Explore · 每周精选
发现最受关注的研究论文,追踪研究趋势,订阅感兴趣的期刊与关键词。
686 篇论文 · 按点赞排序
Zihan Qiu, Zekun Wang, Xiao Li +33 authors
Qwen3.8-Flash-Next is a sparse mixture-of-experts architecture combining hybrid gated delta-net and sparse attention layers, gated residual branches, and off-accelerator n-gram embeddings to improve efficiency, capability, and training stability.
Ling Xu, Chuyu Han, Borui Li +6 authors
Embodied.cpp is a portable C++ runtime that enables efficient deployment of vision-language-action and world-action models across heterogeneous edge devices through modular execution layers and optimized inference.
Hongyu Qu, Jianzhe Gao, Xiaobin Hu +6 authors
LaMem-VLA introduces a latent-memory-native framework that integrates historical experience into vision-language-action reasoning through coordinated memory components operating in the same latent space.
Yiming Zhang, Zhonghan Zhao, Wenwei Zhang +13 authors
The rapid progress of large foundation models has been driven predominantly by pretraining on large-scale text corpora. However, many forms of knowledge are conveyed through visual representations, where figures, typeset equations, and page layouts carry rich information that cannot be faithfully or completely captured by text alone. Yet current pretraining approaches discard these visual cues by converting visually rich sources, such as documents and web pages, into plain text for learning language intelligence. This paper challenges the default assumption that language models must be trained on text-only representations and shows that Visual Pretraining is a scalable learner for foundation model intelligence. To this end, we conduct a systematic study of unsupervised visual pretraining paradigms that directly leverage visual documents without text extraction. Across multiple backbones and benchmarks, visual pretraining on the same underlying corpora consistently outperforms text-only pretraining, offering an efficient pathway to scalable language intelligence.
Hao He, Chengcheng Hu, Zirun Su +17 authors
DriveZero is an end-to-end autonomous driving system that combines a vision foundation model for perception with a closed-loop reinforcement learning action model to learn driving behaviors beyond human demonstrations.
Kejian Zhu, Zhuoran Jin, Shangqing Tu +5 authors
Reinforcement learning enables stable multi-task reasoning in large language models through sparse, near-orthogonal parameter updates, unlike supervised fine-tuning which suffers from task conflicts, motivating a decoupled Parallel-RL training approach.
Shiyin Lu, Yinglun Li, Yu Xia +10 authors
We introduce OvisOCR2, a 0.8B document parsing model. OvisOCR2 is designed as an end-to-end parser: given a document page image, it generates a Markdown representation in natural reading order, covering text, formulas, tables, and visual regions. We build a data engine that combines filtered real-document annotations with synthetic pages whose rendered images and Markdown targets are derived from the same HTML source. The training recipe includes supervised fine-tuning, reinforcement learning on a 4B branch with a multi-component reward design, on-policy distillation into the 0.8B model, and model fusion. On OmniDocBench v1.6, OvisOCR2 achieves a state-of-the-art overall score of 96.58, placing an end-to-end model at the top of this leaderboard previously dominated by pipeline methods and highlighting the potential of end-to-end document parsing. On PureDocBench, OvisOCR2 also achieves the highest Avg3 score of 75.06. Beyond these two public benchmarks, we evaluate OvisOCR2 on an in-house benchmark designed to cover a broader set of long-tail and challenging scenarios. OvisOCR2 obtains the best overall performance among the compared methods, providing further evidence of its generalization and robustness. OvisOCR2 is available at https://huggingface.co/ATH-MaaS/OvisOCR2.
Geo Ahn, Inwoong Lee, Taeoh Kim +3 authors
RCORE addresses object-driven shortcuts in zero-shot compositional action recognition by using co-occurrence prior regularization and temporal order regularization to improve compositional generalization.
Mijin Yoo, In Cho, Subin Jeon +3 authors
A feed-forward framework decomposes 3D scenes into instance-structured token groups from multi-view images, enabling direct object-level reconstruction, segmentation, and manipulation without 3D annotations.
Igor Pavlovic, Thiemo Wandel, Anton Obukhov +6 authors
Marigold V2 repurposes diffusion transformers for monocular depth estimation via single-step flow-matching inference, semantic alignment, and a Sinkhorn-based two-stage fine-tuning protocol, yielding sharper out-of-distribution depth maps and strong results on related dense regression tasks.
Minhyuk Hwang, Sangmin Kim, Seunguk Do +2 authors
Existing volumetric capture of dynamic human performance achieves high fidelity with dense camera arrays. However, in real-world scenarios, only a handful of low-overlap cameras are available, which degrades the output quality and leaves large areas unobserved. Recent 4D reconstruction methods have focused on low-overlap settings, yet they still produce noticeable artifacts in under-observed regions. Video diffusion models have emerged as another option, but they show geometrically inconsistent results for humans. To address these limitations, we propose StudioRecon, a pipeline that reconstructs 4D human scenes from sparse, low-overlap cameras by decoupling background and humans. We densify background supervision by synthesizing hundreds of camera-controlled novel views with a video diffusion model. We also robustly initialize deformable Gaussian humans with cross-view identity association and triangulated multi-view keypoint fitting. Finally, our recursive enhancement module with motion-adaptive consistency injection harmonizes the composed output, thereby further avoiding remaining artifacts. We achieve state-of-the-art novel view synthesis across four real-world datasets and demonstrate applications such as novel trajectory rendering and human replacement.
Jiale Kang, Ziyin Yue, Zheng Zhan +2 authors
Normalized Low-Rank Adaptation stabilizes LoRA training by normalizing down-projection matrices, accelerating convergence and improving performance without extra parameters or inference cost.
Matteo Farina, Vishaal Udandarao, Thao Nguyen +33 authors
DataComp for VLMs (DCVLM) establishes a comprehensive benchmark for evaluating data curation strategies in vision-language models, demonstrating that data mixing rather than filtering significantly improves model performance at scale.
RadixArk, Tom Chen, Mao Cheng +10 authors
Miles is an open-source, production-ready system for large-scale reinforcement learning and post-training that supports diverse backends, weight synchronization, LoRA, distillation, and diffusion models.
AgiBot Research Team, Renhang Liu, Wenzhi Zhao +42 authors
GE-Act 2.0 is a world-action model trained from scratch with a control-oriented autoencoder, single-step visual planner, and inverse dynamics model, using knowledge-aligned selective optimization to enable scalable zero-shot robot manipulation across diverse skills and conditions.
Yansong Ning, Jingwen Ye, Zhongkai Wu +5 authors
A unified framework benchmarks and trains multimodal agents that infer intent, plan 3D scenes, invoke tools, and reflect on feedback, revealing that reinforcement learning improves open-source models beyond closed-source frontiers.
Disen Lan, Jianbin Zheng, Yuxi Ren +5 authors
FlashMorph is an efficient layer selection method that formulates hybrid layer selection as a budget-constrained optimization problem, using morphable models and linearization regularization to improve long-context efficiency in Transformers.
Zhilin Wang, Han Song, Runzhe Zhan +13 authors
Autonomous agents evaluate policy improvement through iterative editing within fixed budgets, revealing that successful policy evolution requires both task-specific mechanisms and feedback-constrained refinement.
Xionghao Wu, Yijun Yang, Shiyang Zhou +17 authors
ZimaBlue learns generalizable world action models from large-scale egocentric video via a three-stage curriculum and a slow-fast architecture, substantially improving zero-shot robotic manipulation.
Fangxu Yu, Zinan Lin, Xiaodong Liu +4 authors
On-policy distillation (OPD), which aligns a student with the teacher's token-level distribution on the student's own rollouts, is an effective paradigm for transferring capabilities across LLMs. Prevailing approaches assume a teacher at least as capable as the student: they either distill a larger model into a smaller one, which fails at the frontier where no larger teacher exists, or consolidate multiple domain experts trained from a shared base, which requires costly training at the student's scale. We introduce Weak-to-Strong On-Policy Distillation (W2S-OPD), a simple yet effective OPD framework that improves the strong student by distilling from multiple weak models. W2S-OPD constructs a proxy teacher in logit space from a contrast pair of a positive and a negative model, both smaller than the student and cheap to obtain. Their logit difference isolates the capability direction, which is added to the student's own base model, yielding a proxy teacher that couples this direction while staying distributionally adjacent to the student. The student then distills it by minimizing the per-token reverse KL on its own rollouts. We instantiate the contrast pair as i) a post-RL expert against its pre-RL initialization, isolating the skill RL instills, ii) a larger against a smaller base model, isolating the capability from scale, and iii) a small base model with correct versus wrong hints, isolating the instance-level direction toward the solution. Across four math and three code benchmarks, W2S-OPD outperforms OPD, enables the student to surpass the domain teacher, and keeps improving the student even when every supervision source is weaker. Analysis shows different contrasts yield distinct signals: the post-RL and hint contrasts emphasize reasoning frameworks, while the scale contrast emphasizes the solving procedure. Our code will be available at https://github.com/Yu-Fangxu/W2S-OPD.
Tingyu Song, Mingxin Li, Yanzhao Zhang +5 authors
UEmbed is a decoder-only multimodal model that jointly produces dense and sparse embeddings in a single forward pass, extending sparse retrieval to unified text and multimodal inputs.
Zhen Fang, Yu Zeng, Wenxuan Huang +17 authors
Video-DeepResearch extends multimodal agents to continuous video via decoupled perception-exploration and GRPO training, achieving state-of-the-art results on a new video reasoning benchmark.
Danze Chen, Zeqing Wang, Ziyue Lin +2 authors
We present H3-World, an efficient framework that turns the 33B MiniMax-H3 video generator into an interactive world model. Our key finding is that, as large video generators become more capable, language is emerging as a natural interface for control. MiniMax-H3, for example, already supports zero-shot control of character behavior and camera motion through natural-language instructions. Building on this, H3-World turns this coarse language interface into precise, temporally grounded world control, without introducing dedicated action modules. Specifically, we represent each action as a structured combination of character and camera instructions, and align them with the corresponding temporal video latents. To make the control temporally precise, we further introduce temporal attention routing, which restricts each instruction to its intended time interval and reduces control leakage across actions. Importantly, H3-World directly reuses the semantic representations learned during large-scale video pretraining and requires only lightweight adaptation. With only 8,000 gameplay samples, 10,000 LoRA optimization steps, and 0.199% trainable parameters, H3-World achieves effective character and camera control while preserving strong generation quality. It also generalizes to unseen scenarios. These results show that the control capabilities emerging in large video generators can be efficiently transformed into interactive world control.
Taeil Kim, Kangsan Kim, Sung Ju Hwang
Agent Memory Distillation improves small language model tool-use performance by transferring structured hierarchical memory from a large teacher agent without additional training.
Yaxin Luo, Haobin Jiang, Jialv Zou +11 authors
AutoDesign uses a meta-harness optimizer to recursively improve a code agent for structured media generation, achieving state-of-the-art results on paper-to-poster synthesis.
Matan Avitan, Yoav Goldberg, Yanai Elazar
Manifold constraint hypothesis enables improved concept erasure by projecting updates onto estimated representation manifolds, achieving state-of-the-art results in nonlinear concept removal.
Ming Li, Chenguang Wang, Xirui Li +5 authors
Rhetorical framing significantly biases AI scientific review scores in structured ways, with effects shaped by reviewer identity, score range, and evaluation strictness rather than rewriting complexity.
NeoteAI Team, Fudan TEAI Team
N0-TWAM is a large-scale tactile-native world-action model that predicts future vision, contact, and actions for contact-rich manipulation using a unified force-based tactile representation, contact events for task staging, and an asymmetric mixture-of-transformers architecture.
Zixuan Wang, Yuhong Chen, Yuxuan Zhu +10 authors
Knowledge-Geometry Decoupling separates behavioral pretraining from task-specific geometry via multi-token prediction and orthogonal residual adaptation, improving recommender transfer under continual refresh.
Qi Li, Xingyi Yang, Xinchao Wang
World-action models (WAMs) are emerging as a promising foundation for embodied control: rather than predicting actions alone, they learn representations that couple action generation with future world prediction. This coupling is often viewed as a source of robustness, interpretability, and safety, as a robot's action can in principle be checked against its imagined future. In this paper, we show that this assumption is fragile. We introduce BadWAM, a unified framework for modeling and evaluating World-Action Drift Attacks: a new class of WAM-specific adversarial attacks that use small visual perturbations to break the alignment between what a WAM imagines and what it executes. BadWAM characterizes this attack surface along two natural criteria: attack strength and stealthiness. When the adversary prioritizes disruption, BadWAM instantiates an action-only adversarial attack, which directly drives the model toward task-failing actions. When the adversary additionally prioritizes stealth, BadWAM instantiates an imagination-preserving adversarial attack, which seeks to induce harmful action shifts while keeping the model's predicted future close to its clean imagination. Together, these two attacks capture a spectrum of WAM-specific failures: from overt action hijacking to stealthier cases where the model appears to imagine a plausible future but executes a desynchronized action. We evaluate BadWAM across different variants of WAMs. Results show that our attacks substantially reduce task success rates under closed-loop execution. For example, our action-only attack reduces the model performance from 96.5% to 43.1% success. The results of our imagination-preserving attack further exposes a WAM-specific vulnerability: moderate future-preserving regularization can maintain strong attack performance while reducing future imagination drift.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号