Siqi Yang

Shurui Xu, Siqi Yang, Weiping Ding et al.

Clinical diagnosis of meniscus injuries requires radiologists to integrate volumetric MRI evidence with patient context (e.g., sex, age, BMI) and to produce structured diagnostic reports. Existing knee MRI benchmarks are typically unimodal and rely on coarse labels, limiting their ability to evaluate holistic clinical reasoning. We introduce MeniOmni, a structured multimodal benchmark for meniscus injury assessment, consisting of 746 multi-center MRI studies with tri-planar volumetric inputs, Clinical Priors, and expert-annotated clinical text. MeniOmni supports two tasks: (1) fine-grained Stoller severity grading and (2) diagnostic report generation. We further propose risk-aware ordinal evaluation and a semantic consistency metric (Meni-Score) to better reflect clinical relevance. Baseline experiments show that incorporating Clinical Priors improves grading performance and reduces severe errors, highlighting the value of multimodal context for safer assessment. Code and data are available at https://github.com/ShuruiXu/MeniOmni.

Siqi Yang, Zilve Gao, Haibo Qiu et al.

Multimodal Large Language Models (MLLMs) demonstrate significant potential but remain brittle in complex, long-chain visual reasoning tasks. A critical failure mode is "visual forgetting", where models progressively lose visual grounding as reasoning extends, a phenomenon aptly described as "think longer, see less". We posit this failure stems from current training paradigms prematurely entangling two distinct cognitive skills: (1) abstract logical reasoning "how-to-think") and (2) strategic visual perception ("when-to-look"). This creates a foundational cold-start deficiency -- weakening abstract reasoning -- and a strategic perception deficit, as models lack a policy for when to perceive. In this paper, we propose a novel curriculum-based framework to disentangle these skills. First, we introduce a disentangled Supervised Fine-Tuning (SFT) curriculum that builds a robust abstract reasoning backbone on text-only data before anchoring it to vision with a novel Perception-Grounded Chain-of-Thought (PG-CoT) paradigm. Second, we resolve the strategic perception deficit by formulating timing as a reinforcement learning problem. We design a Pivotal Perception Reward that teaches the model when to look by coupling perceptual actions to linguistic markers of cognitive uncertainty (e.g., "wait", "verify"), thereby learning an autonomous grounding policy. Our contributions include the formalization of these two deficiencies and the development of a principled, two-stage framework to address them, transforming the model from a heuristic-driven observer to a strategic, grounded reasoner. \textbf{Code}: \url{https://github.com/gaozilve-max/learning-when-to-look}.

Meituan LongCat Team, Bairui Wang, Bayan et al.

We introduce LongCat-Flash-Omni, a state-of-the-art open-source omni-modal model with 560 billion parameters, excelling at real-time audio-visual interaction. By adopting a curriculum-inspired progressive training strategy that transitions from simpler to increasingly complex modality sequence modeling tasks, LongCat-Flash-Omni attains comprehensive multimodal capabilities while maintaining strong unimodal capability. Building upon LongCat-Flash, which adopts a high-performance Shortcut-connected Mixture-of-Experts (MoE) architecture with zero-computation experts, LongCat-Flash-Omni integrates efficient multimodal perception and speech reconstruction modules. Despite its immense size of 560B parameters (with 27B activated), LongCat-Flash-Omni achieves low-latency real-time audio-visual interaction. For training infrastructure, we developed a modality-decoupled parallelism scheme specifically designed to manage the data and model heterogeneity inherent in large-scale multimodal training. This innovative approach demonstrates exceptional efficiency by sustaining over 90% of the throughput achieved by text-only training. Extensive evaluations show that LongCat-Flash-Omni achieves state-of-the-art performance on omni-modal benchmarks among open-source models. Furthermore, it delivers highly competitive results across a wide range of modality-specific tasks, including text, image, and video understanding, as well as audio understanding and generation. We provide a comprehensive overview of the model architecture design, training procedures, and data strategies, and open-source the model to foster future research and development in the community.

Siqi Yang, Zejun Yang, Zhisheng Wang

Dancing with music is always an essential human art form to express emotion. Due to the high temporal-spacial complexity, long-term 3D realist dance generation synchronized with music is challenging. Existing methods suffer from the freezing problem when generating long-term dances due to error accumulation and training-inference discrepancy. To address this, we design a conditional diffusion model, LongDanceDiff, for this sequence-to-sequence long-term dance generation, addressing the challenges of temporal coherency and spatial constraint. LongDanceDiff contains a transformer-based diffusion model, where the input is a concatenation of music, past motions, and noised future motions. This partial noising strategy leverages the full-attention mechanism and learns the dependencies among music and past motions. To enhance the diversity of generated dance motions and mitigate the freezing problem, we introduce a mutual information minimization objective that regularizes the dependency between past and future motions. We also address common visual quality issues in dance generation, such as foot sliding and unsmooth motion, by incorporating spatial constraints through a Global-Trajectory Modulation (GTM) layer and motion perceptual losses, thereby improving the smoothness and naturalness of motion generation. Extensive experiments demonstrate a significant improvement in our approach over the existing state-of-the-art methods. We plan to release our codes and models soon.

Chu Zhou, Siqi Yang, Kailong Zhang et al.

Conventional RGB-based high dynamic range (HDR) imaging faces a fundamental trade-off between motion artifacts in multi-exposure captures and irreversible information loss in single-shot techniques. Modulo sensors offer a promising alternative by encoding theoretically unbounded dynamic range into wrapped measurements. However, existing modulo solutions remain bottlenecked by iterative unwrapping overhead and hardware constraints limiting them to low-speed, grayscale capture. In this work, we present a complete modulo-based HDR imaging system that enables high-speed, full-color HDR acquisition by synergistically advancing both the sensing formulation and the unwrapping algorithm. At the core of our approach is an exposure-decoupled formulation of modulo imaging that allows multiple measurements to be interleaved in time, preserving a clean, observation-wise measurement model. Building upon this, we introduce an iteration-free unwrapping algorithm that integrates diffusion-based generative priors with the physical least absolute remainder property of modulo images, supporting highly efficient, physics-consistent HDR reconstruction. Finally, to validate the practical viability of our system, we demonstrate a proof-of-concept hardware implementation based on modulo-encoded spike streams. This setup preserves the native high temporal resolution of spike cameras, achieving 1000 FPS full-color imaging while reducing output data bandwidth from approximately 20 Gbps to 6 Gbps. Extensive evaluations indicate that our coordinated approach successfully overcomes key systemic bottlenecks, demonstrating the feasibility of deploying modulo imaging in dynamic scenarios.

Hanxuan Chen, Jie Zheng, Siqi Yang et al.

Vision-and-Language Navigation for Unmanned Aerial Vehicles (UAV-VLN) represents a pivotal challenge in embodied artificial intelligence, focused on enabling UAVs to interpret high-level human commands and execute long-horizon tasks in complex 3D environments. This paper provides a comprehensive and structured survey of the field, from its formal task definition to the current state of the art. We establish a methodological taxonomy that charts the technological evolution from early modular and deep learning approaches to contemporary agentic systems driven by large foundation models, including Vision-Language Models (VLMs), Vision-Language-Action (VLA) models, and the emerging integration of generative world models with VLA architectures for physically-grounded reasoning. The survey systematically reviews the ecosystem of essential resources simulators, datasets, and evaluation metrics that facilitates standardized research. Furthermore, we conduct a critical analysis of the primary challenges impeding real-world deployment: the simulation-to-reality gap, robust perception in dynamic outdoor settings, reasoning with linguistic ambiguity, and the efficient deployment of large models on resource-constrained hardware. By synthesizing current benchmarks and limitations, this survey concludes by proposing a forward-looking research roadmap to guide future inquiry into key frontiers such as multi-agent swarm coordination and air-ground collaborative robotics.

Chao Yi, Dian Chen, Gaoyang Guo et al.

Large language models (LLMs) have demonstrated remarkable potential in transforming recommender systems from implicit behavioral pattern matching to explicit intent reasoning. While RecGPT-V1 successfully pioneered this paradigm by integrating LLM-based reasoning into user interest mining and item tag prediction, it suffers from four fundamental limitations: (1) computational inefficiency and cognitive redundancy across multiple reasoning routes; (2) insufficient explanation diversity in fixed-template generation; (3) limited generalization under supervised learning paradigms; and (4) simplistic outcome-focused evaluation that fails to match human standards. To address these challenges, we present RecGPT-V2 with four key innovations. First, a Hierarchical Multi-Agent System restructures intent reasoning through coordinated collaboration, eliminating cognitive duplication while enabling diverse intent coverage. Combined with Hybrid Representation Inference that compresses user-behavior contexts, our framework reduces GPU consumption by 60% and improves exclusive recall from 9.39% to 10.99%. Second, a Meta-Prompting framework dynamically generates contextually adaptive prompts, improving explanation diversity by +7.3%. Third, constrained reinforcement learning mitigates multi-reward conflicts, achieving +24.1% improvement in tag prediction and +13.0% in explanation acceptance. Fourth, an Agent-as-a-Judge framework decomposes assessment into multi-step reasoning, improving human preference alignment. Online A/B tests on Taobao demonstrate significant improvements: +2.98% CTR, +3.71% IPV, +2.19% TV, and +11.46% NER. RecGPT-V2 establishes both the technical feasibility and commercial viability of deploying LLM-powered intent reasoning at scale, bridging the gap between cognitive exploration and industrial utility.

Haojie Zheng, Yixin Yang, Siqi Yang et al.

Recent diffusion-based methods have achieved impressive progress in video content manipulation. However, they typically ignore the accompanying audio, leaving the audio disjointed from the edited results. In this paper, we propose InstructAV2AV, the first end-to-end framework for instruction-guided audio-video joint editing. We first develop a scalable data synthesis pipeline and construct InsAVE-80K, the first large-scale audio-video editing dataset with high-quality source-to-target pairs. With this data foundation, we adapt an audio-video generation backbone to leverage its robust priors. We concatenate the audio-video input with noisy latent codes to anchor the source context, propose the source-instruction gated attention to improve instruction following and content preservation, and introduce a two-stage training strategy to effectively transfer these pre-trained priors. Extensive experiments demonstrate that InstructAV2AV outperforms state-of-the-art methods across 11 metrics spanning three aspects on two evaluation sets, highlighting its potential for controllable content creation. Project page: https://hjzheng.net/projects/InstructAV2AV/.

Haotong Yang, Zitong Wang, Shijia Kang et al.

While Large Language Models (LLMs) have demonstrated strong math reasoning abilities through Reinforcement Learning with *Verifiable Rewards* (RLVR), many advanced mathematical problems are proof-based, with no guaranteed way to determine the authenticity of a proof by simple answer matching. To enable automatic verification, a Reward Model (RM) capable of reliably evaluating full proof processes is required. In this work, we design a *scalable* data-construction pipeline that, with minimal human effort, leverages LLMs to generate a large quantity of high-quality "**question-proof-check**" triplet data. By systematically varying problem sources, generation methods, and model configurations, we create diverse problem-proof pairs spanning multiple difficulty levels, linguistic styles, and error types, subsequently filtered through hierarchical human review for label alignment. Utilizing these data, we train a proof-checking RM, incorporating additional process reward and token weight balance to stabilize the RL process. Our experiments validate the model's scalability and strong performance from multiple perspectives, including reward accuracy, generalization ability and test-time guidance, providing important practical recipes and tools for strengthening LLM mathematical capabilities.

Kun Chen, Peng Shi, Fanfan Liu et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a critical method for enhancing the reasoning capabilities of Large Language Models (LLMs). However, continuous training often leads to policy entropy collapse, characterized by a rapid decay in entropy that results in premature overconfidence, reduced output diversity, and vanishing gradient norms that inhibit learning. Gradient-Preserving Clipping is a primary factor influencing these dynamics, but existing mitigation strategies are largely static and lack a framework connecting clipping mechanisms to precise entropy control. This paper proposes reshaping entropy control in RL from the perspective of Gradient-Preserving Clipping. We first theoretically and empirically verify the contributions of specific importance sampling ratio regions to entropy growth and reduction. Leveraging these findings, we introduce a novel regulation mechanism using dynamic clipping threshold to precisely manage entropy. Furthermore, we design and evaluate dynamic entropy control strategies, including increase-then-decrease, decrease-increase-decrease, and oscillatory decay. Experimental results demonstrate that these strategies effectively mitigate entropy collapse, and achieve superior performance across multiple benchmarks.

Fanfan Liu, Youyang Yin, Peng Shi et al.

Recent applications of Reinforcement Learning with Verifiable Rewards (RLVR) to Large Language Models (LLMs) and Vision-Language Models (VLMs) have demonstrated significant success in enhancing reasoning capabilities for complex tasks. During RLVR training, an increase in response length is often regarded as a key factor contributing to the growth of reasoning ability. However, the patterns of change in response length vary significantly across different RLVR algorithms during the training process. To provide a fundamental explanation for these variations, this paper conducts an in-depth analysis of the components of mainstream RLVR algorithms. We present a theoretical analysis of the factors influencing response length and validate our theory through extensive experimentation. Building upon these theoretical findings, we propose the Length-Unbiased Sequence Policy Optimization (LUSPO) algorithm. Specifically, we rectify the length bias inherent in Group Sequence Policy Optimization (GSPO), rendering its loss function unbiased with respect to response length and thereby resolving the issue of response length collapse. We conduct extensive experiments across mathematical reasoning benchmarks and multimodal reasoning scenarios, where LUSPO consistently achieves superior performance. Empirical results demonstrate that LUSPO represents a novel, state-of-the-art optimization strategy compared to existing methods such as GRPO and GSPO.

Siqi Yang, Zonghui Li, Chaoqun You et al.

Low-Earth Orbit (LEO) satellite networks are a key enabler for the 6G Non-Terrestrial Network (NTN) architecture. However, supporting time-sensitive services in LEO networks is challenging due to highly dynamic topologies and the difficulty of maintaining precise global time synchronization. Existing Time-Sensitive Networking (TSN) mechanisms largely rely on static topologies and strict synchronization, which makes them ill-suited to dynamic LEO environments. To address this issue, we propose CRT, a deterministic transmission framework tailored for LEO networks. CRT regulates per-hop residence time using local clocks, thereby compensating for link-delay variations without requiring strict global synchronization. To handle asynchronous collisions, CRT adopts a collision-tolerant scheduling strategy that maximizes the number of schedulable flows while bounding collision-induced jitter. We formalize the corresponding scheduling problem and show that it is NP-hard. We further develop CRT-Fast, an efficient heuristic algorithm. It combines iterative layering with path continuity to control collision intensity and improve path stability under topology changes. Simulations on Iridium and Starlink constellations show that the proposed method achieves lower delay jitter and high schedulability under heavy traffic loads.

Chongkai Gao, Zhengrong Xue, Shuying Deng et al.

We present RiEMann, an end-to-end near Real-time SE(3)-Equivariant Robot Manipulation imitation learning framework from scene point cloud input. Compared to previous methods that rely on descriptor field matching, RiEMann directly predicts the target poses of objects for manipulation without any object segmentation. RiEMann learns a manipulation task from scratch with 5 to 10 demonstrations, generalizes to unseen SE(3) transformations and instances of target objects, resists visual interference of distracting objects, and follows the near real-time pose change of the target object. The scalable action space of RiEMann facilitates the addition of custom equivariant actions such as the direction of turning the faucet, which makes articulated object manipulation possible for RiEMann. In simulation and real-world 6-DOF robot manipulation experiments, we test RiEMann on 5 categories of manipulation tasks with a total of 25 variants and show that RiEMann outperforms baselines in both task success rates and SE(3) geodesic distance errors on predicted poses (reduced by 68.6%), and achieves a 5.4 frames per second (FPS) network inference speed. Code and video results are available at https://riemann-web.github.io/.

Ziteng Wang, Yujie He, Guanliang Li et al.

Recent advancements in Multimodal Large Language Models (MLLMs) have demonstrated impressive performance on standard visual reasoning benchmarks. However, there is growing concern that these models rely excessively on linguistic shortcuts rather than genuine visual grounding, a phenomenon we term Text Bias. In this paper, we investigate the fundamental tension between visual perception and linguistic priors. We decouple the sources of this bias into two dimensions: Internal Corpus Bias, stemming from statistical correlations in pretraining, and External Instruction Bias, arising from the alignment-induced tendency toward sycophancy. To quantify this effect, we introduce V-FAT (Visual Fidelity Against Text-bias), a diagnostic benchmark comprising 4,026 VQA instances across six semantic domains. V-FAT employs a Three-Level Evaluation Framework that systematically increases the conflict between visual evidence and textual information: (L1) internal bias from atypical images, (L2) external bias from misleading instructions, and (L3) synergistic bias where both coincide. We introduce the Visual Robustness Score (VRS), a metric designed to penalize "lucky" linguistic guesses and reward true visual fidelity. Our evaluation of 12 frontier MLLMs reveals that while models excel in existing benchmarks, they experience significant visual collapse under high linguistic dominance.

Yifei Xia, Shuchen Weng, Siqi Yang et al.

Panoramic video generation enables immersive 360° content creation, valuable in applications that demand scene-consistent world exploration. However, existing panoramic video generation models struggle to leverage pre-trained generative priors from conventional text-to-video models for high-quality and diverse panoramic videos generation, due to limited dataset scale and the gap in spatial feature representations. In this paper, we introduce PanoWan to effectively lift pre-trained text-to-video models to the panoramic domain, equipped with minimal modules. PanoWan employs latitude-aware sampling to avoid latitudinal distortion, while its rotated semantic denoising and padded pixel-wise decoding ensure seamless transitions at longitude boundaries. To provide sufficient panoramic videos for learning these lifted representations, we contribute PanoVid, a high-quality panoramic video dataset with captions and diverse scenarios. Consequently, PanoWan achieves state-of-the-art performance in panoramic video generation and demonstrates robustness for zero-shot downstream tasks. Our project page is available at https://panowan.variantconst.com.

Limeng Qiao, Yiyang Gan, Bairui Wang et al.

Conventional Vision Transformer simplifies visual modeling by standardizing input resolutions, often disregarding the variability of natural visual data and compromising spatial-contextual fidelity. While preliminary explorations have superficially investigated native resolution modeling, existing approaches still lack systematic analysis from a visual representation perspective. To bridge this gap, we introduce UniViTAR, a family of homogeneous vision foundation models tailored for unified visual modality and native resolution scenario in the era of multimodal. Our framework first conducts architectural upgrades to the vanilla paradigm by integrating multiple advanced components. Building upon these improvements, a progressive training paradigm is introduced, which strategically combines two core mechanisms: (1) resolution curriculum learning, transitioning from fixed-resolution pretraining to native resolution tuning, thereby leveraging ViT's inherent adaptability to variable-length sequences, and (2) visual modality adaptation via inter-batch image-video switching, which balances computational efficiency with enhanced temporal reasoning. In parallel, a hybrid training framework further synergizes sigmoid-based contrastive loss with feature distillation from a frozen teacher model, thereby accelerating early-stage convergence. Finally, trained exclusively on public datasets, externsive experiments across multiple model scales from 0.3B to 1B demonstrate its effectiveness.

Yufeng Zhong, Lei Chen, Zhixiong Zeng et al.

Reading text from images or scanned documents via OCR models has been a longstanding focus of researchers. Intuitively, text reading is perceived as a straightforward perceptual task, and existing work primarily focuses on constructing enriched data engineering to enhance SFT capabilities. In this work, we observe that even advanced OCR models exhibit significantly higher entropy in formatted text (\emph{e.g.}, formula, table, etc.) compared to plain text, often by an order of magnitude. These statistical patterns reveal that advanced OCR models struggle with high output uncertainty when dealing with format sensitive document, suggesting that reasoning over diverse reading pathways may improve OCR performance. To address this, we propose format decoupled reinforcement learning (FD-RL), which leverages high-entropy patterns for targeted optimization. Our approach employs entropy-based data filtration strategy to identify format-intensive instances, and adopt format decoupled rewards tailored to different format types, enabling format-level validation rather than token-level memorization. FD-RL achieves an average score of 90.41 on OmniDocBench, setting a new record for end-to-end models on this highly popular benchmark. More importantly, we conduct comprehensive ablation studies over data, training, filtering, and rewarding strategies, thoroughly validating their effectiveness.

Haojie Zheng, Shuchen Weng, Jingqi Liu et al.

Recent advancements in video generation highlight that realistic audio-visual synchronization is crucial for engaging content creation. However, existing video editing methods largely overlook audio-visual synchronization and lack the fine-grained spatial and temporal controllability required for precise instance-level edits. In this paper, we propose AVI-Edit, a framework for audio-sync video instance editing. We propose a granularity-aware mask refiner that iteratively refines coarse user-provided masks into precise instance-level regions. We further design a self-feedback audio agent to curate high-quality audio guidance, providing fine-grained temporal control. To facilitate this task, we additionally construct a large-scale dataset with instance-centric correspondence and comprehensive annotations. Extensive experiments demonstrate that AVI-Edit outperforms state-of-the-art methods in visual quality, condition following, and audio-visual synchronization. Project page: https://hjzheng.net/projects/AVI-Edit/.

Siqi Yang, Kai Yan, Alexander G. Schwing et al.

Imitation Learning (IL) enables agents to mimic expert behavior by learning from demonstrations. However, traditional IL methods require large amounts of medium-to-high-quality demonstrations as well as actions of expert demonstrations, both of which are often unavailable. To reduce this need, we propose Latent Wasserstein Adversarial Imitation Learning (LWAIL), a novel adversarial imitation learning framework that focuses on state-only distribution matching. It benefits from the Wasserstein distance computed in a dynamics-aware latent space. This dynamics-aware latent space differs from prior work and is obtained via a pre-training stage, where we train the Intention Conditioned Value Function (ICVF) to capture a dynamics-aware structure of the state space using a small set of randomly generated state-only data. We show that this enhances the policy's understanding of state transitions, enabling the learning process to use only one or a few state-only expert episodes to achieve expert-level performance. Through experiments on multiple MuJoCo environments, we demonstrate that our method outperforms prior Wasserstein-based IL methods and prior adversarial IL methods, achieving better results across various tasks.

Chi Zhang, Haibo Qiu, Qiming Zhang et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capabilities of Large Language Models (LLMs) and is now being applied to Vision-Language Models (VLMs). However, vanilla RLVR for VLMs verifies only the final textual output, critically neglecting the foundational step of visual perception. This oversight leads to visual hallucinations and reward hacking, as reasoning built upon flawed perception is inherently unreliable. To address this, we propose PEARL (Perceptual-Evidence Anchored Reinforced Learning), a dual-branch, perception-reasoning synergistic that strengthens multimodal reasoning by explicitly anchoring it to verified visual evidence. For each reasoning-oriented QA instance, PEARL first derive a perception checklist -- a set of perception-oriented sub-questions with verifiable answers that probe the model's understanding of key visual evidence. During training, auxiliary rollouts on this checklist yield a perceptual reward that both directly reinforces the model's perception ability and acts as a fidelity gate for reasoning. If the model passes the perception check, its policy update is biased towards evidence-anchored reasoning. Otherwise, the process is halted to prevent reasoning from flawed premises. PEARL can be seamlessly integrated with popular RL methods like GRPO and DAPO. Comprehensive experiments show PEARL achieves substantial gains on multimodal reasoning benchmarks, e.g., a +9.7% improvement over the baseline and +6.6% over GRPO on MathVerse.

Kun Chen, Peng Shi, Haibo Qiu et al.

Reinforcement learning (RL) with verifiable rewards has recently catalyzed a wave of "MLLM-r1" approaches that bring RL to vision language models. Most representative paradigms begin with a cold start, typically employing supervised fine-tuning (SFT), to initialize the policy before RL. However, SFT-based cold start adopts the reasoning paradigm intertwined with task solution and output format, which may induce instruction-style overfitting, weakens out-of-distribution generalization, and ultimately affects downstream RL. We revisit the cold start along two views, its training method and data construction, and introduce the Generalization Factor (GF) coefficient to quantify the generalization capability under different methods. Our empirical study finds that preference-based training methods (e.g. DPO) generalizes better than SFT-based methods in cold start. Motivated by this, we propose SPECS-a Self-distilled, Preference-based Cold Start framework that decouples multimodal learning: (1) generates introspective preference data pairs via self-distillation, avoiding reliance on larger teachers or manual annotation; (2) performs preference-based training to learn, focusing on shallow, transferable surface-form criteria (format, structure, style) rather than memorizing content; and (3) hands off to RL with verifiable rewards for deep reasoning results. Experimental results across multiple multimodal benchmarks show that our decoupling learning framework yields consistent performance gains over strong baselines, improving MEGA-Bench by 4.1% and MathVista by 12.2%. Additional experiments indicate that SPECS contributes to reducing in-distribution "stuckness," improving exploration, stabilizing training, and raising the performance ceiling.

Xiaohan Lan, Fanfan Liu, Haibo Qiu et al.

Inspired by recent advancements in LLM reasoning, the field of multimodal reasoning has seen remarkable progress, achieving significant performance gains on intricate tasks such as mathematical problem-solving. Despite this progress, current multimodal large reasoning models exhibit two key limitations. They tend to employ computationally expensive reasoning even for simple queries, leading to inefficiency. Furthermore, this focus on specialized reasoning often impairs their broader, more general understanding capabilities. In this paper, we propose Metis-HOME: a Hybrid Optimized Mixture-of-Experts framework designed to address this trade-off. Metis-HOME enables a ''Hybrid Thinking'' paradigm by structuring the original dense model into two distinct expert branches: a thinking branch tailored for complex, multi-step reasoning, and a non-thinking branch optimized for rapid, direct inference on tasks like general VQA and OCR. A lightweight, trainable router dynamically allocates queries to the most suitable expert. We instantiate Metis-HOME by adapting the Qwen2.5-VL-7B into an MoE architecture. Comprehensive evaluations reveal that our approach not only substantially enhances complex reasoning abilities but also improves the model's general capabilities, reversing the degradation trend observed in other reasoning-specialized models. Our work establishes a new paradigm for building powerful and versatile MLLMs, effectively resolving the prevalent reasoning-vs-generalization dilemma.

Xiaoxiao Ma, Haibo Qiu, Guohui Zhang et al.

Reinforcement learning has recently been explored to improve text-to-image generation, yet applying existing GRPO algorithms to autoregressive (AR) image models remains challenging. The instability of the training process easily disrupts the pretrained model capability during long runs, resulting in marginal gains, degraded image quality, and poor generalization. In this work, we revisit GRPO for AR image generation and identify two key issues: contradictory gradients from unnecessary tokens and unstable policy entropy dynamics. To address these, we introduce STAGE, a stable and generalizable framework that leverages two targeted solutions: 1) Advantage/KL reweighting. Similarity-aware reweighting to alleviate conflicting updates; and 2) Entropy reward. An entropy-based reward corresponding to reference model to stabilize learning. With the help of alleviating conflicts between tokens and an entropy reward for stabilizing training, we reduce disruption of the pretrained distribution and mitigate reward hacking, which in turn improves generalization and transfer better to other benchmarks. Experiments across multiple benchmarks show that STAGE consistently improves visual quality, stability, and cross-task generalization compared to baseline GRPO.

Ziteng Wang, Siqi Yang, Limeng Qiao et al.

Despite the success of Vision-Language Models (VLMs) like CLIP in aligning vision and language, their proficiency in detailed, fine-grained visual comprehension remains a key challenge. We present CLIP-IN, a novel framework that bolsters CLIP's fine-grained perception through two core innovations. Firstly, we leverage instruction-editing datasets, originally designed for image manipulation, as a unique source of hard negative image-text pairs. Coupled with a symmetric hard negative contrastive loss, this enables the model to effectively distinguish subtle visual-semantic differences. Secondly, CLIP-IN incorporates long descriptive captions, utilizing rotary positional encodings to capture rich semantic context often missed by standard CLIP. Our experiments demonstrate that CLIP-IN achieves substantial gains on the MMVP benchmark and various fine-grained visual recognition tasks, without compromising robust zero-shot performance on broader classification and retrieval tasks. Critically, integrating CLIP-IN's visual representations into Multimodal Large Language Models significantly reduces visual hallucinations and enhances reasoning abilities. This work underscores the considerable potential of synergizing targeted, instruction-based contrastive learning with comprehensive descriptive information to elevate the fine-grained understanding of VLMs.

Yufeng Zhong, Zhixiong Zeng, Lei Chen et al.

Optical Character Recognition (OCR) for mathematical formula is essential for the intelligent analysis of scientific literature. However, both task-specific and general vision-language models often struggle to handle the structural diversity, complexity, and real-world variability inherent in mathematical content. In this work, we present DocTron-Formula, a unified framework built upon general vision-language models, thereby eliminating the need for specialized architectures. Furthermore, we introduce CSFormula, a large-scale and challenging dataset that encompasses multidisciplinary and structurally complex formulas at the line, paragraph, and page levels. Through straightforward supervised fine-tuning, our approach achieves state-of-the-art performance across a variety of styles, scientific domains, and complex layouts. Experimental results demonstrate that our method not only surpasses specialized models in terms of accuracy and robustness, but also establishes a new paradigm for the automated understanding of complex scientific documents.

Qi Wen, Xiang Wen, Hao Jiang et al.

With the rise of the ``metaverse'' and the rapid development of games, it has become more and more critical to reconstruct characters in the virtual world faithfully. The immersive experience is one of the most central themes of the ``metaverse'', while the reducibility of the avatar is the crucial point. Meanwhile, the game is the carrier of the metaverse, in which players can freely edit the facial appearance of the game character. In this paper, we propose a simple but powerful cross-domain framework that can reconstruct fine-grained 3D game characters from single-view images in an end-to-end manner. Different from the previous methods, which do not resolve the cross-domain gap, we propose an effective regressor that can greatly reduce the discrepancy between the real-world domain and the game domain. To figure out the drawbacks of no ground truth, our unsupervised framework has accomplished the knowledge transfer of the target domain. Additionally, an innovative contrastive loss is proposed to solve the instance-wise disparity, which keeps the person-specific details of the reconstructed character. In contrast, an auxiliary 3D identity-aware extractor is activated to make the results of our model more impeccable. Then a large set of physically meaningful facial parameters is generated robustly and exquisitely. Experiments demonstrate that our method yields state-of-the-art performance in 3D game character reconstruction.

Gordon Dai, Weijia Zhang, Jinhan Li et al.

The emergence of Large Language Models (LLMs) and advancements in Artificial Intelligence (AI) offer an opportunity for computational social science research at scale. Building upon prior explorations of LLM agent design, our work introduces a simulated agent society where complex social relationships dynamically form and evolve over time. Agents are imbued with psychological drives and placed in a sandbox survival environment. We conduct an evaluation of the agent society through the lens of Thomas Hobbes's seminal Social Contract Theory (SCT). We analyze whether, as the theory postulates, agents seek to escape a brutish "state of nature" by surrendering rights to an absolute sovereign in exchange for order and security. Our experiments unveil an alignment: Initially, agents engage in unrestrained conflict, mirroring Hobbes's depiction of the state of nature. However, as the simulation progresses, social contracts emerge, leading to the authorization of an absolute sovereign and the establishment of a peaceful commonwealth founded on mutual cooperation. This congruence between our LLM agent society's evolutionary trajectory and Hobbes's theoretical account indicates LLMs' capability to model intricate social dynamics and potentially replicate forces that shape human societies. By enabling such insights into group behavior and emergent societal phenomena, LLM-driven multi-agent simulations, while unable to simulate all the nuances of human behavior, may hold potential for advancing our understanding of social structures, group dynamics, and complex human systems.

Siqi Yang, Lin Wu, Arnold Wiliem et al.

We present a novel approach to perform the unsupervised domain adaptation for object detection through forward-backward cyclic (FBC) training. Recent adversarial training based domain adaptation methods have shown their effectiveness on minimizing domain discrepancy via marginal feature distributions alignment. However, aligning the marginal feature distributions does not guarantee the alignment of class conditional distributions. This limitation is more evident when adapting object detectors as the domain discrepancy is larger compared to the image classification task, e.g. various number of objects exist in one image and the majority of content in an image is the background. This motivates us to learn domain invariance for category level semantics via gradient alignment. Intuitively, if the gradients of two domains point in similar directions, then the learning of one domain can improve that of another domain. To achieve gradient alignment, we propose Forward-Backward Cyclic Adaptation, which iteratively computes adaptation from source to target via backward hopping and from target to source via forward passing. In addition, we align low-level features for adapting holistic color/texture via adversarial training. However, the detector performs well on both domains is not ideal for target domain. As such, in each cycle, domain diversity is enforced by maximum entropy regularization on the source domain to penalize confident source-specific learning and minimum entropy regularization on target domain to intrigue target-specific learning. Theoretical analysis of the training process is provided, and extensive experiments on challenging cross-domain object detection datasets have shown the superiority of our approach over the state-of-the-art.

Siqi Yang, Arnold Wiliem, Shaokang Chen et al.

This work shows that it is possible to fool/attack recent state-of-the-art face detectors which are based on the single-stage networks. Successfully attacking face detectors could be a serious malware vulnerability when deploying a smart surveillance system utilizing face detectors. We show that existing adversarial perturbation methods are not effective to perform such an attack, especially when there are multiple faces in the input image. This is because the adversarial perturbation specifically generated for one face may disrupt the adversarial perturbation for another face. In this paper, we call this problem the Instance Perturbation Interference (IPI) problem. This IPI problem is addressed by studying the relationship between the deep neural network receptive field and the adversarial perturbation. As such, we propose the Localized Instance Perturbation (LIP) that uses adversarial perturbation constrained to the Effective Receptive Field (ERF) of a target to perform the attack. Experiment results show the LIP method massively outperforms existing adversarial perturbation generation methods -- often by a factor of 2 to 10.

Teng Zhang, Arnold Wiliem, Siqi Yang et al.

This work tackles the face recognition task on images captured using thermal camera sensors which can operate in the non-light environment. While it can greatly increase the scope and benefits of the current security surveillance systems, performing such a task using thermal images is a challenging problem compared to face recognition task in the Visible Light Domain (VLD). This is partly due to the much smaller amount number of thermal imagery data collected compared to the VLD data. Unfortunately, direct application of the existing very strong face recognition models trained using VLD data into the thermal imagery data will not produce a satisfactory performance. This is due to the existence of the domain gap between the thermal and VLD images. To this end, we propose a Thermal-to-Visible Generative Adversarial Network (TV-GAN) that is able to transform thermal face images into their corresponding VLD images whilst maintaining identity information which is sufficient enough for the existing VLD face recognition models to perform recognition. Some examples are presented in Figure 1. Unlike the previous methods, our proposed TV-GAN uses an explicit closed-set face recognition loss to regularize the discriminator network training. This information will then be conveyed into the generator network in the forms of gradient loss. In the experiment, we show that by using this additional explicit regularization for the discriminator network, the TV-GAN is able to preserve more identity information when translating a thermal image of a person which is not seen before by the TV-GAN.