Xing Chen

Jingtao Li, Adnan Siraj Rakin, Xing Chen et al.

This work aims to tackle Model Inversion (MI) attack on Split Federated Learning (SFL). SFL is a recent distributed training scheme where multiple clients send intermediate activations (i.e., feature map), instead of raw data, to a central server. While such a scheme helps reduce the computational load at the client end, it opens itself to reconstruction of raw data from intermediate activation by the server. Existing works on protecting SFL only consider inference and do not handle attacks during training. So we propose ResSFL, a Split Federated Learning Framework that is designed to be MI-resistant during training. It is based on deriving a resistant feature extractor via attacker-aware training, and using this extractor to initialize the client-side model prior to standard SFL training. Such a method helps in reducing the computational complexity due to use of strong inversion model in client-side adversarial training as well as vulnerability of attacks launched in early training epochs. On CIFAR-100 dataset, our proposed framework successfully mitigates MI attack on a VGG-11 model with a high reconstruction Mean-Square-Error of 0.050 compared to 0.005 obtained by the baseline system. The framework achieves 67.5% accuracy (only 1% accuracy drop) with very low computation overhead. Code is released at: https://github.com/zlijingtao/ResSFL.

Xing Chen, Dongcui Diao, Hechang Chen et al.

The popular Proximal Policy Optimization (PPO) algorithm approximates the solution in a clipped policy space. Does there exist better policies outside of this space? By using a novel surrogate objective that employs the sigmoid function (which provides an interesting way of exploration), we found that the answer is ``YES'', and the better policies are in fact located very far from the clipped space. We show that PPO is insufficient in ``off-policyness'', according to an off-policy metric called DEON. Our algorithm explores in a much larger policy space than PPO, and it maximizes the Conservative Policy Iteration (CPI) objective better than PPO during training. To the best of our knowledge, all current PPO methods have the clipping operation and optimize in the clipped policy space. Our method is the first of this kind, which advances the understanding of CPI optimization and policy gradient methods. Code is available at https://github.com/raincchio/P3O.

Han Zhu, Chengkun Cai, Yuanfeng Song et al.

Large Language Models (LLMs) have become increasingly adopted in daily applications, with deep research standing out as a particularly important capability. Unlike traditional question-answering (QA) tasks, deep research report generation lacks definitive ground-truth, making reward design inherently unverifiable and limiting effective reinforcement learning. Existing approaches mitigate this challenge with LLM-as-a-judge and query-dependent evaluation rubrics, but they still rely on static evaluators that cannot adapt their standards as the solver improves, leading to insufficient and eventually saturated optimization pressure. We address this limitation with a \textbf{s}elf-evolving \textbf{co}-evolutionary training framework for deep \textbf{re}search evaluation and generation (SCORE), which tightly couples an evaluator and a solver in a shared-parameter learning process. Rather than treating generation and evaluation as isolated modules, we leverage their intrinsic connection to enable joint improvement within a single shared-parameter model. To restrict this process, we introduce a meta-harness, which dynamically controls the evaluation environment based on solver performance, encouraging valid evaluation dimensions and sufficiently deep evaluator search. Extensive experiments on deep research benchmarks demonstrate consistent improvement in report generation quality, showing that co-evolving evaluation and generation is a promising direction for training open-ended research agents.

Ailin Huang, Chengyuan Yao, Chunrui Han et al.

We present STEP3-VL-10B, a lightweight open-source foundation model designed to redefine the trade-off between compact efficiency and frontier-level multimodal intelligence. STEP3-VL-10B is realized through two strategic shifts: first, a unified, fully unfrozen pre-training strategy on 1.2T multimodal tokens that integrates a language-aligned Perception Encoder with a Qwen3-8B decoder to establish intrinsic vision-language synergy; and second, a scaled post-training pipeline featuring over 1k iterations of reinforcement learning. Crucially, we implement Parallel Coordinated Reasoning (PaCoRe) to scale test-time compute, allocating resources to scalable perceptual reasoning that explores and synthesizes diverse visual hypotheses. Consequently, despite its compact 10B footprint, STEP3-VL-10B rivals or surpasses models 10$\times$-20$\times$ larger (e.g., GLM-4.6V-106B, Qwen3-VL-235B) and top-tier proprietary flagships like Gemini 2.5 Pro and Seed-1.5-VL. Delivering best-in-class performance, it records 92.2% on MMBench and 80.11% on MMMU, while excelling in complex reasoning with 94.43% on AIME2025 and 75.95% on MathVision. We release the full model suite to provide the community with a powerful, efficient, and reproducible baseline.

Jingtao Li, Jian Zhou, Yan Xiong et al.

Sampling is an essential part of raw point cloud data processing such as in the popular PointNet++ scheme. Farthest Point Sampling (FPS), which iteratively samples the farthest point and performs distance updating, is one of the most popular sampling schemes. Unfortunately it suffers from low efficiency and can become the bottleneck of point cloud applications. We propose adjustable FPS (AFPS), parameterized by M, to aggressively reduce the complexity of FPS without compromising on the sampling performance. Specifically, it divides the original point cloud into M small point clouds and samples M points simultaneously. It exploits the dimensional locality of an approximately sorted point cloud data to minimize its performance degradation. AFPS method can achieve 22 to 30x speedup over original FPS. Furthermore, we propose the nearest-point-distance-updating (NPDU) method to limit the number of distance updates to a constant number. The combined NPDU on AFPS method can achieve a 34-280x speedup on a point cloud with 2K-32K points with algorithmic performance that is comparable to the original FPS. For instance, for the ShapeNet part segmentation task, it achieves 0.8490 instance average mIoU (mean Intersection of Union), which is only 0.0035 drop compared to the original FPS.

Jingtao Li, Adnan Siraj Rakin, Xing Chen et al.

Federated Learning (FL) is a popular collaborative learning scheme involving multiple clients and a server. FL focuses on protecting clients' data but turns out to be highly vulnerable to Intellectual Property (IP) threats. Since FL periodically collects and distributes the model parameters, a free-rider can download the latest model and thus steal model IP. Split Federated Learning (SFL), a recent variant of FL that supports training with resource-constrained clients, splits the model into two, giving one part of the model to clients (client-side model), and the remaining part to the server (server-side model). Thus SFL prevents model leakage by design. Moreover, by blocking prediction queries, it can be made resistant to advanced IP threats such as traditional Model Extraction (ME) attacks. While SFL is better than FL in terms of providing IP protection, it is still vulnerable. In this paper, we expose the vulnerability of SFL and show how malicious clients can launch ME attacks by querying the gradient information from the server side. We propose five variants of ME attack which differs in the gradient usage as well as in the data assumptions. We show that under practical cases, the proposed ME attacks work exceptionally well for SFL. For instance, when the server-side model has five layers, our proposed ME attack can achieve over 90% accuracy with less than 2% accuracy degradation with VGG-11 on CIFAR-10.

Ailin Huang, Ang Li, Aobo Kong et al.

We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.

Jiadi Yao, Xing Chen, Xiao-Lei Zhang et al.

Adversarial attack approaches to speaker identification either need high computational cost or are not very effective, to our knowledge. To address this issue, in this paper, we propose a novel generation-network-based approach, called symmetric saliency-based encoder-decoder (SSED), to generate adversarial voice examples to speaker identification. It contains two novel components. First, it uses a novel saliency map decoder to learn the importance of speech samples to the decision of a targeted speaker identification system, so as to make the attacker focus on generating artificial noise to the important samples. It also proposes an angular loss function to push the speaker embedding far away from the source speaker. Our experimental results demonstrate that the proposed SSED yields the state-of-the-art performance, i.e. over 97% targeted attack success rate and a signal-to-noise level of over 39 dB on both the open-set and close-set speaker identification tasks, with a low computational cost.

Xing Chen, Jie Wang, Xiao-Lei Zhang et al.

Although the security of automatic speaker verification (ASV) is seriously threatened by recently emerged adversarial attacks, there have been some countermeasures to alleviate the threat. However, many defense approaches not only require the prior knowledge of the attackers but also possess weak interpretability. To address this issue, in this paper, we propose an attacker-independent and interpretable method, named learnable mask detector (LMD), to separate adversarial examples from the genuine ones. It utilizes score variation as an indicator to detect adversarial examples, where the score variation is the absolute discrepancy between the ASV scores of an original audio recording and its transformed audio synthesized from its masked complex spectrogram. A core component of the score variation detector is to generate the masked spectrogram by a neural network. The neural network needs only genuine examples for training, which makes it an attacker-independent approach. Its interpretability lies that the neural network is trained to minimize the score variation of the targeted ASV, and maximize the number of the masked spectrogram bins of the genuine training examples. Its foundation is based on the observation that, masking out the vast majority of the spectrogram bins with little speaker information will inevitably introduce a large score variation to the adversarial example, and a small score variation to the genuine example. Experimental results with 12 attackers and two representative ASV systems show that our proposed method outperforms five state-of-the-art baselines. The extensive experimental results can also be a benchmark for the detection-based ASV defenses.

Guoqing Ma, Haoyang Huang, Kun Yan et al.

We present Step-Video-T2V, a state-of-the-art text-to-video pre-trained model with 30B parameters and the ability to generate videos up to 204 frames in length. A deep compression Variational Autoencoder, Video-VAE, is designed for video generation tasks, achieving 16x16 spatial and 8x temporal compression ratios, while maintaining exceptional video reconstruction quality. User prompts are encoded using two bilingual text encoders to handle both English and Chinese. A DiT with 3D full attention is trained using Flow Matching and is employed to denoise input noise into latent frames. A video-based DPO approach, Video-DPO, is applied to reduce artifacts and improve the visual quality of the generated videos. We also detail our training strategies and share key observations and insights. Step-Video-T2V's performance is evaluated on a novel video generation benchmark, Step-Video-T2V-Eval, demonstrating its state-of-the-art text-to-video quality when compared with both open-source and commercial engines. Additionally, we discuss the limitations of current diffusion-based model paradigm and outline future directions for video foundation models. We make both Step-Video-T2V and Step-Video-T2V-Eval available at https://github.com/stepfun-ai/Step-Video-T2V. The online version can be accessed from https://yuewen.cn/videos as well. Our goal is to accelerate the innovation of video foundation models and empower video content creators.

Xing Chen, Dongshu Liu, Jeremie Laydevant et al.

Agents that operate autonomously benefit from lifelong learning capabilities. However, compatible training algorithms must comply with the decentralized nature of these systems, which imposes constraints on both the parameter counts and the computational resources. The Forward-Forward (FF) algorithm is one of these. FF relies only on feedforward operations, the same used for inference, for optimizing layer-wise objectives. This purely forward approach eliminates the need for transpose operations required in traditional backpropagation. Despite its potential, FF has failed to reach state-of-the-art performance on most standard benchmark tasks, in part due to unreliable negative data generation methods for unsupervised learning. In this work, we propose the Self-Contrastive Forward-Forward (SCFF) algorithm, a competitive training method aimed at closing this performance gap. Inspired by standard self-supervised contrastive learning for vision tasks, SCFF generates positive and negative inputs applicable across various datasets. The method demonstrates superior performance compared to existing unsupervised local learning algorithms on several benchmark datasets, including MNIST, CIFAR-10, STL-10, and Tiny ImageNet. We extend FF's application to training recurrent neural networks, expanding its utility to sequential data tasks. These findings pave the way for high-accuracy, real-time learning on resource-constrained edge devices.

Hengshuai Yao, Xing Chen, Ahmed Murtadha et al.

Standard transformer attention uses identical dimensionality for queries, keys, and values, yet these components serve different roles: queries and keys produce scalar attention weights (selection), while values carry rich representations (value transfer). We show that selection requires only $O(\log N)$ dimensions to distinguish among $N$ relevant token categories (e.g., syntactic roles, semantic clusters, positional patterns) -- far fewer than value transfer needs. We introduce factored keys, which exploit this asymmetry to physically shrink the KV cache of any pretrained model without retraining from scratch -- unlike GQA and MLA, which must be designed into the architecture before pretraining. We factorize each key projection $W_K \approx A_{d \times r} B_{r \times d}$ via truncated SVD (where $r = d_{\text{select}}$), set $W_K' = A$ as the new key projection producing compact $r$-dimensional keys for the cache, and absorb $B^\top$ into the query projection ($W_Q' = W_Q B^\top$) at zero cost -- since queries are never cached. At 7B scale, training from scratch with $r = d_{\text{model}}/4$ matches full-attention perplexity (9.2 vs 9.3 PPL after 20B tokens) while using 12% fewer parameters and training 8% faster. For existing models, SVD + QK fine-tuning (3 epochs, less than 1% of pretraining data) achieves 75% key cache savings at approximately 2% quality cost on both GPT-2 and Mistral-7B. The approach composes with GQA and quantization for up to $16\times$ combined key cache compression. For a 7B model serving 128K context, factored keys save 25 GB of KV cache per user, enabling approximately 60% more concurrent users on identical hardware.

Luying Zhong, Yueyang Pi, Zheyi Chen et al.

Federated Graph Learning (FGL) has garnered widespread attention by enabling collaborative training on multiple clients for semi-supervised classification tasks. However, most existing FGL studies do not well consider the missing inter-client topology information in real-world scenarios, causing insufficient feature aggregation of multi-hop neighbor clients during model training. Moreover, the classic FGL commonly adopts the FedAvg but neglects the high training costs when the number of clients expands, resulting in the overload of a single edge server. To address these important challenges, we propose a novel FGL framework, named SpreadFGL, to promote the information flow in edge-client collaboration and extract more generalized potential relationships between clients. In SpreadFGL, an adaptive graph imputation generator incorporated with a versatile assessor is first designed to exploit the potential links between subgraphs, without sharing raw data. Next, a new negative sampling mechanism is developed to make SpreadFGL concentrate on more refined information in downstream tasks. To facilitate load balancing at the edge layer, SpreadFGL follows a distributed training manner that enables fast model convergence. Using real-world testbed and benchmark graph datasets, extensive experiments demonstrate the effectiveness of the proposed SpreadFGL. The results show that SpreadFGL achieves higher accuracy and faster convergence against state-of-the-art algorithms.

Xing Chen, Yijun Liu, Zhaogeng Liu et al.

Exploration strategies in continuous action space are often heuristic due to the infinite actions, and these kinds of methods cannot derive a general conclusion. In prior work, it has been shown that policy-based exploration is beneficial for continuous action space in deterministic policy reinforcement learning(DPRL). However, policy-based exploration in DPRL has two prominent issues: aimless exploration and policy divergence, and the policy gradient for exploration is only sometimes helpful due to inaccurate estimation. Based on the double-Q function framework, we introduce a novel exploration strategy to mitigate these issues, separate from the policy gradient. We first propose the greedy Q softmax update schema for Q value update. The expected Q value is derived by weighted summing the conservative Q value over actions, and the weight is the corresponding greedy Q value. Greedy Q takes the maximum value of the two Q functions, and conservative Q takes the minimum value of the two different Q functions. For practicality, this theoretical basis is then extended to allow us to combine action exploration with the Q value update, except for the premise that we have a surrogate policy that behaves like this exploration policy. In practice, we construct such an exploration policy with a few sampled actions, and to meet the premise, we learn such a surrogate policy by minimizing the KL divergence between the target policy and the exploration policy constructed by the conservative Q. We evaluate our method on the Mujoco benchmark and demonstrate superior performance compared to previous state-of-the-art methods across various environments, particularly in the most complex Humanoid environment.

Haoyang Huang, Guoqing Ma, Nan Duan et al.

We present Step-Video-TI2V, a state-of-the-art text-driven image-to-video generation model with 30B parameters, capable of generating videos up to 102 frames based on both text and image inputs. We build Step-Video-TI2V-Eval as a new benchmark for the text-driven image-to-video task and compare Step-Video-TI2V with open-source and commercial TI2V engines using this dataset. Experimental results demonstrate the state-of-the-art performance of Step-Video-TI2V in the image-to-video generation task. Both Step-Video-TI2V and Step-Video-TI2V-Eval are available at https://github.com/stepfun-ai/Step-Video-TI2V.

Zhou Liu, Zhaoyang Han, Guochen Yan et al.

Data governance ensures data quality, security, and compliance through policies and standards, a critical foundation for scaling modern AI development. Recently, large language models (LLMs) have emerged as a promising solution for automating data governance by translating user intent into executable transformation code. However, existing benchmarks for automated data science often emphasize snippet-level coding or high-level analytics, failing to capture the unique challenge of data governance: ensuring the correctness and quality of the data itself. To bridge this gap, we introduce DataGovBench, a benchmark featuring 150 diverse tasks grounded in real-world scenarios, built on data from actual cases. DataGovBench employs a novel "reversed-objective" methodology to synthesize realistic noise and utilizes rigorous metrics to assess end-to-end pipeline reliability. Our analysis on DataGovBench reveals that current models struggle with complex, multi-step workflows and lack robust error-correction mechanisms. Consequently, we propose DataGovAgent, a framework utilizing a Planner-Executor-Evaluator architecture that integrates constraint-based planning, retrieval-augmented generation, and sandboxed feedback-driven debugging. Experimental results show that DataGovAgent significantly boosts the Average Task Score (ATS) on complex tasks from 39.7 to 54.9 and reduces debugging iterations by over 77.9 percent compared to general-purpose baselines.

Zhenghao Zhu, Chuxue Cao, Sirui Han et al.

In medical data analysis, extracting deep insights from complex, multi-modal datasets is essential for improving patient care, increasing diagnostic accuracy, and optimizing healthcare operations. However, there is currently a lack of high-quality datasets specifically designed to evaluate the ability of large multi-modal models (LMMs) to discover medical insights. In this paper, we introduce MedInsightBench, the first benchmark that comprises 332 carefully curated medical cases, each annotated with thoughtfully designed insights. This benchmark is intended to evaluate the ability of LMMs and agent frameworks to analyze multi-modal medical image data, including posing relevant questions, interpreting complex findings, and synthesizing actionable insights and recommendations. Our analysis indicates that existing LMMs exhibit limited performance on MedInsightBench, which is primarily attributed to their challenges in extracting multi-step, deep insights and the absence of medical expertise. Therefore, we propose MedInsightAgent, an automated agent framework for medical data analysis, composed of three modules: Visual Root Finder, Analytical Insight Agent, and Follow-up Question Composer. Experiments on MedInsightBench highlight pervasive challenges and demonstrate that MedInsightAgent can improve the performance of general LMMs in medical data insight discovery.

Shuai Shao, Yuhao Sun, Xing Chen et al.

Large-scale sparse multi-objective optimization problems (LSMOPs) are prevalent in real-world applications, where optimal solutions typically contain only a few nonzero variables, such as in adversarial attacks, critical node detection, and sparse signal reconstruction. Since the function evaluation of LSMOPs often relies on large-scale datasets involving a large number of decision variables, the search space becomes extremely high-dimensional. The coexistence of sparsity and high dimensionality greatly intensifies the conflict between exploration and exploitation, making it difficult for existing multi-objective evolutionary algorithms (MOEAs) to identify the critical nonzero decision variables within limited function evaluations. To address this challenge, this paper proposes an evolutionary algorithm with probabilistic annealing for large-scale sparse multi-objective optimization. The algorithm is driven by two probability vectors with distinct entropy characteristics: a convergence-oriented probability vector with relatively low entropy ensures stable exploitation, whereas an annealed probability vector with gradually decreasing entropy enables an adaptive transition from global exploration to local refinement. By integrating these complementary search dynamics, the proposed algorithm achieves a dynamic equilibrium between exploration and exploitation. Experimental results on benchmark problems and real-world applications demonstrate that the proposed algorithm outperforms state-of-the-art evolutionary algorithms in terms of both convergence and diversity.

Zishu Yao, Xiang-Xiang Su, Shengning Zhou et al.

Event cameras, with their high dynamic range, show great promise for Low-light Image Enhancement (LLIE). Existing works primarily focus on designing effective modal fusion strategies. However, a key challenge is the dual degradation from intrinsic background activity (BA) noise in events and low signal-to-noise ratio (SNR) in images, which causes severe noise coupling during modal fusion, creating a critical performance bottleneck. We therefore posit that precise event denoising is the prerequisite to unlocking the full potential of event-based fusion. To this end, we propose BiEvLight, a hierarchical and task-aware framework that collaboratively optimizes enhancement and denoising by exploiting their intrinsic interdependence. Specifically, BiEvLight exploits the strong gradient correlation between images and events to build a gradient-guided event denoising prior that alleviates insufficient denoising in heavily noisy regions. Moreover, instead of treating event denoising as a static pre-processing stage-which inevitably incurs a trade-off between over- and under-denoising and cannot adapt to the requirements of a specific enhancement objective-we recast it as a bilevel optimization problem constrained by the enhancement task. Through cross-task interaction, the upper-level denoising problem learns event representations tailored to the lower-level enhancement objective, thereby substantially improving overall enhancement quality. Extensive experiments on the Real-world noise Dataset SDE demonstrate that our method significantly outperforms state-of-the-art (SOTA) approaches, with average improvements of 1.30dB in PSNR, 2.03dB in PSNR* and 0.047 in SSIM, respectively. The code will be publicly available at https://github.com/iijjlk/BiEvlight.

Hengshuai Yao, Xing Chen, Ahmed Murtadha et al.

We introduce Focus, a method that learns which token pairs matter rather than approximating all of them. Learnable centroids assign tokens to groups; distant attention is restricted to same-group pairs while local attention operates at full resolution. Because all model weights stay frozen, Focus is purely additive: centroid-only training (as few as 148K parameters) improves domain perplexity with zero degradation on downstream benchmarks--from 124M to 70B parameters, across five attention architectures. No existing efficient attention method achieves this in the retrofit setting. At 124M, Focus surpasses full attention (30.3 vs 31.4 PPL); trained from scratch at 7B scale (2B tokens), Focus again beats full attention (13.82 vs 13.89 PPL). At inference, restricting each token to its top-k highest-scoring groups discretizes the soft routing into a hard sparsity pattern, yielding 2x speedup while beating the pretrained baseline (41.3 vs 42.8 PPL); decomposing this pattern into two standard FlashAttention calls reaches 8.6x wall-clock speedup at 1M tokens with no custom kernels. Unlike LoRA, centroid routing preserves alignment: instruction-tuned models retain TruthfulQA scores after adaptation, while LoRA degrades at every learning rate and rank. Sinkhorn normalization enforces balanced groups as a hard constraint, and the resulting groups discover interpretable linguistic categories without supervision.

Yang Li, Xing Chen, Yutao Liu et al.

Large Language Models (LLMs) have achieved strong performance on static reasoning benchmarks, yet their effectiveness as interactive agents operating in adversarial, time-sensitive environments remains poorly understood. Existing evaluations largely treat reasoning as a single-shot capability, overlooking the challenges of opponent-aware decision-making, temporal constraints, and execution under pressure. This paper introduces Strategic Tactical Agent Reasoning (STAR) Benchmark, a multi-agent evaluation framework that assesses LLMs through 1v1 zero-sum competitive interactions, framing reasoning as an iterative, adaptive decision-making process. STAR supports both turn-based and real-time settings, enabling controlled analysis of long-horizon strategic planning and fast-paced tactical execution within a unified environment. Built on a modular architecture with a standardized API and fully implemented execution engine, STAR facilitates reproducible evaluation and flexible task customization. To move beyond binary win-loss outcomes, we introduce a Strategic Evaluation Suite that assesses not only competitive success but also the quality of strategic behavior, such as execution efficiency and outcome stability. Extensive pairwise evaluations reveal a pronounced strategy-execution gap: while reasoning-intensive models dominate turn-based settings, their inference latency often leads to inferior performance in real-time scenarios, where faster instruction-tuned models prevail. These results show that strategic intelligence in interactive environments depends not only on reasoning depth, but also on the ability to translate plans into timely actions, positioning STAR as a principled benchmark for studying this trade-off in competitive, dynamic settings.

Jinhao Li, Yuhao Sun, Zhiyuan Ma et al.

Recurrent spiking neural networks (RSNNs) are a promising substrate for energy-efficient control policies, but training them for high-dimensional, long-horizon reinforcement learning remains challenging. Population-based, gradient-free optimization circumvents backpropagation through non-differentiable spike dynamics by estimating gradients. However, with finite populations, high variance of these estimates can induce harmful and overly aggressive update steps. Inspired by trust-region methods in reinforcement learning that constrain policy updates in distribution space, we propose \textbf{Signal-Adaptive Trust Regions (SATR)}, a distributional update rule that constrains relative change by bounding KL divergence normalized by an estimated signal energy. SATR automatically expands the trust region under strong signals and contracts it when updates are noise-dominated. We instantiate SATR for Bernoulli connectivity distributions, which have shown strong empirical performance for RSNN optimization. Across a suite of high-dimensional continuous-control benchmarks, SATR improves stability under limited populations and reaches competitive returns against strong baselines including PPO-LSTM. In addition, to make SATR practical at scale, we introduce a bitset implementation for binary spiking and binary weights, substantially reducing wall-clock training time and enabling fast RSNN policy search.

Guan Wang, Jin Li, Yuhao Sun et al.

Reasoning, the process of devising and executing complex goal-oriented action sequences, remains a critical challenge in AI. Current large language models (LLMs) primarily employ Chain-of-Thought (CoT) techniques, which suffer from brittle task decomposition, extensive data requirements, and high latency. Inspired by the hierarchical and multi-timescale processing in the human brain, we propose the Hierarchical Reasoning Model (HRM), a novel recurrent architecture that attains significant computational depth while maintaining both training stability and efficiency. HRM executes sequential reasoning tasks in a single forward pass without explicit supervision of the intermediate process, through two interdependent recurrent modules: a high-level module responsible for slow, abstract planning, and a low-level module handling rapid, detailed computations. With only 27 million parameters, HRM achieves exceptional performance on complex reasoning tasks using only 1000 training samples. The model operates without pre-training or CoT data, yet achieves nearly perfect performance on challenging tasks including complex Sudoku puzzles and optimal path finding in large mazes. Furthermore, HRM outperforms much larger models with significantly longer context windows on the Abstraction and Reasoning Corpus (ARC), a key benchmark for measuring artificial general intelligence capabilities. These results underscore HRM's potential as a transformative advancement toward universal computation and general-purpose reasoning systems.

StepFun, Bin Wang, Bojun Wang et al.

Large language models (LLMs) face low hardware efficiency during decoding, especially for long-context reasoning tasks. This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs. Step-3 innovates in two key dimensions: (1) A novel Multi-Matrix Factorization Attention (MFA) mechanism that significantly reduces both KV cache size and computation while maintaining high attention expressiveness, and (2) Attention-FFN Disaggregation (AFD), a distributed inference system that decouples attention and Feed-Forward Network (FFN) layers into specialized subsystems. This co-design achieves unprecedented cost efficiency: Step-3 significantly reduces theoretical decoding costs compared with models like DeepSeek-V3 and Qwen3 MoE 235B, with the gains widening at longer context. Step-3 achieves low cost while activating 38B parameters per token (more than DeepSeek-V3 and Qwen3 MoE 235B), demonstrating that hardware-aligned attention arithmetic intensity, MoE sparsity, and AFD are critical to cost-effectiveness. We perform a head-to-head comparison with DeepSeek-V3 in its favorable scenarios. Our implementation on Hopper GPUs achieves a decoding throughput of up to 4,039 tokens per second per GPU under 50ms TPOT SLA (4K context, FP8, no MTP). It is higher than DeepSeek-V3's 2,324 in the same setup and sets a new Pareto frontier for LLM decoding.

Xin Tan, Yuetao Chen, Yimin Jiang et al.

Diffusion Transformers (DiTs) have shown remarkable performance in generating high-quality videos. However, the quadratic complexity of 3D full attention remains a bottleneck in scaling DiT training, especially with high-definition, lengthy videos, where it can consume up to 95% of processing time and demand specialized context parallelism. This paper introduces DSV to accelerate video DiT training by leveraging the dynamic attention sparsity we empirically observe. DSV uses a two-stage algorithm to capture the dynamic sparsity patterns via low-rank based approximation of the original query and key. It employs custom kernels to efficiently identify critical key-value pairs and compute the sparse attention. To accommodate the new sparsity dimension, DSV adopts a hybrid sparsity-aware context parallelism that re-balances the skewed workload across attention heads and blocks due to sparsity heterogeneity. DSV achieves up to 3.02x higher training throughput, scaling to 128 GPUs and 520k token lengths, without quality loss.

Jing Ye, Yiwen Duan, Yonghong Yu et al.

SQL is central to enterprise data engineering, yet generating fully correct SQL code in a single attempt remains difficult, even for experienced developers and advanced text-to-SQL LLMs, often requiring multiple debugging iterations. We introduce OurBench, the first benchmark for enterprise-level SQL reasoning and debugging. Our benchmark is built on two key innovations: (1) an automated construction workflow that uses reverse engineering to systematically inject realistic bugs into large-scale SQL code, enabling scalable and diverse benchmark generation; and (2) an execution-free evaluation framework tailored to enterprise settings, providing fast, accurate, and resource-efficient assessment. OurBench comprises 469 OurBenchSyn queries featuring syntax errors with explicit error messages, and 516 OurBenchSem queries targeting semantic errors in which the code fails to meet user intent. The queries are highly complex, averaging over 140 lines and featuring deep and wide abstract syntax trees. Evaluation of nearly 30 LLMs reveals a substantial performance gap: the best-performing model, Claude-4-Sonnet, achieves only 36.46 percent accuracy on OurBenchSyn and 32.17 percent on OurBenchSem, while most models score below 20 percent. We further explore four solution strategies, identify key challenges, and outline promising directions for enterprise SQL debugging with LLMs.

Hengshuai Yao, Xing Chen, Ahmed Murtadha et al.

Adapting LLMs to new domains causes forgetting because standard methods (full fine-tuning, LoRA) inject new directions into the weight space. We propose GAIN, which re-emphasizes existing features through multiplicative modulation W_new = S * W. The learned diagonal matrix S is applied to the attention output projection and optionally the FFN. The principle mirrors gain modulation in neuroscience, where neurons adapt to context by scaling response strength while preserving selectivity. We evaluate GAIN on five models from four families (774M to 70B), adapting sequentially across eight domains. GAIN-FFN matches LoRA's in-domain adaptation, but their effects on previously trained domains are opposite: GAIN-FFN improves them by 7-13% (validation PPL), while LoRA degrades them by 18-36%. Downstream accuracy confirms the pattern: for example, after seven sequential adaptations on Qwen2.5, GAIN-FFN degrades BoolQ by only 0.8% while LoRA damages it by 14.9%. GAIN adds 46K-230K parameters per model and can be absorbed into the pretrained weights for zero inference cost.

Han Weng, Puzhen Wu, Longjie Cui et al.

Reinforcement learning (RL) has been widely adopted to enhance the performance of large language models (LLMs) on Text-to-SQL tasks. However, existing methods often rely on execution-based or LLM-based Bradley-Terry reward models. The former suffers from high execution latency caused by repeated database calls, whereas the latter imposes substantial GPU memory overhead, both of which significantly hinder the efficiency and scalability of RL pipelines. To this end, we propose a novel reward model framework for RL-based Text-to-SQL named Graph-Reward-SQL, which employs the GMNScore outcome reward model. We leverage SQL graph representations to provide accurate reward signals while significantly reducing time cost and GPU memory usage. Building on this foundation, we further introduce StepRTM, a stepwise reward model that provides intermediate supervision over Common Table Expression (CTE) subqueries. This encourages both functional correctness and readability of SQL. Extensive comparative and ablation experiments on standard benchmarks, including Spider and BIRD, demonstrate that our method consistently outperforms existing reward models.

Zhenliang Xue, Hanpeng Hu, Xing Chen et al.

Large multimodal models (LMMs) have demonstrated excellent capabilities in both understanding and generation tasks with various modalities. While these models can accept flexible combinations of input data, their training efficiency suffers from two major issues: pipeline stage imbalance caused by heterogeneous model architectures, and training data dynamicity stemming from the diversity of multimodal data. In this paper, we present PipeWeaver, a dynamic pipeline scheduling framework designed for LMM training. The core of PipeWeaver is dynamic interleaved pipeline, which searches for pipeline schedules dynamically tailored to current training batches. PipeWeaver addresses issues of LMM training with two techniques: adaptive modality-aware partitioning and efficient pipeline schedule search within a hierarchical schedule space. Meanwhile, PipeWeaver utilizes SEMU (Step Emulator), a training simulator for multimodal models, for accurate performance estimations, accelerated by spatial-temporal subgraph reuse to improve search efficiency. Experiments show that PipeWeaver can enhance LMM training efficiency by up to 97.3% compared to state-of-the-art systems, and demonstrate excellent adaptivity to LMM training's data dynamicity.

Xiaochuan Liu, Yuanfeng Song, Xiaoming Yin et al.

In today's data-driven era, fully automated end-to-end data analytics, particularly insight discovery, is critical for discovering actionable insights that assist organizations in making effective decisions. With the rapid advancement of large language models (LLMs), LLM-driven agents have emerged as a promising paradigm for automating data analysis and insight discovery. However, existing data insight agents remain limited in several key aspects, often failing to deliver satisfactory results due to: (1) insufficient utilization of domain knowledge, (2) shallow analytical depth, and (3) error-prone code generation during insight generation. To address these issues, we propose DataSage, a novel multi-agent framework that incorporates three innovative features including external knowledge retrieval to enrich the analytical context, a multi-role debating mechanism to simulate diverse analytical perspectives and deepen analytical depth, and multi-path reasoning to improve the accuracy of the generated code and insights. Extensive experiments on InsightBench demonstrate that DataSage consistently outperforms existing data insight agents across all difficulty levels, offering an effective solution for automated data insight discovery.

Hao Wang, Yuanfeng Song, Xiaoming Yin et al.

Text-to-SQL datasets are essential for training and evaluating text-to-SQL models, but existing datasets often suffer from limited coverage and fail to capture the diversity of real-world applications. To address this, we propose a novel taxonomy for text-to-SQL classification based on dimensions including core intents, statement types, syntax structures, and key actions. Using this taxonomy, we evaluate widely used public text-to-SQL datasets (e.g., Spider and Bird) and reveal limitations in their coverage and diversity. We then introduce a taxonomy-guided dataset synthesis pipeline, yielding a new dataset named SQL-Synth. This approach combines the taxonomy with Large Language Models (LLMs) to ensure the dataset reflects the breadth and complexity of real-world text-to-SQL applications. Extensive analysis and experimental results validate the effectiveness of our taxonomy, as SQL-Synth exhibits greater diversity and coverage compared to existing benchmarks. Moreover, we uncover that existing LLMs typically fall short in adequately capturing the full range of scenarios, resulting in limited performance on SQL-Synth. However, fine-tuning can substantially improve their performance in these scenarios. The proposed taxonomy has significant potential impact, as it not only enables comprehensive analysis of datasets and the performance of different LLMs, but also guides the construction of training data for LLMs.

Guifeng Wang, Yuanfeng Song, Meng Yang et al.

Text-to-SQL, a pivotal natural language processing (NLP) task that converts textual queries into executable SQL, has seen substantial progress in recent years. However, existing evaluation and reward mechanisms used to train and assess the text-to-SQL models remain a critical bottleneck. Current approaches heavily rely on manually annotated gold SQL queries, which are costly to produce and impractical for large-scale evaluation. More importantly, most reinforcement learning (RL) methods in text-to-SQL leverage only the final binary execution outcome as the reward signal, a coarse-grained supervision that overlooks detailed structural and semantic errors from the perspective of rubrics. To address these challenges, we propose RuCo-C, a novel generative judge model for fine-grained, query-specific automatic evaluation using interpretable critiques without human intervention. Our framework first automatically generates query-specific evaluation rubrics for human-free annotation, linking them to interpretable critiques. Subsequently, it integrates densified reward feedback through a "progressive exploration" strategy during the RL training process, which dynamically adjusts the rewards to enhance the model's performance. Comprehensive experiments demonstrate that RuCo-C outperforms existing methods in text-to-SQL evaluation, yielding significant performance gains.

Kuntai Cai, Juncheng Liu, Xianglin Yang et al.

Large language model (LLM) agents typically receive two kinds of context: (i) environment-level manuals that define interaction interfaces and global rules, and (ii) task-level guidance or demonstrations tied to specific goals. In this work, we identify a crucial but overlooked third type of context, instance-level context, which consists of verifiable and reusable facts tied to a specific environment instance, such as object locations, crafting recipes, and local rules. We argue that the absence of instance-level context is a common source of failure for LLM agents in complex tasks, as success often depends not only on reasoning over global rules or task prompts but also on making decisions based on precise and persistent facts. Acquiring such context requires more than memorization: the challenge lies in efficiently exploring, validating, and formatting these facts under tight interaction budgets. We formalize this problem as Instance-Level Context Learning (ILCL) and introduce our task-agnostic method to solve it. Our method performs a guided exploration, using a compact TODO forest to intelligently prioritize its next actions and a lightweight plan-act-extract loop to execute them. This process automatically produces a high-precision context document that is reusable across many downstream tasks and agents, thereby amortizing the initial exploration cost. Experiments across TextWorld, ALFWorld, and Crafter demonstrate consistent gains in both success and efficiency: for instance, ReAct's mean success rate in TextWorld rises from 37% to 95%, while IGE improves from 81% to 95%. By transforming one-off exploration into persistent, reusable knowledge, our method complements existing contexts to enable more reliable and efficient LLM agents.

Xing Chen, Flavio Abreu Araujo, Mathieu Riou et al.

Deep learning has an increasing impact to assist research, allowing, for example, the discovery of novel materials. Until now, however, these artificial intelligence techniques have fallen short of discovering the full differential equation of an experimental physical system. Here we show that a dynamical neural network, trained on a minimal amount of data, can predict the behavior of spintronic devices with high accuracy and an extremely efficient simulation time, compared to the micromagnetic simulations that are usually employed to model them. For this purpose, we re-frame the formalism of Neural Ordinary Differential Equations (ODEs) to the constraints of spintronics: few measured outputs, multiple inputs and internal parameters. We demonstrate with Spin-Neural ODEs an acceleration factor over 200 compared to micromagnetic simulations for a complex problem -- the simulation of a reservoir computer made of magnetic skyrmions (20 minutes compared to three days). In a second realization, we show that we can predict the noisy response of experimental spintronic nano-oscillators to varying inputs after training Spin-Neural ODEs on five milliseconds of their measured response to different excitations. Spin-Neural ODE is a disruptive tool for developing spintronic applications in complement to micromagnetic simulations, which are time-consuming and cannot fit experiments when noise or imperfections are present. Spin-Neural ODE can also be generalized to other electronic devices involving dynamics.

Siyuan Yi, Xing Chen, Chuanming Tang

AI-based methods have been widely applied to tourism demand forecasting. However, current AI-based methods are short of the ability to process long-term dependency, and most of them lack interpretability. The Transformer used initially for machine translation shows an incredible ability to long-term dependency processing. Based on the Transformer, we proposed a time series Transformer (Tsformer) with Encoder-Decoder architecture for tourism demand forecasting. The proposed Tsformer encodes long-term dependency with encoder, captures short-term dependency with decoder, and simplifies the attention interactions under the premise of highlighting dominant attention through a series of attention masking mechanisms. These improvements make the multi-head attention mechanism process the input sequence according to the time relationship, contributing to better interpretability. What's more, the context processing ability of the Encoder-Decoder architecture allows adopting the calendar of days to be forecasted to enhance the forecasting performance. Experiments conducted on the Jiuzhaigou valley and Siguniang mountain tourism demand datasets with other nine baseline methods indicate that the proposed Tsformer outperformed all baseline models in the short-term and long-term tourism demand forecasting tasks. Moreover, ablation studies demonstrate that the adoption of the calendar of days to be forecasted contributes to the forecasting performance of the proposed Tsformer. For better interpretability, the attention weight matrix visualization is performed. It indicates that the Tsformer concentrates on seasonal features and days close to days to be forecast in short-term forecasting.

Xing Chen, Jingtao Li, Chaitali Chakrabarti

Split learning is a promising privacy-preserving distributed learning scheme that has low computation requirement at the edge device but has the disadvantage of high communication overhead between edge device and server. To reduce the communication overhead, this paper proposes a loss-based asynchronous training scheme that updates the client-side model less frequently and only sends/receives activations/gradients in selected epochs. To further reduce the communication overhead, the activations/gradients are quantized using 8-bit floating point prior to transmission. An added benefit of the proposed communication reduction method is that the computations at the client side are reduced due to reduction in the number of client model updates. Furthermore, the privacy of the proposed communication reduction based split learning method is almost the same as traditional split learning. Simulation results on VGG11, VGG13 and ResNet18 models on CIFAR-10 show that the communication cost is reduced by 1.64x-106.7x and the computations in the client are reduced by 2.86x-32.1x when the accuracy degradation is less than 0.5% for the single-client case. For 5 and 10-client cases, the communication cost reduction is 11.9x and 11.3x on VGG11 for 0.5% loss in accuracy.

Xinyu Huang, Lijun He, Xing Chen et al.

For in-vehicle application,task type and vehicle state information, i.e., vehicle speed, bear a significant impact on the task delay requirement. However, the joint impact of task type and vehicle speed on the task delay constraint has not been studied, and this lack of study may cause a mismatch between the requirement of the task delay and allocated computation and wireless resources. In this paper, we propose a joint task type and vehicle speed-aware task offloading and resource allocation strategy to decrease the vehicl's energy cost for executing tasks and increase the revenue of the vehicle for processing tasks within the delay constraint. First, we establish the joint task type and vehicle speed-aware delay constraint model. Then, the delay, energy cost and revenue for task execution in the vehicular edge computing (VEC) server, local terminal and terminals of other vehicles are calculated. Based on the energy cost and revenue from task execution,the utility function of the vehicle is acquired. Next, we formulate a joint optimization of task offloading and resource allocation to maximize the utility level of the vehicles subject to the constraints of task delay, computation resources and wireless resources. To obtain a near-optimal solution of the formulated problem, a joint offloading and resource allocation based on the multi-agent deep deterministic policy gradient (JORA-MADDPG) algorithm is proposed to maximize the utility level of vehicles. Simulation results show that our algorithm can achieve superior performance in task completion delay, vehicles' energy cost and processing revenue.

Xing Chen, Lijun He, Shang Xu et al.

More and more scholars focus on mobile edge computing (MEC) technology, because the strong storage and computing capabilities of MEC servers can reduce the long transmission delay, bandwidth waste, energy consumption, and privacy leaks in the data transmission process. In this paper, we study the cache placement problem to determine how to cache videos and which videos to be cached in a mobile edge computing system. First, we derive the video request probability by taking into account video popularity, user preference and the characteristic of video representations. Second, based on the acquired request probability, we formulate a cache placement problem with the objective to maximize the cache hit ratio subject to the storage capacity constraints. Finally, in order to solve the formulated problem, we transform it into a grouping knapsack problem and develop a dynamic programming algorithm to obtain the optimal caching strategy. Simulation results show that the proposed algorithm can greatly improve the cache hit ratio.

Kaidong Wu, Xiao He, Xing Chen et al.

When developing smart home systems, developers integrate and compose smart devices and software applications. Because of their diversity and heterogeneity, developers usually encounter many problems. In this paper, we present a runtime model based approach to smart home system development. First, we analyze mobile applications associated with smart devices and then extract some device control APIs. Second, we use SM@RT framework to build the device runtime model. Third, we define the scenario model, that is an abstraction of devices and objects which the system consists of. Fourth, we specify mapping rules from the scenario model to the runtime model and employ a synchronizer, which can interpret the mapping rules, to keep the synchronization between the scenario model and the device runtime model. The mapping handler reads the mapping rules that are defined by developers and does the mapping in terms of them. At last, developers can program smart home systems upon the MOF-compliant scenario model using the state-of-the-art model driven technologies. https://youtu.be/SP12OtmHj50

Richard J. Tomsett, Matt Ainsworth, Alexander Thiele et al.

Local field potentials (LFPs) sampled with extracellular electrodes are frequently used as a measure of population neuronal activity. However, relating such measurements to underlying neuronal behaviour and connectivity is non-trivial. To help study this link, we developed the Virtual Electrode Recording Tool for EXtracellular potentials (VERTEX). We first identified a reduced neuron model that retained the spatial and frequency filtering characteristics of extracellular potentials from neocortical neurons. We then developed VERTEX as an easy-to-use Matlab tool for simulating LFPs from large populations (>100 000 neurons). A VERTEX-based simulation successfully reproduced features of the LFPs from an in vitro multi-electrode array recording of macaque neocortical tissue. Our model, with virtual electrodes placed anywhere in 3D, allows direct comparisons with the in vitro recording setup. We envisage that VERTEX will stimulate experimentalists, clinicians, and computational neuroscientists to use models to understand the mechanisms underlying measured brain dynamics in health and disease.