Showing 1–50 of 455 results for author: Xia, W

$M^3-Verse$: A "Spot the Difference" Challenge for Large Multimodal Models

Authors: Kewei Wei, Bocheng Hu, Jie Cao, Xiaohan Chen, Zhengxi Lu, Wubing Xia, Weili Xu, Jiaao Wu, Junchen He, Mingyu Jia, Ciyun Zhao, Ye Sun, Yizhi Li, Zhonghan Zhao, Jian Zhang, Gaoang Wang

Abstract: Modern Large Multimodal Models (LMMs) have demonstrated extraordinary ability in static image and single-state spatial-temporal understanding. However, their capacity to comprehend the dynamic changes of objects within a shared spatial context between two distinct video observations, remains largely unexplored. This ability to reason about transformations within a consistent environment is particu… ▽ More Modern Large Multimodal Models (LMMs) have demonstrated extraordinary ability in static image and single-state spatial-temporal understanding. However, their capacity to comprehend the dynamic changes of objects within a shared spatial context between two distinct video observations, remains largely unexplored. This ability to reason about transformations within a consistent environment is particularly crucial for advancements in the field of spatial intelligence. In this paper, we introduce $M^3-Verse$, a Multi-Modal, Multi-State, Multi-Dimensional benchmark, to formally evaluate this capability. It is built upon paired videos that provide multi-perspective observations of an indoor scene before and after a state change. The benchmark contains a total of 270 scenes and 2,932 questions, which are categorized into over 50 subtasks that probe 4 core capabilities. We evaluate 16 state-of-the-art LMMs and observe their limitations in tracking state transitions. To address these challenges, we further propose a simple yet effective baseline that achieves significant performance improvements in multi-state perception. $M^3-Verse$ thus provides a challenging new testbed to catalyze the development of next-generation models with a more holistic understanding of our dynamic visual world. You can get the construction pipeline from https://github.com/Wal-K-aWay/M3-Verse_pipeline and full benchmark data from https://www.modelscope.cn/datasets/WalKaWay/M3-Verse. △ Less

Submitted 21 December, 2025; originally announced December 2025.

ISS Policy : Scalable Diffusion Policy with Implicit Scene Supervision

Authors: Wenlong Xia, Jinhao Zhang, Ce Zhang, Yaojia Wang, Youmin Gong, Jie Mei

Abstract: Vision-based imitation learning has enabled impressive robotic manipulation skills, but its reliance on object appearance while ignoring the underlying 3D scene structure leads to low training efficiency and poor generalization. To address these challenges, we introduce \emph{Implicit Scene Supervision (ISS) Policy}, a 3D visuomotor DiT-based diffusion policy that predicts sequences of continuous… ▽ More Vision-based imitation learning has enabled impressive robotic manipulation skills, but its reliance on object appearance while ignoring the underlying 3D scene structure leads to low training efficiency and poor generalization. To address these challenges, we introduce \emph{Implicit Scene Supervision (ISS) Policy}, a 3D visuomotor DiT-based diffusion policy that predicts sequences of continuous actions from point cloud observations. We extend DiT with a novel implicit scene supervision module that encourages the model to produce outputs consistent with the scene's geometric evolution, thereby improving the performance and robustness of the policy. Notably, ISS Policy achieves state-of-the-art performance on both single-arm manipulation tasks (MetaWorld) and dexterous hand manipulation (Adroit). In real-world experiments, it also demonstrates strong generalization and robustness. Additional ablation studies show that our method scales effectively with both data and parameters. Code and videos will be released. △ Less

Submitted 16 December, 2025; originally announced December 2025.

MMGR: Multi-Modal Generative Reasoning

Authors: Zefan Cai, Haoyi Qiu, Tianyi Ma, Haozhe Zhao, Gengze Zhou, Kung-Hsiang Huang, Parisa Kordjamshidi, Minjia Zhang, Wen Xiao, Jiuxiang Gu, Nanyun Peng, Junjie Hu

Abstract: Video foundation models generate visually realistic and temporally coherent content, but their reliability as world simulators depends on whether they capture physical, logical, and spatial constraints. Existing metrics such as Frechet Video Distance (FVD) emphasize perceptual quality and overlook reasoning failures, including violations of causality, physics, and global consistency. We introduce… ▽ More Video foundation models generate visually realistic and temporally coherent content, but their reliability as world simulators depends on whether they capture physical, logical, and spatial constraints. Existing metrics such as Frechet Video Distance (FVD) emphasize perceptual quality and overlook reasoning failures, including violations of causality, physics, and global consistency. We introduce MMGR (Multi-Modal Generative Reasoning Evaluation and Benchmark), a principled evaluation framework based on five reasoning abilities: Physical, Logical, 3D Spatial, 2D Spatial, and Temporal. MMGR evaluates generative reasoning across three domains: Abstract Reasoning (ARC-AGI, Sudoku), Embodied Navigation (real-world 3D navigation and localization), and Physical Commonsense (sports and compositional interactions). MMGR applies fine-grained metrics that require holistic correctness across both video and image generation. We benchmark leading video models (Veo-3, Sora-2, Wan-2.2) and image models (Nano-banana, Nano-banana Pro, GPT-4o-image, Qwen-image), revealing strong performance gaps across domains. Models show moderate success on Physical Commonsense tasks but perform poorly on Abstract Reasoning (below 10 percent accuracy on ARC-AGI) and struggle with long-horizon spatial planning in embodied settings. Our analysis highlights key limitations in current models, including overreliance on perceptual data, weak global state consistency, and objectives that reward visual plausibility over causal correctness. MMGR offers a unified diagnostic benchmark and a path toward reasoning-aware generative world models. △ Less

Submitted 17 December, 2025; v1 submitted 16 December, 2025; originally announced December 2025.

Comments: work in progress

Taylor-Lagrange Control for Safety-Critical Systems

Authors: Wei Xiao, Anni Li

Abstract: This paper proposes a novel Taylor-Lagrange Control (TLC) method for nonlinear control systems to ensure the safety and stability through Taylor's theorem with Lagrange remainder. To achieve this, we expand a safety or stability function with respect to time along the system dynamics using the Lie derivative and Taylor's theorem. This expansion enables the control input to appear in the Taylor ser… ▽ More This paper proposes a novel Taylor-Lagrange Control (TLC) method for nonlinear control systems to ensure the safety and stability through Taylor's theorem with Lagrange remainder. To achieve this, we expand a safety or stability function with respect to time along the system dynamics using the Lie derivative and Taylor's theorem. This expansion enables the control input to appear in the Taylor series at an order equivalent to the relative degree of the function. We show that the proposed TLC provides necessary and sufficient conditions for system safety and is applicable to systems and constraints of arbitrary relative degree. The TLC exhibits connections with existing Control Barrier Function (CBF) and Control Lyapunov Function (CLF) methods, and it further extends the CBF and CLF methods to the complex domain, especially for higher order cases. Compared to High-Order CBFs (HOCBFs), TLC is less restrictive as it does not require forward invariance of the intersection of a set of safe sets while HOCBFs do. We employ TLC to reformulate a constrained optimal control problem as a sequence of quadratic programs with a zero-order hold implementation method, and demonstrate the safety of zero-order hold TLC using an event-triggered control method to address inter-sampling effects. Finally, we illustrate the effectiveness of the proposed TLC method through an adaptive cruise control system and a robot control problem, and compare it with existing CBF methods. △ Less

Submitted 12 December, 2025; originally announced December 2025.

Comments: 13 pages

Learning Robust and Correct Controllers Guided by Feasibility-Aware Signal Temporal Logic via BarrierNet

Authors: Shuo Liu, Wenliang Liu, Wei Xiao, Calin A. Belta

Abstract: Control Barrier Functions (CBFs) have emerged as a powerful tool for enforcing safety in optimization-based controllers, and their integration with Signal Temporal Logic (STL) has enabled the specification-driven synthesis of complex robotic behaviors. However, existing CBF-STL approaches typically rely on fixed hyperparameters and myopic, per-time step optimization, which can lead to overly conse… ▽ More Control Barrier Functions (CBFs) have emerged as a powerful tool for enforcing safety in optimization-based controllers, and their integration with Signal Temporal Logic (STL) has enabled the specification-driven synthesis of complex robotic behaviors. However, existing CBF-STL approaches typically rely on fixed hyperparameters and myopic, per-time step optimization, which can lead to overly conservative behavior, infeasibility near tight input limits, and difficulty satisfying long-horizon STL tasks. To address these limitations, we propose a feasibility-aware learning framework that embeds trainable, time-varying High Order Control Barrier Functions (HOCBFs) into a differentiable Quadratic Program (dQP). Our approach provides a systematic procedure for constructing time-varying HOCBF constraints for a broad fragment of STL and introduces a unified robustness measure that jointly captures STL satisfaction, QP feasibility, and control-bound compliance. Three neural networks-InitNet, RefNet, and an extended BarrierNet-collaborate to generate reference inputs and adapt constraint-related hyperparameters automatically over time and across initial conditions, reducing conservativeness while maximizing robustness. The resulting controller achieves STL satisfaction with strictly feasible dQPs and requires no manual tuning. Simulation results demonstrate that the proposed framework maintains high STL robustness under tight input bounds and significantly outperforms fixed-parameter and non-adaptive baselines in complex environments. △ Less

Submitted 16 December, 2025; v1 submitted 7 December, 2025; originally announced December 2025.

Comments: 16 pages, 11 figures

EgoLCD: Egocentric Video Generation with Long Context Diffusion

Authors: Liuzhou Zhang, Jiarui Ye, Yuanlei Wang, Ming Zhong, Mingju Cao, Wanke Xia, Bowen Zeng, Zeyu Zhang, Hao Tang

Abstract: Generating long, coherent egocentric videos is difficult, as hand-object interactions and procedural tasks require reliable long-term memory. Existing autoregressive models suffer from content drift, where object identity and scene semantics degrade over time. To address this challenge, we introduce EgoLCD, an end-to-end framework for egocentric long-context video generation that treats long video… ▽ More Generating long, coherent egocentric videos is difficult, as hand-object interactions and procedural tasks require reliable long-term memory. Existing autoregressive models suffer from content drift, where object identity and scene semantics degrade over time. To address this challenge, we introduce EgoLCD, an end-to-end framework for egocentric long-context video generation that treats long video synthesis as a problem of efficient and stable memory management. EgoLCD combines a Long-Term Sparse KV Cache for stable global context with an attention-based short-term memory, extended by LoRA for local adaptation. A Memory Regulation Loss enforces consistent memory usage, and Structured Narrative Prompting provides explicit temporal guidance. Extensive experiments on the EgoVid-5M benchmark demonstrate that EgoLCD achieves state-of-the-art performance in both perceptual quality and temporal consistency, effectively mitigating generative forgetting and representing a significant step toward building scalable world models for embodied AI. Code: https://github.com/AIGeeksGroup/EgoLCD. Website: https://aigeeksgroup.github.io/EgoLCD. △ Less

Submitted 4 December, 2025; originally announced December 2025.

Belobog: Move Language Fuzzing Framework For Real-World Smart Contracts

Authors: Wanxu Xia, Ziqiao Kong, Zhengwei Li, Yi Lu, Pan Li, Liqun Yang, Yang Liu, Xiapu Luo, Shaohua Li

Abstract: Move is a research-oriented programming language design for secure and verifiable smart contract development and has been widely used in managing billions of digital assets in blockchains, such as Sui and Aptos. Move features a strong static type system and explicit resource semantics to enforce safety properties such as the prevention of data races, invalid asset transfers, and entry vulnerabilit… ▽ More Move is a research-oriented programming language design for secure and verifiable smart contract development and has been widely used in managing billions of digital assets in blockchains, such as Sui and Aptos. Move features a strong static type system and explicit resource semantics to enforce safety properties such as the prevention of data races, invalid asset transfers, and entry vulnerabilities. However, smart contracts written in Move may still contain certain vulnerabilities that are beyond the reach of its type system. It is thus essential to validate Move smart contracts. Unfortunately, due to its strong type system, existing smart contract fuzzers are ineffective in producing syntactically or semantically valid transactions to test Move smart contracts. This paper introduces the first fuzzing framework, Belobog, for Move smart contracts. Belobog is type-aware and ensures that all generated and mutated transactions are well-typed. More specifically, for a target Move smart contract, Belobog first constructs a type graph based on Move's type system, and then generates or mutates a transaction based on the graph trace derived from the type graph. In order to overcome the complex checks in Move smart contracts, we further design and implement a concolic executor in Belobog. We evaluated Belobog on 109 real-world Move smart contract projects. The experimental results show that Belobog is able to detect 100\% critical and 79\% major vulnerabilities manually audited by human experts. We further selected two recent notorious incidents in Move smart contracts, i.e., Cetus and Nemo. Belobog successfully reproduced full exploits for both of them, without any prior knowledge. △ Less

Submitted 2 December, 2025; originally announced December 2025.

Comments: Slight revision and under review

StockMem: An Event-Reflection Memory Framework for Stock Forecasting

Authors: He Wang, Wenyilin Xiao, Songqiao Han, Hailiang Huang

Abstract: Stock price prediction is challenging due to market volatility and its sensitivity to real-time events. While large language models (LLMs) offer new avenues for text-based forecasting, their application in finance is hindered by noisy news data and the lack of explicit answers in text. General-purpose memory architectures struggle to identify the key drivers of price movements. To address this, we… ▽ More Stock price prediction is challenging due to market volatility and its sensitivity to real-time events. While large language models (LLMs) offer new avenues for text-based forecasting, their application in finance is hindered by noisy news data and the lack of explicit answers in text. General-purpose memory architectures struggle to identify the key drivers of price movements. To address this, we propose StockMem, an event-reflection dual-layer memory framework. It structures news into events and mines them along two dimensions: horizontal consolidation integrates daily events, while longitudinal tracking captures event evolution to extract incremental information reflecting market expectation discrepancies. This builds a temporal event knowledge base. By analyzing event-price dynamics, the framework further forms a reflection knowledge base of causal experiences. For prediction, it retrieves analogous historical scenarios and reasons with current events, incremental data, and past experiences. Experiments show StockMem outperforms existing memory architectures and provides superior, explainable reasoning by tracing the information chain affecting prices, enhancing decision transparency in financial forecasting. △ Less

Submitted 2 December, 2025; originally announced December 2025.

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

Authors: DeepSeek-AI, Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenhao Xu, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Erhang Li, Fangqi Zhou, Fangyun Lin, Fucong Dai, Guangbo Hao , et al. (239 additional authors not shown)

Abstract: We introduce DeepSeek-V3.2, a model that harmonizes high computational efficiency with superior reasoning and agent performance. The key technical breakthroughs of DeepSeek-V3.2 are as follows: (1) DeepSeek Sparse Attention (DSA): We introduce DSA, an efficient attention mechanism that substantially reduces computational complexity while preserving model performance in long-context scenarios. (2)… ▽ More We introduce DeepSeek-V3.2, a model that harmonizes high computational efficiency with superior reasoning and agent performance. The key technical breakthroughs of DeepSeek-V3.2 are as follows: (1) DeepSeek Sparse Attention (DSA): We introduce DSA, an efficient attention mechanism that substantially reduces computational complexity while preserving model performance in long-context scenarios. (2) Scalable Reinforcement Learning Framework: By implementing a robust reinforcement learning protocol and scaling post-training compute, DeepSeek-V3.2 performs comparably to GPT-5. Notably, our high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and exhibits reasoning proficiency on par with Gemini-3.0-Pro, achieving gold-medal performance in both the 2025 International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI). (3) Large-Scale Agentic Task Synthesis Pipeline: To integrate reasoning into tool-use scenarios, we developed a novel synthesis pipeline that systematically generates training data at scale. This methodology facilitates scalable agentic post-training, yielding substantial improvements in generalization and instruction-following robustness within complex, interactive environments. △ Less

Submitted 2 December, 2025; originally announced December 2025.

Opening the Sim-to-Real Door for Humanoid Pixel-to-Action Policy Transfer

Authors: Haoru Xue, Tairan He, Zi Wang, Qingwei Ben, Wenli Xiao, Zhengyi Luo, Xingye Da, Fernando Castañeda, Guanya Shi, Shankar Sastry, Linxi "Jim" Fan, Yuke Zhu

Abstract: Recent progress in GPU-accelerated, photorealistic simulation has opened a scalable data-generation path for robot learning, where massive physics and visual randomization allow policies to generalize beyond curated environments. Building on these advances, we develop a teacher-student-bootstrap learning framework for vision-based humanoid loco-manipulation, using articulated-object interaction as… ▽ More Recent progress in GPU-accelerated, photorealistic simulation has opened a scalable data-generation path for robot learning, where massive physics and visual randomization allow policies to generalize beyond curated environments. Building on these advances, we develop a teacher-student-bootstrap learning framework for vision-based humanoid loco-manipulation, using articulated-object interaction as a representative high-difficulty benchmark. Our approach introduces a staged-reset exploration strategy that stabilizes long-horizon privileged-policy training, and a GRPO-based fine-tuning procedure that mitigates partial observability and improves closed-loop consistency in sim-to-real RL. Trained entirely on simulation data, the resulting policy achieves robust zero-shot performance across diverse door types and outperforms human teleoperators by up to 31.7% in task completion time under the same whole-body control stack. This represents the first humanoid sim-to-real policy capable of diverse articulated loco-manipulation using pure RGB perception. △ Less

Submitted 30 November, 2025; originally announced December 2025.

Comments: https://doorman-humanoid.github.io/

TrojanLoC: LLM-based Framework for RTL Trojan Localization

Authors: Weihua Xiao, Zeng Wang, Minghao Shao, Raghu Vamshi Hemadri, Ozgur Sinanoglu, Muhammad Shafique, Johann Knechtel, Siddharth Garg, Ramesh Karri

Abstract: Hardware Trojans (HT s) are a persistent threat to integrated circuits, especially when inserted at the register-transfer level (RTL). Existing methods typically first convert the design into a graph, such as a gate-level netlist or an RTL-derived dataflow graph (DFG), and then use a graph neural network (GNN ) to obtain an embedding of that graph, which (i) loses compact RTL semantics, (ii) relie… ▽ More Hardware Trojans (HT s) are a persistent threat to integrated circuits, especially when inserted at the register-transfer level (RTL). Existing methods typically first convert the design into a graph, such as a gate-level netlist or an RTL-derived dataflow graph (DFG), and then use a graph neural network (GNN ) to obtain an embedding of that graph, which (i) loses compact RTL semantics, (ii) relies on shallow GNNs with limited receptive field, and (iii) is largely restricted to coarse, module-level binary HT detection. We propose TrojanLoC, an LLM-based framework for RTL-level HT localization. We use an RTL-finetuned LLM to derive module-level and line-level embeddings directly from RTL code, capturing both global design context and local semantics. Next, we train task-specific classifiers on these embeddings to perform module-level Trojan detection, type prediction, and fine-grained line-level localization. We also introduce TrojanInS, a large synthetic dataset of RTL designs with systematically injected Trojans from four effect-based categories, each accompanied by precise line-level annotations. Our experiments show that TrojanLoC achieves strong module-level performance, reaching 0.99 F1-score for Trojan detection, up to 0.68 higher than baseline, and 0.84 macro-F1 for Trojan-type classification. At the line level, TrojanLoc further achieves up to 0.93 macro-F1, enabling fine-grained localization of Trojan-relevant RTL lines △ Less

Submitted 29 November, 2025; originally announced December 2025.

LAP: Fast LAtent Diffusion Planner with Fine-Grained Feature Distillation for Autonomous Driving

Authors: Jinhao Zhang, Wenlong Xia, Zhexuan Zhou, Youmin Gong, Jie Mei

Abstract: Diffusion models have demonstrated strong capabilities for modeling human-like driving behaviors in autonomous driving, but their iterative sampling process induces substantial latency, and operating directly on raw trajectory points forces the model to spend capacity on low-level kinematics, rather than high-level multi-modal semantics. To address these limitations, we propose LAtent Planner (LAP… ▽ More Diffusion models have demonstrated strong capabilities for modeling human-like driving behaviors in autonomous driving, but their iterative sampling process induces substantial latency, and operating directly on raw trajectory points forces the model to spend capacity on low-level kinematics, rather than high-level multi-modal semantics. To address these limitations, we propose LAtent Planner (LAP), a framework that plans in a VAE-learned latent space that disentangles high-level intents from low-level kinematics, enabling our planner to capture rich, multi-modal driving strategies. We further introduce a fine-grained feature distillation mechanism to guide a better interaction and fusion between the high-level semantic planning space and the vectorized scene context. Notably, LAP can produce high-quality plans in one single denoising step, substantially reducing computational overhead. Through extensive evaluations on the large-scale nuPlan benchmark, LAP achieves state-of-the-art closed-loop performance among learning-based planning methods, while demonstrating an inference speed-up of at most 10 times over previous SOTA approaches. Code will be released at: https://github.com/jhz1192/Latent-Planner. △ Less

Submitted 2 December, 2025; v1 submitted 29 November, 2025; originally announced December 2025.

VeriDispatcher: Multi-Model Dispatching through Pre-Inference Difficulty Prediction for RTL Generation Optimization

Authors: Zeng Wang, Weihua Xiao, Minghao Shao, Raghu Vamshi Hemadri, Ozgur Sinanoglu, Muhammad Shafique, Ramesh Karri

Abstract: Large Language Models (LLMs) show strong performance in RTL generation, but different models excel on different tasks because of architecture and training differences. Prior work mainly prompts or finetunes a single model. What remains not well studied is how to coordinate multiple different LLMs so they jointly improve RTL quality while also reducing cost, instead of running all models and choosi… ▽ More Large Language Models (LLMs) show strong performance in RTL generation, but different models excel on different tasks because of architecture and training differences. Prior work mainly prompts or finetunes a single model. What remains not well studied is how to coordinate multiple different LLMs so they jointly improve RTL quality while also reducing cost, instead of running all models and choosing the best output. We define this as the multi-LLM RTL generation problem. We propose VeriDispatcher, a multi-LLM RTL generation framework that dispatches each RTL task to suitable LLMs based on pre-inference difficulty prediction. For each model, we train a compact classifier over semantic embeddings of task descriptions, using difficulty scores derived from benchmark variants that combine syntax, structural similarity, and functional correctness. At inference, VeriDispatcher uses these predictors to route tasks to a selected subset of LLMs. Across 10 diverse LLMs on RTLLM and VerilogEval, VeriDispatcher achieves up to 18% accuracy improvement on RTLLM using only 40% of commercial calls, and on VerilogEval maintains accuracy while reducing commercial usage by 25%, enabling cost-effective, high-quality LLM deployment in hardware design automation. △ Less

Submitted 27 November, 2025; originally announced November 2025.

Gated KalmaNet: A Fading Memory Layer Through Test-Time Ridge Regression

Authors: Liangzu Peng, Aditya Chattopadhyay, Luca Zancato, Elvis Nunez, Wei Xia, Stefano Soatto

Abstract: As efficient alternatives to softmax Attention, linear State-Space Models (SSMs) achieve constant memory and linear compute, but maintain only a lossy, fading summary of the past, often leading to inferior performance in recall-oriented tasks. We propose Gated KalmaNet (GKA), a layer that accounts for the full past while maintaining SSM-style efficiency. We ground our approach in the Kalman Filter… ▽ More As efficient alternatives to softmax Attention, linear State-Space Models (SSMs) achieve constant memory and linear compute, but maintain only a lossy, fading summary of the past, often leading to inferior performance in recall-oriented tasks. We propose Gated KalmaNet (GKA), a layer that accounts for the full past while maintaining SSM-style efficiency. We ground our approach in the Kalman Filter (KF) framework, which provides a principled solution for optimal inference in dynamical systems. We show that several existing SSM layers (DeltaNet, Gated DeltaNet, and Kimi Delta Attention) are approximations to the KF recurrence that assume identity error covariance, thereby ignoring how past measurements (keys and values) should optimally influence state updates. In contrast, GKA computes the exact Kalman gain by maintaining the full error covariance. Under a steady-state assumption that enables parallelization, this reduces to solving an online ridge regression problem with constant memory and linear compute cost. A critical insight is that standard KF equations are numerically unstable in low-precision environments (like bfloat16) and hard to parallelize on modern hardware. We address this through: (1) adaptive regularization with input-dependent gating to control the condition number of the ridge regression for numerical stability, and (2) Chebyshev Iteration, which we show is more stable than conventional iterative solvers in low-precision settings. We further develop hardware-aware chunk-wise kernels to enable efficient training. Empirically, GKA outperforms existing SSM layers (like Mamba2 and Gated DeltaNet) on short-context tasks and achieves more than 10\% relative improvement on long-context RAG and LongQA tasks up to 128k tokens. △ Less

Submitted 18 December, 2025; v1 submitted 25 November, 2025; originally announced November 2025.

Comments: 30 pages, 10 figures

NeAR: Coupled Neural Asset-Renderer Stack

Authors: Hong Li, Chongjie Ye, Houyuan Chen, Weiqing Xiao, Ziyang Yan, Lixing Xiao, Zhaoxi Chen, Jianfeng Xiang, Shaocong Xu, Xuhui Liu, Yikai Wang, Baochang Zhang, Xiaoguang Han, Jiaolong Yang, Hao Zhao

Abstract: Neural asset authoring and neural rendering have traditionally evolved as disjoint paradigms: one generates digital assets for fixed graphics pipelines, while the other maps conventional assets to images. However, treating them as independent entities limits the potential for end-to-end optimization in fidelity and consistency. In this paper, we bridge this gap with NeAR, a Coupled Neural Asset--R… ▽ More Neural asset authoring and neural rendering have traditionally evolved as disjoint paradigms: one generates digital assets for fixed graphics pipelines, while the other maps conventional assets to images. However, treating them as independent entities limits the potential for end-to-end optimization in fidelity and consistency. In this paper, we bridge this gap with NeAR, a Coupled Neural Asset--Renderer Stack. We argue that co-designing the asset representation and the renderer creates a robust "contract" for superior generation. On the asset side, we introduce the Lighting-Homogenized SLAT (LH-SLAT). Leveraging a rectified-flow model, NeAR lifts casually lit single images into a canonical, illumination-invariant latent space, effectively suppressing baked-in shadows and highlights. On the renderer side, we design a lighting-aware neural decoder tailored to interpret these homogenized latents. Conditioned on HDR environment maps and camera views, it synthesizes relightable 3D Gaussian splats in real-time without per-object optimization. We validate NeAR on four tasks: (1) G-buffer-based forward rendering, (2) random-lit reconstruction, (3) unknown-lit relighting, and (4) novel-view relighting. Extensive experiments demonstrate that our coupled stack outperforms state-of-the-art baselines in both quantitative metrics and perceptual quality. We hope this coupled asset-renderer perspective inspires future graphics stacks that view neural assets and renderers as co-designed components instead of independent entities. △ Less

Submitted 18 December, 2025; v1 submitted 23 November, 2025; originally announced November 2025.

Comments: 20 pages, 19 figures. The project page: https://near-project.github.io/

SDA: Steering-Driven Distribution Alignment for Open LLMs without Fine-Tuning

Authors: Wei Xia, Zhi-Hong Deng

Abstract: With the rapid advancement of large language models (LLMs), their deployment in real-world applications has become increasingly widespread. LLMs are expected to deliver robust performance across diverse tasks, user preferences, and practical scenarios. However, as demands grow, ensuring that LLMs produce responses aligned with human intent remains a foundational challenge. In particular, aligning… ▽ More With the rapid advancement of large language models (LLMs), their deployment in real-world applications has become increasingly widespread. LLMs are expected to deliver robust performance across diverse tasks, user preferences, and practical scenarios. However, as demands grow, ensuring that LLMs produce responses aligned with human intent remains a foundational challenge. In particular, aligning model behavior effectively and efficiently during inference, without costly retraining or extensive supervision, is both a critical requirement and a non-trivial technical endeavor. To address the challenge, we propose SDA (Steering-Driven Distribution Alignment), a training-free and model-agnostic alignment framework designed for open-source LLMs. SDA dynamically redistributes model output probabilities based on user-defined alignment instructions, enhancing alignment between model behavior and human intents without fine-tuning. The method is lightweight, resource-efficient, and compatible with a wide range of open-source LLMs. It can function independently during inference or be integrated with training-based alignment strategies. Moreover, SDA supports personalized preference alignment, enabling flexible control over the model response behavior. Empirical results demonstrate that SDA consistently improves alignment performance across 8 open-source LLMs with varying scales and diverse origins, evaluated on three key alignment dimensions, helpfulness, harmlessness, and honesty (3H). Specifically, SDA achieves average gains of 64.4% in helpfulness, 30% in honesty and 11.5% in harmlessness across the tested models, indicating its effectiveness and generalization across diverse models and application scenarios. △ Less

Submitted 20 November, 2025; originally announced November 2025.

Fast LLM Post-training via Decoupled and Fastest-of-N Speculation

Authors: Rongxin Cheng, Kai Zhou, Xingda Wei, Siyuan Liu, Mingcong Han, Mingjing Ai, Yeju Zhou, Baoquan Zhong, Wencong Xiao, Rong Chen, Haibo Chen

Abstract: Rollout dominates the training time in large language model (LLM) post-training, where the trained model is used to generate tokens given a batch of prompts. This work, SpecActor, achieves fast rollout with speculative decoding that deploys a fast draft path to accelerate the unparallelizable generation, while the correctness is guaranteed by fast parallel verification of the outputs with the orig… ▽ More Rollout dominates the training time in large language model (LLM) post-training, where the trained model is used to generate tokens given a batch of prompts. This work, SpecActor, achieves fast rollout with speculative decoding that deploys a fast draft path to accelerate the unparallelizable generation, while the correctness is guaranteed by fast parallel verification of the outputs with the original model. SpecActor addresses two foundational challenges that hinder speculation efficiency: (1) a Decoupled speculation method that overcomes the computation inefficiency issue when executing speculative decoding with relative large per-worker batch size -- a common configuration in training but unfriendly to speculation, and (2) a Fastest-of-N speculation method that selects and combines different draft methods according to the rollout progress to approximate the optimal draft method even when the best one is unknown a priori. Extensive evaluations on production traces show that SpecActor accelerates mean rollout speed by 2.0--2.4x, with up to 2.7x speedup, over common post-training baselines. The results are consistent across both dense and MoE models and across different RL algorithms. Notably, SpecActor is 1.1--2.6x faster compared to vanilla speculative rollout in different traces. The accelerated rollout achieves 1.4--2.3x faster end-to-end training time. △ Less

Submitted 23 December, 2025; v1 submitted 20 November, 2025; originally announced November 2025.

VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation

Authors: Tairan He, Zi Wang, Haoru Xue, Qingwei Ben, Zhengyi Luo, Wenli Xiao, Ye Yuan, Xingye Da, Fernando Castañeda, Shankar Sastry, Changliu Liu, Guanya Shi, Linxi Fan, Yuke Zhu

Abstract: A key barrier to the real-world deployment of humanoid robots is the lack of autonomous loco-manipulation skills. We introduce VIRAL, a visual sim-to-real framework that learns humanoid loco-manipulation entirely in simulation and deploys it zero-shot to real hardware. VIRAL follows a teacher-student design: a privileged RL teacher, operating on full state, learns long-horizon loco-manipulation us… ▽ More A key barrier to the real-world deployment of humanoid robots is the lack of autonomous loco-manipulation skills. We introduce VIRAL, a visual sim-to-real framework that learns humanoid loco-manipulation entirely in simulation and deploys it zero-shot to real hardware. VIRAL follows a teacher-student design: a privileged RL teacher, operating on full state, learns long-horizon loco-manipulation using a delta action space and reference state initialization. A vision-based student policy is then distilled from the teacher via large-scale simulation with tiled rendering, trained with a mixture of online DAgger and behavior cloning. We find that compute scale is critical: scaling simulation to tens of GPUs (up to 64) makes both teacher and student training reliable, while low-compute regimes often fail. To bridge the sim-to-real gap, VIRAL combines large-scale visual domain randomization over lighting, materials, camera parameters, image quality, and sensor delays--with real-to-sim alignment of the dexterous hands and cameras. Deployed on a Unitree G1 humanoid, the resulting RGB-based policy performs continuous loco-manipulation for up to 54 cycles, generalizing to diverse spatial and appearance variations without any real-world fine-tuning, and approaching expert-level teleoperation performance. Extensive ablations dissect the key design choices required to make RGB-based humanoid loco-manipulation work in practice. △ Less

Submitted 27 November, 2025; v1 submitted 19 November, 2025; originally announced November 2025.

Comments: Project website: https://viral-humanoid.github.io/

Experience-Guided Adaptation of Inference-Time Reasoning Strategies

Authors: Adam Stein, Matthew Trager, Benjamin Bowman, Michael Kleinman, Aditya Chattopadhyay, Wei Xia, Stefano Soatto

Abstract: Enabling agentic AI systems to adapt their problem-solving approaches based on post-training interactions remains a fundamental challenge. While systems that update and maintain a memory at inference time have been proposed, existing designs only steer the system by modifying textual input to a language model or agent, which means that they cannot change sampling parameters, remove tools, modify s… ▽ More Enabling agentic AI systems to adapt their problem-solving approaches based on post-training interactions remains a fundamental challenge. While systems that update and maintain a memory at inference time have been proposed, existing designs only steer the system by modifying textual input to a language model or agent, which means that they cannot change sampling parameters, remove tools, modify system prompts, or switch between agentic and workflow paradigms. On the other hand, systems that adapt more flexibly require offline optimization and remain static once deployed. We present Experience-Guided Reasoner (EGuR), which generates tailored strategies -- complete computational procedures involving LLM calls, tools, sampling parameters, and control logic -- dynamically at inference time based on accumulated experience. We achieve this using an LLM-based meta-strategy -- a strategy that outputs strategies -- enabling adaptation of all strategy components (prompts, sampling parameters, tool configurations, and control logic). EGuR operates through two components: a Guide generates multiple candidate strategies conditioned on the current problem and structured memory of past experiences, while a Consolidator integrates execution feedback to improve future strategy generation. This produces complete, ready-to-run strategies optimized for each problem, which can be cached, retrieved, and executed as needed without wasting resources. Across five challenging benchmarks (AIME 2025, 3-SAT, and three Big Bench Extra Hard tasks), EGuR achieves up to 14% accuracy improvements over the strongest baselines while reducing computational costs by up to 111x, with both metrics improving as the system gains experience. △ Less

Submitted 14 November, 2025; originally announced November 2025.

Comments: 29 pages, 5 figures

SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control

Authors: Zhengyi Luo, Ye Yuan, Tingwu Wang, Chenran Li, Sirui Chen, Fernando Castañeda, Zi-Ang Cao, Jiefeng Li, David Minor, Qingwei Ben, Xingye Da, Runyu Ding, Cyrus Hogg, Lina Song, Edy Lim, Eugene Jeong, Tairan He, Haoru Xue, Wenli Xiao, Zi Wang, Simon Yuen, Jan Kautz, Yan Chang, Umar Iqbal, Linxi "Jim" Fan , et al. (1 additional authors not shown)

Abstract: Despite the rise of billion-parameter foundation models trained across thousands of GPUs, similar scaling gains have not been shown for humanoid control. Current neural controllers for humanoids remain modest in size, target a limited set of behaviors, and are trained on a handful of GPUs over several days. We show that scaling up model capacity, data, and compute yields a generalist humanoid cont… ▽ More Despite the rise of billion-parameter foundation models trained across thousands of GPUs, similar scaling gains have not been shown for humanoid control. Current neural controllers for humanoids remain modest in size, target a limited set of behaviors, and are trained on a handful of GPUs over several days. We show that scaling up model capacity, data, and compute yields a generalist humanoid controller capable of creating natural and robust whole-body movements. Specifically, we posit motion tracking as a natural and scalable task for humanoid control, leveraging dense supervision from diverse motion-capture data to acquire human motion priors without manual reward engineering. We build a foundation model for motion tracking by scaling along three axes: network size (from 1.2M to 42M parameters), dataset volume (over 100M frames, 700 hours of high-quality motion data), and compute (9k GPU hours). Beyond demonstrating the benefits of scale, we show the practical utility of our model through two mechanisms: (1) a real-time universal kinematic planner that bridges motion tracking to downstream task execution, enabling natural and interactive control, and (2) a unified token space that supports various motion input interfaces, such as VR teleoperation devices, human videos, and vision-language-action (VLA) models, all using the same policy. Scaling motion tracking exhibits favorable properties: performance improves steadily with increased compute and data diversity, and learned representations generalize to unseen motions, establishing motion tracking at scale as a practical foundation for humanoid control. △ Less

Submitted 4 December, 2025; v1 submitted 10 November, 2025; originally announced November 2025.

Comments: Project page: https://nvlabs.github.io/SONIC/

TwinOR: Photorealistic Digital Twins of Dynamic Operating Rooms for Embodied AI Research

Authors: Han Zhang, Yiqing Shen, Roger D. Soberanis-Mukul, Ankita Ghosh, Hao Ding, Lalithkumar Seenivasan, Jose L. Porras, Zhekai Mao, Chenjia Li, Wenjie Xiao, Lonny Yarmus, Angela Christine Argento, Masaru Ishii, Mathias Unberath

Abstract: Developing embodied AI for intelligent surgical systems requires safe, controllable environments for continual learning and evaluation. However, safety regulations and operational constraints in operating rooms (ORs) limit embodied agents from freely perceiving and interacting in realistic settings. Digital twins provide high-fidelity, risk-free environments for exploration and training. How we ma… ▽ More Developing embodied AI for intelligent surgical systems requires safe, controllable environments for continual learning and evaluation. However, safety regulations and operational constraints in operating rooms (ORs) limit embodied agents from freely perceiving and interacting in realistic settings. Digital twins provide high-fidelity, risk-free environments for exploration and training. How we may create photorealistic and dynamic digital representations of ORs that capture relevant spatial, visual, and behavioral complexity remains unclear. We introduce TwinOR, a framework for constructing photorealistic, dynamic digital twins of ORs for embodied AI research. The system reconstructs static geometry from pre-scan videos and continuously models human and equipment motion through multi-view perception of OR activities. The static and dynamic components are fused into an immersive 3D environment that supports controllable simulation and embodied exploration. The proposed framework reconstructs complete OR geometry with centimeter level accuracy while preserving dynamic interaction across surgical workflows, enabling realistic renderings and a virtual playground for embodied AI systems. In our experiments, TwinOR simulates stereo and monocular sensor streams for geometry understanding and visual localization tasks. Models such as FoundationStereo and ORB-SLAM3 on TwinOR-synthesized data achieve performance within their reported accuracy on real indoor datasets, demonstrating that TwinOR provides sensor-level realism sufficient for perception and localization challenges. By establishing a real-to-sim pipeline for constructing dynamic, photorealistic digital twins of OR environments, TwinOR enables the safe, scalable, and data-efficient development and benchmarking of embodied AI, ultimately accelerating the deployment of embodied AI from sim-to-real. △ Less

Submitted 10 November, 2025; originally announced November 2025.

Learning to Focus: Focal Attention for Selective and Scalable Transformers

Authors: Dhananjay Ram, Wei Xia, Stefano Soatto

Abstract: Attention is a core component of transformer architecture, whether encoder-only, decoder-only, or encoder-decoder model. However, the standard softmax attention often produces noisy probability distribution, which can impair effective feature selection at every layer of these models, particularly for long contexts. We propose Focal Attention, a simple yet effective modification that sharpens the a… ▽ More Attention is a core component of transformer architecture, whether encoder-only, decoder-only, or encoder-decoder model. However, the standard softmax attention often produces noisy probability distribution, which can impair effective feature selection at every layer of these models, particularly for long contexts. We propose Focal Attention, a simple yet effective modification that sharpens the attention distribution by controlling the softmax temperature, either as a fixed hyperparameter or as a learnable parameter during training. This sharpening enables the model to concentrate on the most relevant tokens while suppressing irrelevant ones. Empirically, Focal Attention scales more favorably than standard transformer with respect to model size, training data, and context length. Across diverse benchmarks, it achieves the same accuracy with up to 42% fewer parameters or 33% less training data. On long-context tasks, it delivers substantial relative improvements ranging from 17% to 82%, demonstrating its effectiveness in real world applications. △ Less

Submitted 10 November, 2025; originally announced November 2025.

DMSORT: An efficient parallel maritime multi-object tracking architecture for unmanned vessel platforms

Authors: Shengyu Tang, Zeyuan Lu, Jiazhi Dong, Changdong Yu, Xiaoyu Wang, Yaohui Lyu, Weihao Xia

Abstract: Accurate perception of the marine environment through robust multi-object tracking (MOT) is essential for ensuring safe vessel navigation and effective maritime surveillance. However, the complicated maritime environment often causes camera motion and subsequent visual degradation, posing significant challenges to MOT. To address this challenge, we propose an efficient Dual-branch Maritime SORT (D… ▽ More Accurate perception of the marine environment through robust multi-object tracking (MOT) is essential for ensuring safe vessel navigation and effective maritime surveillance. However, the complicated maritime environment often causes camera motion and subsequent visual degradation, posing significant challenges to MOT. To address this challenge, we propose an efficient Dual-branch Maritime SORT (DMSORT) method for maritime MOT. The core of the framework is a parallel tracker with affine compensation, which incorporates an object detection and re-identification (ReID) branch, along with a dedicated branch for dynamic camera motion estimation. Specifically, a Reversible Columnar Detection Network (RCDN) is integrated into the detection module to leverage multi-level visual features for robust object detection. Furthermore, a lightweight Transformer-based appearance extractor (Li-TAE) is designed to capture global contextual information and generate robust appearance features. Another branch decouples platform-induced and target-intrinsic motion by constructing a projective transformation, applying platform-motion compensation within the Kalman filter, and thereby stabilizing true object trajectories. Finally, a clustering-optimized feature fusion module effectively combines motion and appearance cues to ensure identity consistency under noise, occlusion, and drift. Extensive evaluations on the Singapore Maritime Dataset demonstrate that DMSORT achieves state-of-the-art performance. Notably, DMSORT attains the fastest runtime among existing ReID-based MOT frameworks while maintaining high identity consistency and robustness to jitter and occlusion. Code is available at: https://github.com/BiscuitsLzy/DMSORT-An-efficient-parallel-maritime-multi-object-tracking-architecture-. △ Less

Submitted 15 November, 2025; v1 submitted 6 November, 2025; originally announced November 2025.

Comments: This version clarifies several citation formatting inconsistencies caused by a technical issue in the reference management software used during manuscript preparation. All scientific data, experiments, and conclusions remain fully valid and unaffected. The clarification is provided to maintain transparency and consistency in the scholarly record

High-Resolution Magnetic Particle Imaging System Matrix Recovery Using a Vision Transformer with Residual Feature Network

Authors: Abuobaida M. Khair, Wenjing Jiang, Yousuf Babiker M. Osman, Wenjun Xia, Xiaopeng Ma

Abstract: This study presents a hybrid deep learning framework, the Vision Transformer with Residual Feature Network (VRF-Net), for recovering high-resolution system matrices in Magnetic Particle Imaging (MPI). MPI resolution often suffers from downsampling and coil sensitivity variations. VRF-Net addresses these challenges by combining transformer-based global attention with residual convolutional refineme… ▽ More This study presents a hybrid deep learning framework, the Vision Transformer with Residual Feature Network (VRF-Net), for recovering high-resolution system matrices in Magnetic Particle Imaging (MPI). MPI resolution often suffers from downsampling and coil sensitivity variations. VRF-Net addresses these challenges by combining transformer-based global attention with residual convolutional refinement, enabling recovery of both large-scale structures and fine details. To reflect realistic MPI conditions, the system matrix is degraded using a dual-stage downsampling strategy. Training employed paired-image super-resolution on the public Open MPI dataset and a simulated dataset incorporating variable coil sensitivity profiles. For system matrix recovery on the Open MPI dataset, VRF-Net achieved nRMSE = 0.403, pSNR = 39.08 dB, and SSIM = 0.835 at 2x scaling, and maintained strong performance even at challenging scale 8x (pSNR = 31.06 dB, SSIM = 0.717). For the simulated dataset, VRF-Net achieved nRMSE = 4.44, pSNR = 28.52 dB, and SSIM = 0.771 at 2x scaling, with stable performance at higher scales. On average, it reduced nRMSE by 88.2%, increased pSNR by 44.7%, and improved SSIM by 34.3% over interpolation and CNN-based methods. In image reconstruction of Open MPI phantoms, VRF-Net further reduced reconstruction error to nRMSE = 1.79 at 2x scaling, while preserving structural fidelity (pSNR = 41.58 dB, SSIM = 0.960), outperforming existing methods. These findings demonstrate that VRF-Net enables sharper, artifact-free system matrix recovery and robust image reconstruction across multiple scales, offering a promising direction for future in vivo applications. △ Less

Submitted 3 November, 2025; originally announced November 2025.

Journal ref: Biomedical Signal Processing and Control 113 (2026) 108990

Re-FORC: Adaptive Reward Prediction for Efficient Chain-of-Thought Reasoning

Authors: Renos Zabounidis, Aditya Golatkar, Michael Kleinman, Alessandro Achille, Wei Xia, Stefano Soatto

Abstract: We propose Re-FORC, an adaptive reward prediction method that, given a context, enables prediction of the expected future rewards as a function of the number of future thinking tokens. Re-FORC trains a lightweight adapter on reasoning models, demonstrating improved prediction with longer reasoning and larger models. Re-FORC enables: 1) early stopping of unpromising reasoning chains, reducing compu… ▽ More We propose Re-FORC, an adaptive reward prediction method that, given a context, enables prediction of the expected future rewards as a function of the number of future thinking tokens. Re-FORC trains a lightweight adapter on reasoning models, demonstrating improved prediction with longer reasoning and larger models. Re-FORC enables: 1) early stopping of unpromising reasoning chains, reducing compute by 26% while maintaining accuracy, 2) optimized model and thinking length selection that achieves 4% higher accuracy at equal compute and 55% less compute at equal accuracy compared to the largest model, 3) adaptive test-time scaling, which increases accuracy by 11% in high compute regime, and 7% in low compute regime. Re-FORC allows dynamic reasoning with length control via cost-per-token thresholds while estimating computation time upfront. △ Less

Submitted 3 November, 2025; originally announced November 2025.

Comments: Accepted at Efficient Reasoning Workshop at NeurIPS 2025

Self-Improving Vision-Language-Action Models with Data Generation via Residual RL

Authors: Wenli Xiao, Haotian Lin, Andy Peng, Haoru Xue, Tairan He, Yuqi Xie, Fengyuan Hu, Jimmy Wu, Zhengyi Luo, Linxi "Jim" Fan, Guanya Shi, Yuke Zhu

Abstract: Supervised fine-tuning (SFT) has become the de facto post-training strategy for large vision-language-action (VLA) models, but its reliance on costly human demonstrations limits scalability and generalization. We propose Probe, Learn, Distill (PLD), a three-stage plug-and-play framework that improves VLAs through residual reinforcement learning (RL) and distribution-aware data collection. In Stage… ▽ More Supervised fine-tuning (SFT) has become the de facto post-training strategy for large vision-language-action (VLA) models, but its reliance on costly human demonstrations limits scalability and generalization. We propose Probe, Learn, Distill (PLD), a three-stage plug-and-play framework that improves VLAs through residual reinforcement learning (RL) and distribution-aware data collection. In Stage 1, we train lightweight residual actors to probe failure regions of the VLA generalist. In Stage 2, we use a hybrid rollout scheme that aligns collected trajectories with the generalist's deployment distribution while capturing recovery behaviors. In Stage 3, we distill the curated trajectories back into the generalist with standard SFT. PLD achieves near-saturated 99% task success on LIBERO, over 50% gains in SimplerEnv, and 100% success on real-world Franka and YAM arm manipulation tasks. Ablations show that residual probing and distribution-aware replay are key to collecting deployment-aligned data that improves both seen and unseen tasks, offering a scalable path toward self-improving VLA models. △ Less

Submitted 30 October, 2025; originally announced November 2025.

Comments: 26 pages

e1: Learning Adaptive Control of Reasoning Effort

Authors: Michael Kleinman, Matthew Trager, Alessandro Achille, Wei Xia, Stefano Soatto

Abstract: Increasing the thinking budget of AI models can significantly improve accuracy, but not all questions warrant the same amount of reasoning. Users may prefer to allocate different amounts of reasoning effort depending on how they value output quality versus latency and cost. To leverage this tradeoff effectively, users need fine-grained control over the amount of thinking used for a particular quer… ▽ More Increasing the thinking budget of AI models can significantly improve accuracy, but not all questions warrant the same amount of reasoning. Users may prefer to allocate different amounts of reasoning effort depending on how they value output quality versus latency and cost. To leverage this tradeoff effectively, users need fine-grained control over the amount of thinking used for a particular query, but few approaches enable such control. Existing methods require users to specify the absolute number of desired tokens, but this requires knowing the difficulty of the problem beforehand to appropriately set the token budget for a query. To address these issues, we propose Adaptive Effort Control, a self-adaptive reinforcement learning method that trains models to use a user-specified fraction of tokens relative to the current average chain-of-thought length for each query. This approach eliminates dataset- and phase-specific tuning while producing better cost-accuracy tradeoff curves compared to standard methods. Users can dynamically adjust the cost-accuracy trade-off through a continuous effort parameter specified at inference time. We observe that the model automatically learns to allocate resources proportionally to the task difficulty and, across model scales ranging from 1.5B to 32B parameters, our approach enables a 2-3x reduction in chain-of-thought length while maintaining or improving performance relative to the base model used for RL training. △ Less

Submitted 11 November, 2025; v1 submitted 30 October, 2025; originally announced October 2025.

Beyond Hidden-Layer Manipulation: Semantically-Aware Logit Interventions for Debiasing LLMs

Authors: Wei Xia

Abstract: We proposed Static and Dynamic -- two zero-shot logits-layer debiasing methods. Dynamic reduces bias by up to 70% with minimal fluency loss. Logits intervention outperforms hidden-layer approaches. We show semantic-aware logits intervention is stable and effective for debiasing aligned LLMs. We proposed Static and Dynamic -- two zero-shot logits-layer debiasing methods. Dynamic reduces bias by up to 70% with minimal fluency loss. Logits intervention outperforms hidden-layer approaches. We show semantic-aware logits intervention is stable and effective for debiasing aligned LLMs. △ Less

Submitted 25 October, 2025; originally announced October 2025.

Poisson Flow Consistency Training

Authors: Anthony Zhang, Mahmut Gokmen, Dennis Hein, Rongjun Ge, Wenjun Xia, Ge Wang, Jin Chen

Abstract: The Poisson Flow Consistency Model (PFCM) is a consistency-style model based on the robust Poisson Flow Generative Model++ (PFGM++) which has achieved success in unconditional image generation and CT image denoising. Yet the PFCM can only be trained in distillation which limits the potential of the PFCM in many data modalities. The objective of this research was to create a method to train the PFC… ▽ More The Poisson Flow Consistency Model (PFCM) is a consistency-style model based on the robust Poisson Flow Generative Model++ (PFGM++) which has achieved success in unconditional image generation and CT image denoising. Yet the PFCM can only be trained in distillation which limits the potential of the PFCM in many data modalities. The objective of this research was to create a method to train the PFCM in isolation called Poisson Flow Consistency Training (PFCT). The perturbation kernel was leveraged to remove the pretrained PFGM++, and the sinusoidal discretization schedule and Beta noise distribution were introduced in order to facilitate adaptability and improve sample quality. The model was applied to the task of low dose computed tomography image denoising and improved the low dose image in terms of LPIPS and SSIM. It also displayed similar denoising effectiveness as models like the Consistency Model. PFCT is established as a valid method of training the PFCM from its effectiveness in denoising CT images, showing potential with competitive results to other generative models. Further study is needed in the precise optimization of PFCT and in its applicability to other generative modeling tasks. The framework of PFCT creates more flexibility for the ways in which a PFCM can be created and can be applied to the field of generative modeling. △ Less

Submitted 22 October, 2025; originally announced October 2025.

Comments: 5 pages, 3 figures, 1 table

MSC Class: 68T07 (Primary); 68T45 (Secondary)

GAPO: Robust Advantage Estimation for Real-World Code LLMs

Authors: Jianqing Zhang, Zhezheng Hao, Wei Xia, Hande Dong, Hong Wang, Chenxing Wei, Yuyan Zhou, Yubin Qi, Qiang Lin, Jian Cao

Abstract: Reinforcement learning (RL) is widely used for post-training large language models (LLMs) in code editing, where group-relative methods like GRPO are popular for their critic-free, normalized advantage estimation. However, in real-world code-editing scenarios, reward distributions are often skewed with unpredictable outliers, leading to distorted advantage computation and increased noise. To addre… ▽ More Reinforcement learning (RL) is widely used for post-training large language models (LLMs) in code editing, where group-relative methods like GRPO are popular for their critic-free, normalized advantage estimation. However, in real-world code-editing scenarios, reward distributions are often skewed with unpredictable outliers, leading to distorted advantage computation and increased noise. To address this issue, we propose Group Adaptive Policy Optimization (GAPO), which adaptively finds an outlier-free highest-density interval (HDI) per prompt and then uses the median of that interval as an adaptive Q to replace the group mean in advantage calculation. This adaptive Q robustly handles skewed distributions while remaining plug-and-play and efficient. We validate GAPO on nine instruction-tuned LLMs (3B-14B) using a large internal dataset of 51,844 real-world, history-aware code-editing tasks across 10 languages, demonstrating consistent improvements in exact match accuracy over GRPO and its variant DAPO. Code is publicly available. △ Less

Submitted 2 December, 2025; v1 submitted 21 October, 2025; originally announced October 2025.

HybridEP: Scaling Expert Parallelism to Cross-Datacenter Scenario via Hybrid Expert/Data Transmission

Authors: Weihao Yang, Hao Huang, Donglei Wu, Ningke Li, Yanqi Pan, Qiyang Zheng, Wen Xia, Shiyi Li, Qiang Wang

Abstract: Mixture-of-Experts (MoE) has become a popular architecture for scaling large models. However, the rapidly growing scale outpaces model training on a single DC, driving a shift toward a more flexible, cross-DC training paradigm. Under this, Expert Parallelism (EP) of MoE faces significant scalability issues due to the limited cross-DC bandwidth. Specifically, existing EP optimizations attempt to ov… ▽ More Mixture-of-Experts (MoE) has become a popular architecture for scaling large models. However, the rapidly growing scale outpaces model training on a single DC, driving a shift toward a more flexible, cross-DC training paradigm. Under this, Expert Parallelism (EP) of MoE faces significant scalability issues due to the limited cross-DC bandwidth. Specifically, existing EP optimizations attempt to overlap data communication and computation, which has little benefit in low-bandwidth scenarios due to a much longer data communication time. Therefore, the trends of cross-DC EP scaling is fast becoming a critical roadblock to the continued growth of MoE models. To address this, we propose HybridEP, a modeling-guided framework to optimize EP under constrained bandwidth. Our key idea is to dynamically transform the spatial placement of experts to reduce data communication traffic and frequency, thereby minimizing EP's communication overheads. However, it is non-trivial to find the optimal solution because it complicates the original communication pattern by mixing data and expert communication. We therefore build a stream-based model to determine the optimal transmission ratio. Guided by this, we incorporate two techniques: (1) domain-based partition to construct the mapping between hybrid patterns and specific communication topology at GPU level, and (2) parameter-efficient migration to further refine this topology by reducing expert transmission overhead and enlarging the domain size. Combining all these designs, HybridEP can be considered as a more general EP with better scalability. Experimental results show that HybridEP outperforms existing state-of-the-art MoE training systems by up to 5.6x under constrained bandwidth. We further compare HybridEP and EP on large-scale simulations. HybridEP achieves up to 1.45x speedup with 1k DCs under different bandwidths. △ Less

Submitted 22 October, 2025; originally announced October 2025.

From Preferences to Prejudice: The Role of Alignment Tuning in Shaping Social Bias in Video Diffusion Models

Authors: Zefan Cai, Haoyi Qiu, Haozhe Zhao, Ke Wan, Jiachen Li, Jiuxiang Gu, Wen Xiao, Nanyun Peng, Junjie Hu

Abstract: Recent advances in video diffusion models have significantly enhanced text-to-video generation, particularly through alignment tuning using reward models trained on human preferences. While these methods improve visual quality, they can unintentionally encode and amplify social biases. To systematically trace how such biases evolve throughout the alignment pipeline, we introduce VideoBiasEval, a c… ▽ More Recent advances in video diffusion models have significantly enhanced text-to-video generation, particularly through alignment tuning using reward models trained on human preferences. While these methods improve visual quality, they can unintentionally encode and amplify social biases. To systematically trace how such biases evolve throughout the alignment pipeline, we introduce VideoBiasEval, a comprehensive diagnostic framework for evaluating social representation in video generation. Grounded in established social bias taxonomies, VideoBiasEval employs an event-based prompting strategy to disentangle semantic content (actions and contexts) from actor attributes (gender and ethnicity). It further introduces multi-granular metrics to evaluate (1) overall ethnicity bias, (2) gender bias conditioned on ethnicity, (3) distributional shifts in social attributes across model variants, and (4) the temporal persistence of bias within videos. Using this framework, we conduct the first end-to-end analysis connecting biases in human preference datasets, their amplification in reward models, and their propagation through alignment-tuned video diffusion models. Our results reveal that alignment tuning not only strengthens representational biases but also makes them temporally stable, producing smoother yet more stereotyped portrayals. These findings highlight the need for bias-aware evaluation and mitigation throughout the alignment process to ensure fair and socially responsible video generation. △ Less

Submitted 20 October, 2025; originally announced October 2025.

Protenix-Mini+: efficient structure prediction model with scalable pairformer

Authors: Bo Qiang, Chengyue Gong, Xinshi Chen, Yuxuan Zhang, Wenzhi Xiao

Abstract: Lightweight inference is critical for biomolecular structure prediction and downstream tasks, enabling efficient real-world deployment and inference-time scaling for large-scale applications. While AF3 and its variants (e.g., Protenix, Chai-1) have advanced structure prediction results, they suffer from critical limitations: high inference latency and cubic time complexity with respect to token co… ▽ More Lightweight inference is critical for biomolecular structure prediction and downstream tasks, enabling efficient real-world deployment and inference-time scaling for large-scale applications. While AF3 and its variants (e.g., Protenix, Chai-1) have advanced structure prediction results, they suffer from critical limitations: high inference latency and cubic time complexity with respect to token count, both of which restrict scalability for large biomolecular complexes. To address the core challenge of balancing model efficiency and prediction accuracy, we introduce three key innovations: (1) compressing non-scalable operations to mitigate cubic time complexity, (2) removing redundant blocks across modules to reduce unnecessary overhead, and (3) adopting a few-step sampler for the atom diffusion module to accelerate inference. Building on these design principles, we develop Protenix-Mini+, a highly lightweight and scalable variant of the Protenix model. Within an acceptable range of performance degradation, it substantially improves computational efficiency. For example, in the case of low-homology single-chain proteins, Protenix-Mini+ experiences an intra-protein LDDT drop of approximately 3% relative to the full Protenix model -- an acceptable performance trade-off given its substantially 90%+ improved computational efficiency. △ Less

Submitted 15 October, 2025; v1 submitted 13 October, 2025; originally announced October 2025.

Self-Verifying Reflection Helps Transformers with CoT Reasoning

Authors: Zhongwei Yu, Wannian Xia, Xue Yan, Bo Xu, Haifeng Zhang, Yali Du, Jun Wang

Abstract: Advanced large language models (LLMs) frequently reflect in reasoning chain-of-thoughts (CoTs), where they self-verify the correctness of current solutions and explore alternatives. However, given recent findings that LLMs detect limited errors in CoTs, how reflection contributes to empirical improvements remains unclear. To analyze this issue, in this paper, we present a minimalistic reasoning fr… ▽ More Advanced large language models (LLMs) frequently reflect in reasoning chain-of-thoughts (CoTs), where they self-verify the correctness of current solutions and explore alternatives. However, given recent findings that LLMs detect limited errors in CoTs, how reflection contributes to empirical improvements remains unclear. To analyze this issue, in this paper, we present a minimalistic reasoning framework to support basic self-verifying reflection for small transformers without natural language, which ensures analytic clarity and reduces the cost of comprehensive experiments. Theoretically, we prove that self-verifying reflection guarantees improvements if verification errors are properly bounded. Experimentally, we show that tiny transformers, with only a few million parameters, benefit from self-verification in both training and reflective execution, reaching remarkable LLM-level performance in integer multiplication and Sudoku. Similar to LLM results, we find that reinforcement learning (RL) improves in-distribution performance and incentivizes frequent reflection for tiny transformers, yet RL mainly optimizes shallow statistical patterns without faithfully reducing verification errors. In conclusion, integrating generative transformers with discriminative verification inherently facilitates CoT reasoning, regardless of scaling and natural language. △ Less

Submitted 14 October, 2025; originally announced October 2025.

Comments: Accepted by NeurIPS2025

Co-TAP: Three-Layer Agent Interaction Protocol Technical Report

Authors: Shunyu An, Miao Wang, Yongchao Li, Dong Wan, Lina Wang, Ling Qin, Liqin Gao, Congyao Fan, Zhiyong Mao, Jiange Pu, Wenji Xia, Dong Zhao, Zhaohui Hao, Rui Hu, Ji Lu, Guiyue Zhou, Baoyu Tang, Yanqin Gao, Yongsheng Du, Daigang Xu, Lingjun Huang, Baoli Wang, Xiwen Zhang, Luyao Wang, Shilong Liu

Abstract: This paper proposes Co-TAP (T: Triple, A: Agent, P: Protocol), a three-layer agent interaction protocol designed to address the challenges faced by multi-agent systems across the three core dimensions of Interoperability, Interaction and Collaboration, and Knowledge Sharing. We have designed and proposed a layered solution composed of three core protocols: the Human-Agent Interaction Protocol (HAI… ▽ More This paper proposes Co-TAP (T: Triple, A: Agent, P: Protocol), a three-layer agent interaction protocol designed to address the challenges faced by multi-agent systems across the three core dimensions of Interoperability, Interaction and Collaboration, and Knowledge Sharing. We have designed and proposed a layered solution composed of three core protocols: the Human-Agent Interaction Protocol (HAI), the Unified Agent Protocol (UAP), and the Memory-Extraction-Knowledge Protocol (MEK). HAI focuses on the interaction layer, standardizing the flow of information between users, interfaces, and agents by defining a standardized, event-driven communication paradigm. This ensures the real-time performance, reliability, and synergy of interactions. As the core of the infrastructure layer, UAP is designed to break down communication barriers among heterogeneous agents through unified service discovery and protocol conversion mechanisms, thereby enabling seamless interconnection and interoperability of the underlying network. MEK, in turn, operates at the cognitive layer. By establishing a standardized ''Memory (M) - Extraction (E) - Knowledge (K)'' cognitive chain, it empowers agents with the ability to learn from individual experiences and form shareable knowledge, thereby laying the foundation for the realization of true collective intelligence. We believe this protocol framework will provide a solid engineering foundation and theoretical guidance for building the next generation of efficient, scalable, and intelligent multi-agent applications. △ Less

Submitted 28 October, 2025; v1 submitted 9 October, 2025; originally announced October 2025.

SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs

Authors: Dachuan Shi, Abedelkadir Asi, Keying Li, Xiangchi Yuan, Leyan Pan, Wenke Lee, Wen Xiao

Abstract: Recent work shows that, beyond discrete reasoning through explicit chain-of-thought steps, which are limited by the boundaries of natural languages, large language models (LLMs) can also reason continuously in latent space, allowing richer information per step and thereby improving token efficiency. Despite this promise, latent reasoning still faces two challenges, especially in training-free sett… ▽ More Recent work shows that, beyond discrete reasoning through explicit chain-of-thought steps, which are limited by the boundaries of natural languages, large language models (LLMs) can also reason continuously in latent space, allowing richer information per step and thereby improving token efficiency. Despite this promise, latent reasoning still faces two challenges, especially in training-free settings: 1) purely latent reasoning broadens the search distribution by maintaining multiple implicit paths, which diffuses probability mass, introduces noise, and impedes convergence to a single high-confidence solution, thereby hurting accuracy; and 2) overthinking persists even without explicit text, wasting tokens and degrading efficiency. To address these issues, we introduce SwiReasoning, a training-free framework for LLM reasoning which features two key innovations: 1) SwiReasoning dynamically switches between explicit and latent reasoning, guided by block-wise confidence estimated from entropy trends in next-token distributions, to balance exploration and exploitation and promote timely convergence. 2) By limiting the maximum number of thinking-block switches, SwiReasoning curbs overthinking and improves token efficiency across varying problem difficulties. On widely used mathematics and STEM benchmarks, SwiReasoning consistently improves average accuracy by 1.5%-2.8% across reasoning LLMs of different model families and scales. Furthermore, under constrained budgets, SwiReasoning improves average token efficiency by 56%-79%, with larger gains as budgets tighten. △ Less

Submitted 6 December, 2025; v1 submitted 6 October, 2025; originally announced October 2025.

Comments: Code: https://github.com/sdc17/SwiReasoning, Website: https://swireasoning.github.io/

What Is The Performance Ceiling of My Classifier? Utilizing Category-Wise Influence Functions for Pareto Frontier Analysis

Authors: Shahriar Kabir Nahin, Wenxiao Xiao, Joshua Liu, Anshuman Chhabra, Hongfu Liu

Abstract: Data-centric learning seeks to improve model performance from the perspective of data quality, and has been drawing increasing attention in the machine learning community. Among its key tools, influence functions provide a powerful framework to quantify the impact of individual training samples on model predictions, enabling practitioners to identify detrimental samples and retrain models on a cle… ▽ More Data-centric learning seeks to improve model performance from the perspective of data quality, and has been drawing increasing attention in the machine learning community. Among its key tools, influence functions provide a powerful framework to quantify the impact of individual training samples on model predictions, enabling practitioners to identify detrimental samples and retrain models on a cleaner dataset for improved performance. However, most existing work focuses on the question: "what data benefits the learning model?" In this paper, we take a step further and investigate a more fundamental question: "what is the performance ceiling of the learning model?" Unlike prior studies that primarily measure improvement through overall accuracy, we emphasize category-wise accuracy and aim for Pareto improvements, ensuring that every class benefits, rather than allowing tradeoffs where some classes improve at the expense of others. To address this challenge, we propose category-wise influence functions and introduce an influence vector that quantifies the impact of each training sample across all categories. Leveraging these influence vectors, we develop a principled criterion to determine whether a model can still be improved, and further design a linear programming-based sample reweighting framework to achieve Pareto performance improvements. Through extensive experiments on synthetic datasets, vision, and text benchmarks, we demonstrate the effectiveness of our approach in estimating and achieving a model's performance improvement across multiple categories of interest. △ Less

Submitted 4 October, 2025; originally announced October 2025.

AP2O-Coder: Human-Inspired Progressive Optimization to Fix LLM Code Errors

Authors: Jianqing Zhang, Wei Xia, Hande Dong, Qiang Lin, Jian Cao

Abstract: LLMs' code generation capabilities have yielded substantial improvements in the effectiveness of programming tasks. However, LLM-generated code still suffers from compilation and runtime errors. Existing offline preference optimization methods primarily focus on enhancing LLMs' coding abilities using pass/fail signals in the preference data, overlooking the deep-level error types in the failed cod… ▽ More LLMs' code generation capabilities have yielded substantial improvements in the effectiveness of programming tasks. However, LLM-generated code still suffers from compilation and runtime errors. Existing offline preference optimization methods primarily focus on enhancing LLMs' coding abilities using pass/fail signals in the preference data, overlooking the deep-level error types in the failed codes. To address this, we propose Adaptively Progressive Preference Optimization (AP2O) for coding (i.e., AP2O-Coder), a method that guides LLMs adaptively and methodically to reduce code errors for code generation. Specifically, we construct an error notebook from failed codes and progressively optimize the LLM to correct errors type by type. Furthermore, we adaptively replay error types to tailor to the LLM's changing weaknesses throughout the training process. Through extensive experiments on both code and general LLMs (Llama, Qwen, and DeepSeek series) with parameters ranging from 0.5B to 34B, our AP2O-Coder improves code generation performance by up to 3% in pass@k while using less preference data. Code: https://github.com/TsingZ0/AP2O △ Less

Submitted 23 November, 2025; v1 submitted 30 September, 2025; originally announced October 2025.

Comments: Accepted by AAAI2026

Safe Motion Planning and Control Using Predictive and Adaptive Barrier Methods for Autonomous Surface Vessels

Authors: Alejandro Gonzalez-Garcia, Wei Xiao, Wei Wang, Alejandro Astudillo, Wilm Decré, Jan Swevers, Carlo Ratti, Daniela Rus

Abstract: Safe motion planning is essential for autonomous vessel operations, especially in challenging spaces such as narrow inland waterways. However, conventional motion planning approaches are often computationally intensive or overly conservative. This paper proposes a safe motion planning strategy combining Model Predictive Control (MPC) and Control Barrier Functions (CBFs). We introduce a time-varyin… ▽ More Safe motion planning is essential for autonomous vessel operations, especially in challenging spaces such as narrow inland waterways. However, conventional motion planning approaches are often computationally intensive or overly conservative. This paper proposes a safe motion planning strategy combining Model Predictive Control (MPC) and Control Barrier Functions (CBFs). We introduce a time-varying inflated ellipse obstacle representation, where the inflation radius is adjusted depending on the relative position and attitude between the vessel and the obstacle. The proposed adaptive inflation reduces the conservativeness of the controller compared to traditional fixed-ellipsoid obstacle formulations. The MPC solution provides an approximate motion plan, and high-order CBFs ensure the vessel's safety using the varying inflation radius. Simulation and real-world experiments demonstrate that the proposed strategy enables the fully-actuated autonomous robot vessel to navigate through narrow spaces in real time and resolve potential deadlocks, all while ensuring safety. △ Less

Submitted 1 October, 2025; originally announced October 2025.

Comments: IROS 2025

DyMoDreamer: World Modeling with Dynamic Modulation

Authors: Boxuan Zhang, Runqing Wang, Wei Xiao, Weipu Zhang, Jian Sun, Gao Huang, Jie Chen, Gang Wang

Abstract: A critical bottleneck in deep reinforcement learning (DRL) is sample inefficiency, as training high-performance agents often demands extensive environmental interactions. Model-based reinforcement learning (MBRL) mitigates this by building world models that simulate environmental dynamics and generate synthetic experience, improving sample efficiency. However, conventional world models process obs… ▽ More A critical bottleneck in deep reinforcement learning (DRL) is sample inefficiency, as training high-performance agents often demands extensive environmental interactions. Model-based reinforcement learning (MBRL) mitigates this by building world models that simulate environmental dynamics and generate synthetic experience, improving sample efficiency. However, conventional world models process observations holistically, failing to decouple dynamic objects and temporal features from static backgrounds. This approach is computationally inefficient, especially for visual tasks where dynamic objects significantly influence rewards and decision-making performance. To address this, we introduce DyMoDreamer, a novel MBRL algorithm that incorporates a dynamic modulation mechanism to improve the extraction of dynamic features and enrich the temporal information. DyMoDreamer employs differential observations derived from a novel inter-frame differencing mask, explicitly encoding object-level motion cues and temporal dynamics. Dynamic modulation is modeled as stochastic categorical distributions and integrated into a recurrent state-space model (RSSM), enhancing the model's focus on reward-relevant dynamics. Experiments demonstrate that DyMoDreamer sets a new state-of-the-art on the Atari $100$k benchmark with a $156.6$\% mean human-normalized score, establishes a new record of $832$ on the DeepMind Visual Control Suite, and gains a $9.5$\% performance improvement after $1$M steps on the Crafter benchmark. Our code is released at https://github.com/Ultraman-Tiga1/DyMoDreamer. △ Less

Submitted 29 September, 2025; originally announced September 2025.

LATTS: Locally Adaptive Test-Time Scaling

Authors: Theo Uscidda, Matthew Trager, Michael Kleinman, Aditya Chattopadhyay, Wei Xia, Stefano Soatto

Abstract: One common strategy for improving the performance of Large Language Models (LLMs) on downstream tasks involves using a \emph{verifier model} to either select the best answer from a pool of candidates or to steer the auto-regressive generation process towards better outputs. This class of methods typically results in improved accuracy at the cost of increased computation at test-time, a paradigm kn… ▽ More One common strategy for improving the performance of Large Language Models (LLMs) on downstream tasks involves using a \emph{verifier model} to either select the best answer from a pool of candidates or to steer the auto-regressive generation process towards better outputs. This class of methods typically results in improved accuracy at the cost of increased computation at test-time, a paradigm known as \emph{test-time scaling}. However, most existing approaches increase computation uniformly across all samples and generation steps, without considering the complexity of individual instances, leading to inefficient resource use. We address this limitation by proposing an approach, called \emph{Locally Adaptive Test-Time Scaling (LATTS)}, that allocates variable compute across generation steps. Specifically, at each generation step, LATTS employs a verifier-based acceptance criterion to decide whether to resample, backtrack, restart, or stop the generation process. This criterion effectively adjusts the per-step computational effort based on a precise notion of \emph{local difficulty} derived from the verifier model. Empirical results show that LATTS achieves significantly superior accuracy--compute tradeoffs compared to standard verifier-based methods. △ Less

Submitted 16 September, 2025; originally announced September 2025.

From Easy to Hard: The MIR Benchmark for Progressive Interleaved Multi-Image Reasoning

Authors: Hang Du, Jiayang Zhang, Guoshun Nan, Wendi Deng, Zhenyan Chen, Chenyang Zhang, Wang Xiao, Shan Huang, Yuqi Pan, Tao Qi, Sicong Leng

Abstract: Multi-image Interleaved Reasoning aims to improve Multi-modal Large Language Models (MLLMs) ability to jointly comprehend and reason across multiple images and their associated textual contexts, introducing unique challenges beyond single-image or non-interleaved multi-image tasks. While current multi-image benchmarks overlook interleaved textual contexts and neglect distinct relationships between… ▽ More Multi-image Interleaved Reasoning aims to improve Multi-modal Large Language Models (MLLMs) ability to jointly comprehend and reason across multiple images and their associated textual contexts, introducing unique challenges beyond single-image or non-interleaved multi-image tasks. While current multi-image benchmarks overlook interleaved textual contexts and neglect distinct relationships between individual images and their associated texts, enabling models to reason over multi-image interleaved data may significantly enhance their comprehension of complex scenes and better capture cross-modal correlations. To bridge this gap, we introduce a novel benchmark MIR, requiring joint reasoning over multiple images accompanied by interleaved textual contexts to accurately associate image regions with corresponding texts and logically connect information across images. To enhance MLLMs ability to comprehend multi-image interleaved data, we introduce reasoning steps for each instance within the benchmark and propose a stage-wise curriculum learning strategy. This strategy follows an "easy to hard" approach, progressively guiding models from simple to complex scenarios, thereby enhancing their ability to handle challenging tasks. Extensive experiments benchmarking multiple MLLMs demonstrate that our method significantly enhances models reasoning performance on MIR and other established benchmarks. We believe that MIR will encourage further research into multi-image interleaved reasoning, facilitating advancements in MLLMs capability to handle complex inter-modal tasks. △ Less

Submitted 15 October, 2025; v1 submitted 21 September, 2025; originally announced September 2025.

Comments: Accepted by ICCV 2025

Robust LLM Training Infrastructure at ByteDance

Authors: Borui Wan, Gaohong Liu, Zuquan Song, Jun Wang, Yun Zhang, Guangming Sheng, Shuguang Wang, Houmin Wei, Chenyuan Wang, Weiqiang Lou, Xi Yang, Mofan Zhang, Kaihua Jiang, Cheng Ren, Xiaoyun Zhi, Menghan Yu, Zhe Nan, Zhuolin Zheng, Baoquan Zhong, Qinlong Wang, Huan Yu, Jinxin Chi, Wang Zhang, Yuhan Li, Zixian Du , et al. (10 additional authors not shown)

Abstract: The training scale of large language models (LLMs) has reached tens of thousands of GPUs and is still continuously expanding, enabling faster learning of larger models. Accompanying the expansion of the resource scale is the prevalence of failures (CUDA error, NaN values, job hang, etc.), which poses significant challenges to training stability. Any large-scale LLM training infrastructure should s… ▽ More The training scale of large language models (LLMs) has reached tens of thousands of GPUs and is still continuously expanding, enabling faster learning of larger models. Accompanying the expansion of the resource scale is the prevalence of failures (CUDA error, NaN values, job hang, etc.), which poses significant challenges to training stability. Any large-scale LLM training infrastructure should strive for minimal training interruption, efficient fault diagnosis, and effective failure tolerance to enable highly efficient continuous training. This paper presents ByteRobust, a large-scale GPU infrastructure management system tailored for robust and stable training of LLMs. It exploits the uniqueness of LLM training process and gives top priorities to detecting and recovering failures in a routine manner. Leveraging parallelisms and characteristics of LLM training, ByteRobust enables high-capacity fault tolerance, prompt fault demarcation, and localization with an effective data-driven approach, comprehensively ensuring continuous and efficient training of LLM tasks. ByteRobust is deployed on a production GPU platform and achieves 97% ETTR for a three-month training job on 9,600 GPUs. △ Less

Submitted 20 October, 2025; v1 submitted 19 September, 2025; originally announced September 2025.

Efficient Pre-Training of LLMs via Topology-Aware Communication Alignment on More Than 9600 GPUs

Authors: Guoliang He, Youhe Jiang, Wencong Xiao, Kaihua Jiang, Shuguang Wang, Jun Wang, Zixian Du, Zhuo Jiang, Xinlei Zhang, Binhang Yuan, Eiko Yoneki

Abstract: The scaling law for large language models (LLMs) depicts that the path towards machine intelligence necessitates training at large scale. Thus, companies continuously build large-scale GPU clusters, and launch training jobs that span over thousands of computing nodes. However, LLM pre-training presents unique challenges due to its complex communication patterns, where GPUs exchange data in sparse… ▽ More The scaling law for large language models (LLMs) depicts that the path towards machine intelligence necessitates training at large scale. Thus, companies continuously build large-scale GPU clusters, and launch training jobs that span over thousands of computing nodes. However, LLM pre-training presents unique challenges due to its complex communication patterns, where GPUs exchange data in sparse yet high-volume bursts within specific groups. Inefficient resource scheduling exacerbates bandwidth contention, leading to suboptimal training performance. This paper presents Arnold, a scheduling system summarizing our experience to effectively align LLM communication patterns with data center topology at scale. An in-depth characteristic study is performed to identify the impact of physical network topology to LLM pre-training jobs. Based on the insights, we develop a scheduling algorithm to effectively align communication patterns with the physical network topology in modern data centers. Through simulation experiments, we show the effectiveness of our algorithm in reducing the maximum spread of communication groups by up to $1.67$x. In production training, our scheduling system improves the end-to-end performance by $10.6\%$ when training with more than $9600$ GPUs, a significant improvement for our training pipeline. △ Less

Submitted 19 September, 2025; originally announced September 2025.

Comments: NeurIPS 2025

Breathing and Semantic Pause Detection and Exertion-Level Classification in Post-Exercise Speech

Authors: Yuyu Wang, Wuyue Xia, Huaxiu Yao, Jingping Nie

Abstract: Post-exercise speech contains rich physiological and linguistic cues, often marked by semantic pauses, breathing pauses, and combined breathing-semantic pauses. Detecting these events enables assessment of recovery rate, lung function, and exertion-related abnormalities. However, existing works on identifying and distinguishing different types of pauses in this context are limited. In this work, b… ▽ More Post-exercise speech contains rich physiological and linguistic cues, often marked by semantic pauses, breathing pauses, and combined breathing-semantic pauses. Detecting these events enables assessment of recovery rate, lung function, and exertion-related abnormalities. However, existing works on identifying and distinguishing different types of pauses in this context are limited. In this work, building on a recently released dataset with synchronized audio and respiration signals, we provide systematic annotations of pause types. Using these annotations, we systematically conduct exploratory breathing and semantic pause detection and exertion-level classification across deep learning models (GRU, 1D CNN-LSTM, AlexNet, VGG16), acoustic features (MFCC, MFB), and layer-stratified Wav2Vec2 representations. We evaluate three setups-single feature, feature fusion, and a two-stage detection-classification cascade-under both classification and regression formulations. Results show per-type detection accuracy up to 89$\%$ for semantic, 55$\%$ for breathing, 86$\%$ for combined pauses, and 73$\%$overall, while exertion-level classification achieves 90.5$\%$ accuracy, outperformin prior work. △ Less

Submitted 18 September, 2025; originally announced September 2025.

Comments: 6 pages, 3rd ACM International Workshop on Intelligent Acoustic Systems and Applications (IASA 25)

Integrating Trajectory Optimization and Reinforcement Learning for Quadrupedal Jumping with Terrain-Adaptive Landing

Authors: Renjie Wang, Shangke Lyu, Xin Lang, Wei Xiao, Donglin Wang

Abstract: Jumping constitutes an essential component of quadruped robots' locomotion capabilities, which includes dynamic take-off and adaptive landing. Existing quadrupedal jumping studies mainly focused on the stance and flight phase by assuming a flat landing ground, which is impractical in many real world cases. This work proposes a safe landing framework that achieves adaptive landing on rough terrains… ▽ More Jumping constitutes an essential component of quadruped robots' locomotion capabilities, which includes dynamic take-off and adaptive landing. Existing quadrupedal jumping studies mainly focused on the stance and flight phase by assuming a flat landing ground, which is impractical in many real world cases. This work proposes a safe landing framework that achieves adaptive landing on rough terrains by combining Trajectory Optimization (TO) and Reinforcement Learning (RL) together. The RL agent learns to track the reference motion generated by TO in the environments with rough terrains. To enable the learning of compliant landing skills on challenging terrains, a reward relaxation strategy is synthesized to encourage exploration during landing recovery period. Extensive experiments validate the accurate tracking and safe landing skills benefiting from our proposed method in various scenarios. △ Less

Submitted 16 September, 2025; originally announced September 2025.

Comments: Accepted by IROS 2025

Robust Online Residual Refinement via Koopman-Guided Dynamics Modeling

Authors: Zhefei Gong, Shangke Lyu, Pengxiang Ding, Wei Xiao, Donglin Wang

Abstract: Imitation learning (IL) enables efficient skill acquisition from demonstrations but often struggles with long-horizon tasks and high-precision control due to compounding errors. Residual policy learning offers a promising, model-agnostic solution by refining a base policy through closed-loop corrections. However, existing approaches primarily focus on local corrections to the base policy, lacking… ▽ More Imitation learning (IL) enables efficient skill acquisition from demonstrations but often struggles with long-horizon tasks and high-precision control due to compounding errors. Residual policy learning offers a promising, model-agnostic solution by refining a base policy through closed-loop corrections. However, existing approaches primarily focus on local corrections to the base policy, lacking a global understanding of state evolution, which limits robustness and generalization to unseen scenarios. To address this, we propose incorporating global dynamics modeling to guide residual policy updates. Specifically, we leverage Koopman operator theory to impose linear time-invariant structure in a learned latent space, enabling reliable state transitions and improved extrapolation for long-horizon prediction and unseen environments. We introduce KORR (Koopman-guided Online Residual Refinement), a simple yet effective framework that conditions residual corrections on Koopman-predicted latent states, enabling globally informed and stable action refinement. We evaluate KORR on long-horizon, fine-grained robotic furniture assembly tasks under various perturbations. Results demonstrate consistent gains in performance, robustness, and generalization over strong baselines. Our findings further highlight the potential of Koopman-based modeling to bridge modern learning methods with classical control theory. △ Less

Submitted 15 September, 2025; originally announced September 2025.

TrajBooster: Boosting Humanoid Whole-Body Manipulation via Trajectory-Centric Learning

Authors: Jiacheng Liu, Pengxiang Ding, Qihang Zhou, Yuxuan Wu, Da Huang, Zimian Peng, Wei Xiao, Weinan Zhang, Lixin Yang, Cewu Lu, Donglin Wang

Abstract: Recent Vision-Language-Action models show potential to generalize across embodiments but struggle to quickly align with a new robot's action space when high-quality demonstrations are scarce, especially for bipedal humanoids. We present TrajBooster, a cross-embodiment framework that leverages abundant wheeled-humanoid data to boost bipedal VLA. Our key idea is to use end-effector trajectories as a… ▽ More Recent Vision-Language-Action models show potential to generalize across embodiments but struggle to quickly align with a new robot's action space when high-quality demonstrations are scarce, especially for bipedal humanoids. We present TrajBooster, a cross-embodiment framework that leverages abundant wheeled-humanoid data to boost bipedal VLA. Our key idea is to use end-effector trajectories as a morphology-agnostic interface. TrajBooster (i) extracts 6D dual-arm end-effector trajectories from real-world wheeled humanoids, (ii) retargets them in simulation to Unitree G1 with a whole-body controller trained via a heuristic-enhanced harmonized online DAgger to lift low-dimensional trajectory references into feasible high-dimensional whole-body actions, and (iii) forms heterogeneous triplets that couple source vision/language with target humanoid-compatible actions to post-pre-train a VLA, followed by only 10 minutes of teleoperation data collection on the target humanoid domain. Deployed on Unitree G1, our policy achieves beyond-tabletop household tasks, enabling squatting, cross-height manipulation, and coordinated whole-body motion with markedly improved robustness and generalization. Results show that TrajBooster allows existing wheeled-humanoid data to efficiently strengthen bipedal humanoid VLA performance, reducing reliance on costly same-embodiment data while enhancing action space understanding and zero-shot skill transfer capabilities. For more details, For more details, please refer to our \href{https://jiachengliu3.github.io/TrajBooster/}. △ Less

Submitted 16 September, 2025; v1 submitted 15 September, 2025; originally announced September 2025.

Mycroft: Tracing Dependencies in Collective Communication Towards Reliable LLM Training

Authors: Yangtao Deng, Lei Zhang, Qinlong Wang, Xiaoyun Zhi, Xinlei Zhang, Zhuo Jiang, Haohan Xu, Lei Wang, Zuquan Song, Gaohong Liu, Yang Bai, Shuguang Wang, Wencong Xiao, Jianxi Ye, Minlan Yu, Hong Xu

Abstract: Reliability is essential for ensuring efficiency in LLM training. However, many real-world reliability issues remain difficult to resolve, resulting in wasted resources and degraded model performance. Unfortunately, today's collective communication libraries operate as black boxes, hiding critical information needed for effective root cause analysis. We propose Mycroft, a lightweight distributed t… ▽ More Reliability is essential for ensuring efficiency in LLM training. However, many real-world reliability issues remain difficult to resolve, resulting in wasted resources and degraded model performance. Unfortunately, today's collective communication libraries operate as black boxes, hiding critical information needed for effective root cause analysis. We propose Mycroft, a lightweight distributed tracing and root cause analysis system designed to address previously hidden reliability issues in collective communication. Mycroft's key idea is to trace collective communication states and leverage internal control and data dependencies to resolve reliability problems in LLM training. Mycroft has been deployed at ByteDance for over six months to debug collective communication related issues at runtime. It detected anomalies within 15 seconds in 90% of cases and identified the root cause within 20 seconds in 60% of cases. We also conducted extensive fault injection experiments to demonstrate Mycroft's capability and efficiency. △ Less

Submitted 3 September, 2025; originally announced September 2025.

A-MHA*: Anytime Multi-Heuristic A*

Authors: Ramkumar Natarajan, Muhammad Suhail Saleem, William Xiao, Sandip Aine, Howie Choset, Maxim Likhachev

Abstract: Designing good heuristic functions for graph search requires adequate domain knowledge. It is often easy to design heuristics that perform well and correlate with the underlying true cost-to-go values in certain parts of the search space but these may not be admissible throughout the domain thereby affecting the optimality guarantees of the search. Bounded suboptimal search using several such part… ▽ More Designing good heuristic functions for graph search requires adequate domain knowledge. It is often easy to design heuristics that perform well and correlate with the underlying true cost-to-go values in certain parts of the search space but these may not be admissible throughout the domain thereby affecting the optimality guarantees of the search. Bounded suboptimal search using several such partially good but inadmissible heuristics was developed in Multi-Heuristic A* (MHA*). Although MHA* leverages multiple inadmissible heuristics to potentially generate a faster suboptimal solution, the original version does not improve the solution over time. It is a one shot algorithm that requires careful setting of inflation factors to obtain a desired one time solution. In this work, we tackle this issue by extending MHA* to an anytime version that finds a feasible suboptimal solution quickly and continually improves it until time runs out. Our work is inspired from the Anytime Repairing A* (ARA*) algorithm. We prove that our precise adaptation of ARA* concepts in the MHA* framework preserves the original suboptimal and completeness guarantees and enhances MHA* to perform in an anytime fashion. Furthermore, we report the performance of A-MHA* in 3-D path planning domain and sliding tiles puzzle and compare against MHA* and other anytime algorithms. △ Less

Submitted 29 August, 2025; originally announced August 2025.