patterns to enable efficient LLM inference across both the prefilling and decoding stages, on resource-constrained edge devices. DeltaLLM introduces an accuracy- and memory-aware delta matrix construction strategy that introduces temporal sparsity, and a context-aware hybrid attention mechanism that combines full attention in a local context window with delta approximation outside it to increase accuracy. We evaluate our framework on the edge-device-friendly BitNet-b1.58-2B-4T model and Llama3.2-1B-Instruct model across diverse language tasks. The results show that on BitNet, our framework increases the attention sparsity from 0% to 60% during the prefilling stage with slight accuracy improvement on the WG task, and 0% to 57% across both the prefilling and decoding stages, with even higher F1 score from 29.63 to 30.97 on SQuAD-v2 task. On the Llama model, it can also achieve up to 60% sparsity during the prefilling stage and around 57% across both stages with negligible accuracy drop. These results demonstrate that DeltaLLM offers a promising solution for efficient edge deployment, requiring no fine-tuning and seamlessly integrating with existing inference pipelines.

Huanqiang Duan, Manno Versluis, Qinyu Chen et al.

Digital predistortion (DPD) is essential for mitigating nonlinearity in RF power amplifiers, particularly for wideband applications. This paper presents TCN-DPD, a parameter-efficient architecture based on temporal convolutional networks, integrating noncausal dilated convolutions with optimized activation functions. Evaluated on the OpenDPD framework with the DPA_200MHz dataset, TCN-DPD achieves simulated ACPRs of -51.58/-49.26 dBc (L/R), EVM of -47.52 dB, and NMSE of -44.61 dB with 500 parameters and maintains superior linearization than prior models down to 200 parameters, making it promising for efficient wideband PA linearization.

Shuai Bai, Yuxuan Cai, Ruizhe Chen et al.

We introduce Qwen3-VL, the most capable vision-language model in the Qwen series to date, achieving superior performance across a broad range of multimodal benchmarks. It natively supports interleaved contexts of up to 256K tokens, seamlessly integrating text, images, and video. The model family includes both dense (2B/4B/8B/32B) and mixture-of-experts (30B-A3B/235B-A22B) variants to accommodate diverse latency-quality trade-offs. Qwen3-VL delivers three core pillars: (i) markedly stronger pure-text understanding, surpassing comparable text-only backbones in several cases; (ii) robust long-context comprehension with a native 256K-token window for both text and interleaved multimodal inputs, enabling faithful retention, retrieval, and cross-referencing across long documents and videos; and (iii) advanced multimodal reasoning across single-image, multi-image, and video tasks, demonstrating leading performance on comprehensive evaluations such as MMMU and visual-math benchmarks (e.g., MathVista and MathVision). Architecturally, we introduce three key upgrades: (i) an enhanced interleaved-MRoPE for stronger spatial-temporal modeling across images and video; (ii) DeepStack integration, which effectively leverages multi-level ViT features to tighten vision-language alignment; and (iii) text-based time alignment for video, evolving from T-RoPE to explicit textual timestamp alignment for more precise temporal grounding. Under comparable token budgets and latency constraints, Qwen3-VL achieves superior performance in both dense and Mixture-of-Experts (MoE) architectures. We envision Qwen3-VL serving as a foundational engine for image-grounded reasoning, agentic decision-making, and multimodal code intelligence in real-world workflows.

Chang Gao, Chujie Zheng, Xiong-Hui Chen et al.

Reinforcement learning (RL) plays an increasingly important role in enhancing the reasoning capabilities of large language models (LLMs), yet stable and performant policy optimization remains challenging. Token-level importance ratios often exhibit high variance-a phenomenon exacerbated in Mixture-of-Experts models-leading to unstable updates. Existing group-based policy optimization methods, such as GSPO and GRPO, alleviate this problem via hard clipping, making it difficult to maintain both stability and effective learning. We propose Soft Adaptive Policy Optimization (SAPO), which replaces hard clipping with a smooth, temperature-controlled gate that adaptively attenuates off-policy updates while preserving useful learning signals. Compared with GSPO and GRPO, SAPO is both sequence-coherent and token-adaptive. Like GSPO, SAPO maintains sequence-level coherence, but its soft gating forms a continuous trust region that avoids the brittle hard clipping band used in GSPO. When a sequence contains a few highly off-policy tokens, GSPO suppresses all gradients for that sequence, whereas SAPO selectively down-weights only the offending tokens and preserves the learning signal from the near-on-policy ones, improving sample efficiency. Relative to GRPO, SAPO replaces hard token-level clipping with smooth, temperature-controlled scaling, enabling more informative and stable updates. Empirical results on mathematical reasoning benchmarks indicate that SAPO exhibits improved training stability and higher Pass@1 performance under comparable training budgets. Moreover, we employ SAPO to train the Qwen3-VL model series, demonstrating that SAPO yields consistent performance gains across diverse tasks and different model sizes. Overall, SAPO provides a more reliable, scalable, and effective optimization strategy for RL training of LLMs.

He Du, Bowen Li, Chengxing Xie et al.

Reward models can significantly enhance the reasoning capabilities of large language models (LLMs), but they typically require extensive curated data and costly training. To mitigate these challenges, training-free approaches such as LLM-as-a-Judge leverage the intrinsic reasoning abilities of LLMs to evaluate responses, achieving promising results. Recent works have also indicated that model confidence can serve effectively as a reward metric, distinguishing between chain-of-thought (CoT) and non-CoT paths. However, the concept of using confidence as a reward has not been comprehensively studied. In this work, we systematically investigate Confidence-as-a-Reward (CRew), a simple yet powerful training-free method that utilizes token-level confidence in the model's final answers as a proxy for reward, especially suitable for close-ended tasks. Through extensive experiments on mathematical reasoning tasks, we demonstrate that CRew outperforms existing training-free reward approaches on the MATH500 and RewardMATH benchmarks, and even surpasses most trained reward models. We further identify a strong correlation between CRew scores and the actual reasoning performance of the model. Additionally, we find that CRew can effectively filter high-quality training data. Building upon these insights, we propose CRew-DPO, a training strategy that constructs preference data from confidence scores combined with correctness signals. Finetuning with CRew-DPO further enhances the model's judging capabilities and consistently outperforms existing self-training methods.