Monthly collection
The papers that drew the most attention during this month of AI research.
Moonshot AI
Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts model with 104 billion activated parameters, native vision, and a 1-million-token context window. It uses Kimi Delta Attention, Attention Residuals, and Stable LatentMoE (16 of 896 routed experts active per token), achieving about 2.5× scaling efficiency over Kimi K2. Post-training includes…
Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, et al.Research paper
The paper introduces Program-as-Weights (PAW), a paradigm for implementing fuzzy functions (tasks like log filtering or intent classification) by compiling natural-language specifications into compact neural programs. A 4B compiler, trained on the new FuzzyBench dataset (10M examples), emits a hybrid program: a discrete pseudo-program (paraphrase plus…
Wentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, et al.Research paper
ABot-World-0 is an action-conditioned video world model for real-time, long-horizon closed-loop interaction, deployable on a single NVIDIA RTX 5090 GPU. It uses raw keyboard inputs as a unified control interface for both scene roaming and third-person character control, with reference-character memory for identity consistency. The model is trained on…
Fan Jiang, Zhaoxu Sun, Mengchao Wang, Ziyu Zhu, et al.Research paper
Qwen-UI-Agent is a foundation GUI agent designed for real-world use across mobile, computer, web, and DeepSearch environments. It integrates scalable sandbox environments with a real-device mobile runtime (over 100 devices, 150+ apps), a unified action space interleaving GUI, CLI, and batched actions, an AutoResearch-style data flywheel, and a harness…
Hanzhang Zhou, Panrong Tong, Xu Zhang, Quyu Kong, et al.Research paper
AskChem is a claim-centered infrastructure for cross-paper chemistry search, shifting the retrieval unit from documents to provenance-carrying claims. Each claim is an atomic, typed assertion extracted from a paper, grounded by a source DOI and a verbatim quote or evidence locator. The system indexes 2.4M claims from 147K papers (1925–2026) and exposes…
Bing Yan, Gregory Wolfe, Stefano Martiniani, Kyunghyun ChoResearch paper
The paper introduces memory foundation models, which internalize memory into the backbone of large foundation models, contrasting with external memory modules like RAG. The authors formalize native memory via a persistent, dynamically evolving memory state within the model's parameters and native memory procedures that autonomously store and utilize…
Zeyu Zhang, Ziliang Guo, Yihang Sun, Xichong Zhang, et al.Research paper
The paper introduces Harness Handbook, a behavior-centric representation for agent harnesses that links system behaviors to their distributed source implementations, addressing the challenge of behavior localization during harness evolution. It is constructed automatically via static analysis and LLM-assisted structuring, organizing knowledge into a…
Ruhan Wang, Yucheng Shi, Zongxia Li, Zhongzhi Li, et al.Research paper
LongStraw is a system for long-context GRPO post-training under fixed GPU budgets, addressing the memory bottleneck of multi-million-token prompts by virtualizing resident state, replaying responses, and executing distributed gradients. It captures the shared prompt without autograd, retains only architecture-required state (e.g., recurrent state, KV…
Changhai Zhou, Kieran Liu, Yuhua Zhou, Qian Qiao, et al.Research paper
RynnBrain 1.1 is a family of embodied foundation models (2B, 9B, 122B-A10B) built on Qwen3.5, introducing contact-point prediction and native 3D grounding (for 2B/9B) to better align with robot manipulation. The 122B-A10B model outperforms all evaluated proprietary and open-source models on VSI-Bench, MMSI, and RefSpatial-Bench. Scaling analysis shows…
Kehan Li, Bohan Hou, Minghao Zhu, Tianyi Zhang, et al.Research paper
This survey provides a unified framework for progress reward modeling in robotic learning, addressing the fragmentation in the field. It organizes the domain into three connected steps: the interface of progress models (input/output structure), the methods for constructing progress rewards, and the data and benchmarks used for evaluation. The interface…
Jianshu Zhang, Keliang Wu, Haoran Lu, Anbang Liu, et al.Research paper
The paper introduces OpenMLE, an open full-stack system for studying recursive self-improvement (RSI) in machine learning engineering (MLE), and Frontis-MA1-35B, a meta-evolution agent trained on this stack. OpenMLE comprises OpenMLE-Gym (5,758 quality-gated executable tasks with sandboxed execution), OpenMLE-ERL (execution-grounded SFT and RL training of…
Junlin Yang, Che Jiang, Yu Fu, Tianwei Luo, et al.Research paper
VideoChat3 is a fully open, efficient, and generalist video multimodal large language model (MLLM) with 4B parameters, designed to address limitations in existing open-source models: limited cross-domain generalization, high computational demands, and incomplete openness. It introduces two key architectural innovations: the Inflated 3D Vision Transformer…
Xinhao Li, Yuhan Zhu, Xiangyu Zeng, Yuhao Dong, et al.Research paper
TimeLens2 introduces a generalist video temporal grounding model that predicts variable-cardinality sets of evidence intervals across diverse video lengths, domains, query forms, and viewpoints. It addresses two structural mismatches: unreliable long-video supervision and optimization that lacks interval-level geometry. The TimeLens2-93K dataset pipeline…
Yuhan Zhu, Changlian Ma, Xiangyu Zeng, Xinhao Li, et al.Research paper
PHIZERO is a physical world model that learns a compact discrete 'physical language' from in-the-wild videos via self-supervision, representing state transitions. It adopts a reason-then-render paradigm: a Physical Language Reasoner (initialized from Qwen3-VL-4B) predicts a physical-language sequence from the first frame and a textual action intent, and a…
Shuyao Shang, Yuqi Wang, Ruopeng Gao, Xu Chen, et al.Research paper
HiFi-UMI is a portable data-production system that co-designs hardware and software to achieve high-fidelity robot-free UMI data, aiming to eliminate the need for real-robot teleoperation in post-training. It uses head-mounted offline stereo-inertial SLAM, native inter-gripper relative pose, microsecond GPIO synchronization, and ultra-wide 200°…
Simple AI, :, Yuteng Wei, Jinming Ma, et al.Research paper
AREX is a family of recursively self-improving deep research agents developed by BAAI. It exploits the discovery-verification asymmetry in deep research by alternating between an inner research loop that gathers evidence and constructs provisional answers, and an outer self-improvement loop that audits answers constraint-wise, identifies unresolved claims,…
Shuqi Lu, Chaofan Li, Kun Luo, Zhang Zhang, et al.Research paper
RAGU is an open-source, modular GraphRAG engine that addresses three obstacles in existing systems: single-pass extraction, dependence on expensive LLMs, and engineering immaturity. It separates extraction from consolidation via a six-stage pipeline: chunking, two-stage typed extraction (entities then relations) under the NEREL schema, DBSCAN-backed…
Mikhail Komarov, Ivan Bondarenko, Stanislav Shtuka, Oleg Sedukhin, et al.Research paper
Vidu S1 is a real-time interactive video generation model that enables users to control digital characters via voice instructions during generation, supporting infinite-length video without blurring or drift. Built with TurboDiffusion and TurboServe, it outputs 540p video at up to 42 FPS on consumer GPUs. The model uses a three-stage training pipeline:…
Jintao Zhang, Kai Jiang, Jintao Chen, Xu Wang, et al.Research paper
DataFlow-Harness is a platform that bridges the NL2Pipeline gap, where coding agents produce scripts that are not persistent, editable platform artifacts. It guides an LLM agent to construct platform-native DAGs via typed, incremental mutations, combining DataFlow-Skills for procedural guidance, an MCP layer for live operator registry and pipeline state,…
Runming He, Zhen Hao Wong, Hao Liang, Zimo Meng, et al.Research paper
The paper introduces Direct On-Policy Distillation (Direct-OPD), a method to transfer the policy shift induced by reinforcement learning (RL) on a small, weak teacher model to a stronger student model, avoiding the high cost of running RL directly on the larger model. Instead of imitating the post-RL teacher's final policy, Direct-OPD uses the log-ratio…
Shiyuan Feng, Huan-ang Gao, Haohan Chi, Hanlin Wu, et al.Research paper
Boogu-Image-0.1 is an open-source family of unified multimodal understanding and generation models (Base, Turbo, Edit, Edit-Turbo) that achieves competitive performance in text-to-image generation, fast inference, instruction-based editing, and bilingual text rendering. The authors argue that strengthening understanding—via a stronger multimodal encoder…
Guoxuan Chen, Chufeng Xiao, Haoran Yang, Siyue Xie, et al.Research paper
TurboVLA introduces a new paradigm for vision-language-action (VLA) models, replacing the conventional LLM-centric V→L→A pathway with a direct V+L→A mapping. It uses a lightweight BERT text encoder and a DINOv3 vision encoder, with bidirectional cross-attention to fuse visual and linguistic features, and an ACT-style decoder to predict continuous action…
Hengyi Xie, Chenfei Yao, Xianjin Wu, Xuanyang Xi, et al.Research paper
JarvisHub is an open, canvas-native agent harness designed for long-horizon multimodal creative tasks, addressing the limitations of prompt-to-output tools, chat-based agents, and node-based workflows that fail to maintain a unified project state. It treats an editable canvas as the shared workspace, external memory, and action space, representing…
Yunlong Lin, Zixu Lin, Zhaohu Xing, Biqiang Li, et al.Research paper
CodeNib is a multi-view data system that serves repository context to coding agents by building reusable lexical, dense, and structural views per repository commit, mapping outputs to repository-relative source ranges, and maintaining views across edits. It addresses three challenges: heterogeneous views (C1), incremental freshness (C2), and agent delivery…
Zhongming Yu, Hengjia Yu, Boqin Yuan, Shuting Zhao, et al.Research paper
This paper introduces function-aware fill-in-the-middle (FIM) mid-training to improve coding agent foundation models. The authors observe that a coding agent's action-observation-continuation loop is structurally similar to a function call site, and they exploit this by masking functions selected via program dependency graph analysis and a…
Yubo Wang, Jiarong Liang, Yuxuan Zhang, Xuye Liu, et al.Research paper
SEED (SElf-Evolving On-Policy Distillation) is a framework for agentic reinforcement learning that converts completed on-policy trajectories into natural-language hindsight skills and distills their behavioral effect back into the policy model. It addresses the supervision gap between sparse trajectory-level rewards and token-level policy learning in…
Jinyang Wu, Shuo Yang, Zhengxi Lu, Fan Zhang, et al.Research paper
ABot-N1 is a general visual language navigation foundation model from AMAP CV Lab (Alibaba Group) that decouples cognition from control using a slow-fast architecture. The slow system (4B-parameter VLM) performs Chain-of-Thought reasoning and outputs pixel goals (affordance and target pixels) in image space, serving as a universal interface for five tasks:…
Ruiyan Gong, Yingnan Guo, Junjun Hu, Jintao Kong, et al.Research paper
The paper presents Ring-Zero, a stable and efficient pipeline for scaling zero reinforcement learning (RL) with verifiable rewards to a trillion-parameter model (Ling-2.5-1T-Base, 1T parameters, 63B activated) without human-annotated data. The pipeline includes three RL stages and a self-distillation phase, using clipped importance sampling,…
Xinyu Tang, Qianggang Cao, Yurou Liu, Yuliang Zhan, et al.Research paper
RynnWorld-4D is a 4D embodied world model that generates synchronized RGB, depth, and optical flow (RGB-DF) videos from a single RGB-D image and a language instruction, using a tri-branch diffusion architecture with cross-modal attention and 3D RoPE. The authors curate Rynn4DDataset 1.0, containing over 254.4 million frames from human and robotic…
Haoyu Zhao, Xingyue Zhao, Siteng Huang, Xin Li, et al.Research paper
The paper introduces the Relevance-Aware RipGrep Search Agent (RARG), which uses relevance as an execution prior to guide corpus interaction in agentic search. RARG ranks documents with an embedding retriever and makes ripgrep traverse them in that order, exposing matches from more relevant documents earlier. It has two extensions: RARG+ seeds the agent…
Jiangnan Li, Yuqing Li, Mo Yu, Jinchao Zhang, et al.Research paper
The paper introduces Mental World Modeling (MWM), a framework for world models that jointly represent physical and mental dynamics to predict human decisions. MWM maintains a coupled physical-mental world state, renders a target-specific partial observation, and simulates how candidate actions update both components. The authors implement MENTIS, a…
Hao Fei, Yiran ZhaoResearch paper
EvolvingWorld is a framework and benchmark for simulating interactive literary worlds where characters and the world co-evolve over long horizons. It addresses limitations of existing systems that treat role-play as static persona imitation or isolated scene generation. The framework uses an open-schema design, coupling a Character Agent for…
Qing Zong, Yue Guo, Mengxin Yang, Yiwen Guo, et al.Research paper
The paper DistillAlign revisits autoregressive video distillation from a distributional perspective. It argues that existing multi-stage pipelines, which separate initialization (e.g., ODE or consistency distillation) from DMD refinement, often have misaligned target distributions. Since DMD is mode-seeking, a good initialization must match the mode…
Jiaxing Li, Kai Zou, Cindy Zhou, Kaichen Huang, et al.Research paper
AlayaWorld is a full-stack, open-source framework for building interactive generative worlds, fine-tuned from LTX-2.3. It addresses four key challenges: control, consistency, stability, and runtime. For control, it combines a 3D cache rendered along the camera trajectory with AdaLN-style camera modulation, and supports prompt-driven actions via a…
AlayaWorld Team, Kaipeng Zhang, Chuanhao Li, Yifan Zhan, et al.Research paper
The paper introduces DeepSearch-World, a deterministic and verifiable offline environment for deep search agents, built on a Wikipedia corpus with 420K multi-hop QA tasks. It also presents DeepSearch-Evolve, a self-distillation framework that iteratively generates, filters, and converts scaffolded trajectories into ReAct-format supervision for student…
Xinyu Geng, Xuanhua He, Sixiang Chen, Yanjing Xiao, et al.Research paper
SciReasoner is a multimodal scientific foundation model for native structural reasoning across proteins, small molecules, and inorganic crystals. It discretizes coordinates, topologies, and periodic connectivities into a unified structure-aware vocabulary, treating structural tokens as addressable evidence units within autoregressive reasoning…
Chen Tang, Yizhou Wang, Jianyu Wu, Lintao Wang, et al.Research paper
This paper introduces RLSVR (Reinforcement Learning with Self-Verifiable Rewards), a training paradigm that extends RLVR to open-ended tasks by transforming them into proxy environments with automatically verifiable rewards, inspired by self-supervised learning. The authors instantiate RLSVR with SpyRL, an information-asymmetric self-play framework based…
Qinsi Wang, Jing Shi, Huazheng Wang, Kun Wan, et al.Research paper
Visual generators fail on requests requiring world knowledge beyond their training data, such as new characters or recent events. The authors introduce SEARCHGEN-20K, a dataset of 20,939 prompts with twelve failure categories, and SEARCHGEN-BENCH, where open generators score only 21–28 out of 100, a 40-point drop from standard benchmarks. Naive search…
Haozhe Wang, Weijia Feng, Jinpeng Yu, Che Liu, et al.Research paper
The paper introduces GenCeption, a general-purpose vision model that uses large-scale text-to-video generation as a pre-training paradigm. By repurposing a pre-trained video diffusion backbone (WAN 2.1) into a feed-forward model, GenCeption performs multiple vision tasks—depth, surface normal, camera pose, segmentation, and 3D keypoint estimation—steered…
Letian Wang, Chuhan Zhang, Rishabh Kabra, Jasper Uijlings, et al.Research paper
ABot-AgentOS is a general robotic Agent Operating System that provides a deliberative layer above low-level controllers, enabling scene-conditioned planning, context-isolated skill execution, multi-stage verification, multi-modal memory, and edge-cloud collaboration. It introduces Universal Multi-modal Graph Memory, a persistent, source-grounded substrate…
Jiayi Tian, Shiao Liu, Yuting Xu, Jia Lu, et al.Research paper
The paper introduces Hierarchical Landmark Sparse (HiLS) Attention, a chunk-wise sparse attention mechanism for large language models that learns chunk selection end-to-end under the language-modeling loss. HiLS factorizes attention hierarchically: each query attends independently to retrieved chunks and fuses outputs using retrieval scores, which are…
Xiang Hu, Xinyu Wei, Hao Gu, Minshen Zhang, et al.Research paper
The paper introduces SpectraReward, a training-free reward function for text-to-image reinforcement learning (RL) that converts pretrained multimodal large language models (MLLMs) into reward models. Instead of asking the MLLM to judge images or answer questions, SpectraReward measures how well the original prompt can be recovered from a generated image…
Runhui Huang, Qihui Zhang, Zhe Liu, Yu Gao, et al.Research paper
The paper introduces Multi-Agent Protocol Distillation (MAPD), a framework for distilling knowledge from proprietary LLMs to open-source student models in agentic search. It addresses two bottlenecks: inaccessible logits and tokenizer mismatches that prevent logit-matching, and style drift from imitating raw natural-language trajectories. MAPD uses a…
Junlin Liu, Jiangwang Chen, Zixin Song, Shuaiyu Zhou, et al.Research paper
CoRT (Counterfactual Replay for Token-level credit weighting) is a method for rubric-conditioned GRPO that addresses the uniform token-level credit assignment in standard GRPO. It rescores the same sampled response under a criteria-free prompt, computes tokenwise log-likelihood contrasts, and maps these to bounded, response-normalized weights. These…
Bo-Wen Zhang, Junwei He, Wen Wang, Song-Lin Lv, et al.Research paper
AlayaRenderer-Flash is a real-time generative world renderer that accelerates the offline AlayaRenderer from 0.56 FPS to 31.54 FPS, enabling interactive, prompt-controllable gameplay. It reformulates the original renderer into a few-step autoregressive streaming model with three key improvements: autoregressive generation over unbounded G-buffer streams,…
Guixu Lin, Zheng-Hui Huang, Siqi Yang, Ming-Hsuan Yang, et al.Research paper
KnowAct-GUIClaw is a personal GUI assistant framework built on the 'Know Deeply, Act Perfectly' paradigm to address OpenClaw's limitations in cross-platform GUI interaction and self-evolution. It uses a Know-Route-Act-Reflect loop: the host agent decomposes tasks and allocates them (Know), a router classifies tasks and transfers data via a blackboard…
Yunxin Li, Jinchao Li, Shibo Su, Zhenran Xu, et al.Research paper
RynnWorld-Teleop introduces digital teleoperation, a paradigm that replaces physical robots with a generative world model for scalable robot data collection. An operator's hand-pose stream drives a robot-centric world model to synthesize high-fidelity egocentric videos from a single reference image, with the pose stream serving as an embodiment-agnostic…
Haoyu Zhao, Xingyue Zhao, Hangyu Li, Biao Gong, et al.Research paper
SWE-Pruner Pro is a method for pruning long tool outputs in coding agents by reading line-level importance directly from the agent's own internal representations, eliminating the need for a separate scoring model or explicit goal-hint query. A lightweight head, trained with a per-sample balanced focal loss and a length-aware embedding, converts the frozen…
Yuhang Wang, Yuling Shi, Shaoqiu Zhang, Jialiang Liang, et al.Research paper
MonkeyOCRv2 is a visual-text foundation model for document AI, addressing the mismatch between natural-image encoders and document images. The authors construct MonkeyDoc v2, a 113-million-image pretraining corpus spanning 17 languages, and propose a dual-objective pretraining strategy combining image-to-text generation with pixel-level reconstruction to…
Yuliang Liu, Zhang Li, Ziyang Zhang, Shuo Zhang, et al.Research paper
OmniOpt is a survey and benchmark framework for modern optimizers, addressing the fragmented landscape of over one hundred methods. It introduces a five-stage meta-pipeline (S0-S5) to describe optimizer updates, and a four-axis decomposition (update domain, state estimator, geometry/precondition operator, finalization) based on linear minimization oracles…
Siyuan Li, Jiabao Pan, Yumou Liu, Zhuoli Ouyang, et al.Research paper
Long-Horizon-Terminal-Bench is a new benchmark of 46 containerized terminal tasks across nine categories, designed to test AI agents on long-horizon workflows. Unlike prior benchmarks that use binary pass/fail grading, it decomposes each task into graded subtasks, providing dense partial-credit rewards. Tasks require an average of 239 episodes, 9.8M…
Zongxia Li, Zhongzhi Li, Yucheng Shi, Ruhan Wang, et al.Research paper
Mage-Flow is a compact 4B-parameter generative stack from Microsoft for efficient text-to-image generation and instruction-based image editing. It comprises Mage-VAE, a lightweight latent tokenizer using one-step diffusion-style encoding/decoding with anchor-latent regularization, and a Native-Resolution Multimodal Diffusion Transformer (NR-MMDiT) trained…
Xinjie Zhang, Peng Zhang, Shicheng Zheng, Jinghao Guo, et al.Research paper
This paper analyzes on-policy distillation (OPD) for diffusion models under classifier-free guidance (CFG). The authors show that the naive objective of matching CFG-composed velocities is under-identified at the branch level, allowing positive- and negative-branch errors to cancel. They identify a failure mode, Negative Branch Asymmetry (NBA), which…
Bingnan Li, Haozhe Wang, Haozhong Xiong, Fangtai Wu, et al.Research paper
HumanCLAW is an evaluation framework that measures whether vision-language models (VLMs) can act through a physical body by decoupling action decision-making from low-level motor execution. A harnessed VLM issues atomic skill commands (e.g., walk, turn, sit) at sub-second intervals, which are converted into continuous full-body motion by a…
Siyao Li, Jiawei Gu, Shuai Liu, Kairui Hu, et al.Google DeepMind
Gemma 4 is a new family of open-weight, natively multimodal language models from Google DeepMind, featuring dense (2.3B, 4.5B, 12B, 31B) and Mixture-of-Experts (26B total, 3.8B active) architectures. Key innovations include a thinking mode for reasoning traces, an encoder-free architecture for the 12B model that processes raw audio and image patches, and…
Gemma Team, Sherif El Abd, Vaibhav Aggarwal, Robin Algayres, et al.Research paper
The paper introduces the Loopie series, two Mixture-of-Experts (MoE) language models: Loopie-20B-A2B and Loopie-6B-A0.6B, which use a novel 'layer-loop' recurrence pattern where each layer is applied twice before moving to the next, unlike prior 'model-loop' approaches. The key contribution is the Loopie Recipe, a compute-matched scaling method that halves…
Zitian Gao, Yilong Chen, Yihao Xiao, Xinyu Yang, et al.Research paper
The paper investigates how text-to-image diffusion transformers (DiTs) incorporate text semantics, focusing on chat-template tokens introduced by LLM-based text encoders. Using a causal interpretability framework on Qwen-Image models, the authors find that template tokens, despite carrying little prompt-specific information, become dominant attention sinks…
Maohua Li, Qirui Li, Yanke Zhou, Yiduo Li, et al.Research paper
This technical report presents SLAI T-Rex, a full-stack framework for post-training the DeepSeek-V4 model family on Ascend SuperPOD. System-level optimizations (parallelism, communication, memory, kernels) increased Model FLOPs Utilization (MFU) from 11.67% to 34.22%, a 2.93x improvement. For Operations Research (OR) specialization, a solver-grounded…
Dongfang Li, Xiaodong Luo, Ruoyu Sun, Xuhui Chen, et al.Research paper
The paper introduces UI-MOPD, a method for continual learning of multi-platform GUI agents, addressing challenges like scarce cross-platform data and behavioral pattern mixing. It constructs Uni-GUI, a dataset of ~10K high-quality trajectories from desktop and mobile environments, and proposes multi-teacher on-policy distillation (MOPD) with…
Niu Lian, Alan Chen, Zhehao Yu, Chengzhen Duan, et al.Research paper
SynthDocBench is a fully synthetic benchmark for long-context visual document understanding, designed to systematically control factors like document length, layout, modality, and question type. It comprises 200 reports (avg. 51.1 pages, 16.7 charts) and 1,788 questions across three subsets: chart-reading, complex multi-hop, and cross-modal. Documents are…
Abhigya Verma, Khyati Mahajan, Amit Kumar Saha, Shruthan Radhakrishna, et al.Research paper
SearchOS is a multi-agent framework for robust open-domain information seeking. It formulates information-seeking tasks as relational schema completion with grounded citations, where agents discover entities, populate attributes across linked tables, and anchor each value to source evidence. To manage long-horizon search, it introduces Search-Oriented…
Yuyao Zhang, Junjie Gao, Zhengxian Wu, Jiaming Fan, et al.Research paper
Xiaomi-Robotics-1 is a foundational vision-language-action (VLA) model trained on over 100,000 hours of real-world manipulation trajectories collected via UMI devices. It uses a two-stage training recipe: pre-training on UMI data with auto-labeled state-transition language prompts (generated by a VLM) to learn generalizable action generation, and…
Xiaomi Robotics Team, Jun Guo, Piaopiao Jin, Jason Li, et al.Research paper
VideoCoCo is an agentic dual-engine framework for physically consistent text-to-video generation. It uses executable Blender code as a process-level chain of thought. A coding agent synthesizes a Blender program from a text prompt, which is executed in a sandbox to produce a deterministic, low-fidelity spatiotemporal draft. A generative video engine then…
Haodong Li, Tianfei Ren, Xiaoxiao Ma, Chunmei Qing, et al.Research paper
Open-AoE is an open, community-oriented egocentric manipulation dataset and toolchain from Ant Group, covering the full pipeline from smartphone capture to model training. Its first release includes approximately 2,000 hours of manipulation video collected in natural environments by 500+ contributors using 400+ smartphone models, spanning 400+ scenes and…
Zishuo Li, Bowen Yang, Changtao Miao, Kai Zhu, et al.Research paper
The paper introduces AgenticSTS, a bounded-memory testbed for long-horizon LLM agents, instantiated in the game Slay the Spire 2. Instead of appending raw transcripts, the agent composes each decision prompt from five typed layers (L1-L5): fixed protocol, state schemas, game rules, episodic summaries, and triggered strategic skills. This contract keeps…
Xiangchen Cheng, Yunwei Jiang, Jianwen Sun, Zizhen Li, et al.Research paper
PixWorld is a unified pixel-space diffusion framework for 3D scene generation and reconstruction. It partitions multi-view inputs into clean and noisy subsets, processes them with a two-stream diffusion transformer, and decodes features into a pixel-aligned 3D Gaussian representation. The diffusion objective is applied directly on rendered images,…
Sensen Gao, Zhaoqing Wang, Qihang Cao, Dongdong Yu, et al.Research paper
ReDesign is an agentic framework that recovers editable design structures (e.g., Figma files) from raster images by growing a layer hierarchy through tool composition. A VLM controller selects actions from a fixed set (text extraction, multi-layer decomposition, connected component labeling, detection/segmentation, vectorization) to expand nodes, with…
Jooyeol Yun, Jintae Park, Hyesu Lim, Junha Hyung, et al.Research paper
DecoEvo is a text-space optimization framework that co-evolves a solver skill and a rubric-generator skill under decoupled objectives, avoiding the pitfalls of score-coupled co-adaptation. The solver skill is updated using criterion-level feedback from generated rubrics, while the generator skill is revised through two audits: a task-conditioned structural…
Jiangwang Chen, Zixin Song, Junlin Liu, Shuaiyu Zhou, et al.Research paper
ResearchStudio-Reel is a native-editable dissemination workspace that automates the last mile of research communication, turning a single paper PDF into a print-ready conference poster, a narration-aligned talk video, and a bilingual blog. Implemented as five composable skills in Claude Code and Codex, it uses a shared Paper2Assets extractor to ground all…
Lingao Xiao, Yalun Dai, Yangyu Huang, Qihao Zhao, et al.Research paper
LingBot-Video is a DiT-based video pretraining paradigm for embodied intelligence, introduced as the first large-scale open-source Mixture-of-Experts (MoE) video foundation model. It addresses domain mismatch in video generation by using a sparse MoE framework for better capacity-efficiency trade-off, a data profiling engine that augments internet videos…
Shuailei Ma, Jiaqi Liao, Xinyang Wang, Jingjing Wang, et al.Research paper
ResearchStudio-Idea is a suite of three skills for evidence-grounded research ideation in machine learning. It includes Paper-Search for multi-source literature grounding, Scoop-Check for prior-art collision checking, and IdeaSpark, an end-to-end skill that composes evidence grounding, pattern-guided generation, collision retrieval, audit, and idea-card…
Qihao Zhao, Yangyu Huang, Yalun Dai, Lingao Xiao, et al.Research paper
StateAct is a code-first, multi-agent harness for long-horizon computer-use tasks that grounds the main agent in program state (files, DOM, backends) rather than screenshots, which are lossy and non-injective. The main agent acts via code, while a dedicated GUI subagent handles visual interaction on only 28 of 108 tasks and 1.1% of main-agent steps. An…
Yan Yang, Xiangru Jian, Ziyang Luo, Zirui Zhao, et al.Research paper
HOMIE is a framework for human-object centric video personalization (HOCVP) that unifies inter-subject (distinct subjects) and intra-subject (multiple references of the same subject) personalization. It addresses limitations of existing methods by integrating Multimodal Large Language Models (MLLMs) while preserving the text encoder, avoiding costly…
Yiyang Cai, Nan Chen, Rongchang Xie, Junwen Pan, et al.Research paper
Embodied.cpp is a portable C++ inference runtime for embodied AI models, addressing the fragmented deployment of vision-language-action (VLA) models and world-action models (WAMs) on heterogeneous edge devices. It identifies three key runtime requirements: multi-rate execution, latency-first batch-1 inference, and extensible embodied interfaces. The…
Ling Xu, Chuyu Han, Borui Li, Hao Wu, et al.Research paper
AlayaWorld is an interactive long-horizon video world model that generates 24-fps video at 540p and 720p, built on a 15B video diffusion transformer. It generates short latent chunks autoregressively under camera trajectories and switchable text prompts, using a bounded visual context that combines a persistent sink frame, compressed temporal history,…
AlayaWorld Team, Kaipeng Zhang, Chuanhao Li, Yifan Zhan, et al.Research paper
LaMem-VLA introduces a dual latent memory framework for Vision-Language-Action (VLA) models to address the temporal short-horizon bias in long-horizon robotic manipulation. Unlike existing memory-augmented VLAs that store history externally and consume it as auxiliary policy-side context, LaMem-VLA represents historical experience as context-native latent…
Hongyu Qu, Jianzhe Gao, Xiaobin Hu, Shaohuan Yang, et al.Research paper
This paper introduces Visual Pretraining (VP), a framework that trains foundation models directly on raw document images without text extraction or image-text pairing, using a next-visual-latent prediction objective. VP consistently outperforms text-only pretraining (TP) on scientific reasoning benchmarks across multiple backbones (Qwen3.5, Qwen3, Llama3.2…
Yiming Zhang, Zhonghan Zhao, Wenwei Zhang, Haiteng Zhao, et al.Research paper
The paper introduces xHC (Expanded Hyper-Connections), a method to scale residual-stream expansion in Transformer models beyond the previous limit of N=4 streams. Existing Hyper-Connections (HC) methods, particularly Manifold-Constrained HC (mHC), show diminishing returns and rising costs when scaling N due to two bottlenecks: insufficient write-back…
Xiangdong Zhang, Xiaohan Qin, Sunan Zou, Tuo Dai, et al.Research paper
N0-VTLA is a vision-tactile-language-action (VTLA) foundation model for contact-rich manipulation. It introduces latent tactile tokens that predict future tactile changes over the action horizon, conditioning a flow-matching action expert, rather than using current tactile readings as observation context. The model is pretrained on NeoData, a large-scale…
NeoteAI Team, Fudan TEAI TeamResearch paper
This paper introduces Memory Decoder at Scale, scaling parametric long-term memory models up to 6.9B parameters and pretraining them on 300B tokens. To handle the computational bottleneck of constructing kNN distributions over 207B tokens, the authors develop a distributed Faiss pipeline using embedding compression, index sharding, and parallel search,…
Rubin Wei, Jiaqi Cao, Jiarui Wang, Junming Zhang, et al.Research paper
OvisOCR2 is a 0.8B end-to-end document parsing model that converts document page images into Markdown, covering text, formulas, tables, and visual regions. It uses a data engine combining filtered real-document annotations with synthetic pages generated from HTML sources. Training includes supervised fine-tuning, reinforcement learning (GRPO) on a 4B…
Shiyin Lu, Yinglun Li, Yu Xia, Yuhui Chen, et al.DeepSeek
DSpark is a speculative decoding framework that combines a semi-autoregressive draft model with confidence-scheduled verification to accelerate LLM inference. The draft model uses a parallel backbone (based on DFlash) to generate tokens in one pass, followed by a lightweight sequential head (Markov or RNN) to model inter-token dependencies, mitigating…
Xin Cheng, Xingkai Yu, Chenze Shao, Jiashi Li, et al.Moonshot AI
PerceptionBench is a benchmark introduced by Moonshot AI to evaluate atomic visual perception in Multimodal Large Language Models (MLLMs). It addresses limitations of existing benchmarks that conflate perception with reasoning or knowledge. The benchmark was constructed bottom-up: failures of frontier MLLMs on 42 existing benchmarks were attributed to…
Zichao Lin, Yifeng Xie, Bowen Qu, Haiming Wang, et al.OpenAI
This paper introduces deployment simulation, a method for predicting LLM safety before release by resampling the next assistant response from de-identified production conversation prefixes using a candidate model. The authors evaluate this approach across GPT-5-series deployments, finding that it produces informative estimates of post-deployment…
Marcus Williams, Hannah Sheahan, Cameron Raymond, Tomek Korbak, et al.