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。
Jiawei Chi, Shangchen Miao, Zhiyuan Shi +11 authors
AgentGarten couples simulators with a shared neural renderer to build real-time interactive virtual worlds, enabling agents to learn efficiently through inherited experience and adversarially distilled visual observations.
50 篇论文 · 按点赞排序
Arman Behnam, Sunglyoung Kim, Liangwei Yang
A benchmark of ten real long-term AI companion relationships reveals that past context is rarely needed, memory detectors fail on real messages, and persona reconstruction costs vary widely.
Deyuan Liu, Yihao Hu, Jingxuan Zhang +13 authors
UltraText Bench evaluates dense visual text generation across bilingual scenes using structured references and a vision-language judge, revealing trade-offs between clarity, fidelity, and workload difficulty.
Haoxiang Zhang, Qinglin Chen, Hiroaki Hayashi +9 authors
Prospective learning via Self-Retrospection Distillation uses hindsight experience to train foresight predictions, improving reinforcement learning when outcome rewards are uniform.
Team Kandinsky, Julia Agafonova, Bulat Akhmatov +85 authors
Kandinsky 6.0 Video is a family of diffusion models that generate synchronized audio-video clips with lip-sync via a dual-stream CrossDiT architecture and multi-stage training.
Bingchen Yao, Haobo Xu, Haokun Lin +6 authors
STEPQuant is a spatial-temporal quantization framework that reduces memory usage of linear attention recurrent states by allocating precision based on error magnitude and memory lifetime.
Tianyu Fu, Tengxuan Liu, Ruoxi Wang +4 authors
TokenRouter is a serving system for token-level routed LLM inference that improves throughput by using request-centric programming with asynchronous subservers and optimized delayed-batching schedulers.
Qiutong Chen, Yuchan Guo, Zhenlong Yuan +10 authors
VepAgent improves video event prediction by combining causal reasoning with tool-augmented reinforcement learning and a diagnostic tool library to recover missing visual evidence.
Guangyi Liu, Yong Liu, Jiangning Zhang
Assistants from 2011 answered and waited, and agents from 2023 did a task and stopped. In September 2026 Meta's Muse showed an agent for one person, with accounts, devices, memory and a conversation that lasts, closed, in a vendor's cloud, in one country. Such an agent is expected to act on a person's accounts and devices, remember them across weeks, speak first when it is worth it, and answer for what it did. It is a kind of software, not a model, and until now had no open counterpart. This report defines the personal agent in five questions and three horizons. It reads how Muse is built from Meta's public record and a copy of its production prompt, each statement marked by its source. It then presents nanoMuse, the open-source counterpart under the GPL-3.0, one agent on every device a person owns, with hands on the phone's screen and the computer's. They share one conversation over a relay anyone can run; every action goes through a Sentinel, memory is files the person can read, and the model is their choice. Its size and cost are given as estimates. What is open, memory with provenance, an evaluation suite for the hands and an open model for them, is set out as a roadmap.
Bingxuan Li, Siqi Song, Yizhuo Wu +3 authors
MotorMind enables general-purpose vision-language models to perform zero-shot robotic manipulation by reasoning directly from observations and issuing mid-level actions with asynchronous feedback, achieving high success without specialized training or external tools.
Wei Huang, Bohan Zhang, Chenzhi Liu +13 authors
Long-WAM scales causal world-action models for real-time robot control by using autoregressive video pretraining to leverage long visual histories, improving success on long-horizon manipulation tasks while maintaining low-latency deployment.
Bingxi Hou, Guochao Jiang, Guofeng Quan +4 authors
On-policy distillation with strict token alignment outperforms broader cross-tokenizer supervision by avoiding weak or conflicting gradient signals.
Zongxia Li, Yucheng Shi, Zhongzhi Li +7 authors
Recursive Self-Rewrite reconstructs diverse harness-assisted solutions into reusable training trajectories, significantly improving terminal task performance under a general harness.
Jiahao Zhan, Yan Wang, Yongrui Ma +5 authors
DuoMatching improves streaming video generation by combining joint and marginal distribution matching with frame-level image supervision via LatentBridge and temporal sampling.
Fei Zhang, Zhaochong An, Duncan Frost +5 authors
ProWAM improves robotic control by predicting sparse visual sub-goals to guide long-horizon action generation, achieving state-of-the-art results in simulation and real-world settings.
Jiajun Chen, Haoyu Wu, Mingda Jia +1 authors
Recursive Game Creator improves agentic game development through recursive design, building, programmatic playtesting, and review to enhance player enjoyment.
Core Team, Zongming Qiao, Ziyue Hua +146 authors
The MiMo-V2.6 series advances omni-modal foundation models through scaled reinforcement learning with larger batches, diverse agent environments, and robust reward grading, supported by stable training infrastructure.
Jiyoung Kim, Paul Hyunbin Cho, Jisu Nam +5 authors
GRACE compresses pretrained video autoencoders via a frozen base latent and learned residual, aligning outputs with a frozen diffusion transformer to enable efficient video generation with minimal adaptation.
Jinkai Zhang, Jingyi Xu, Yuanhong Yu +5 authors
SuperNav equips a pretrained multimodal large language model with a specialized agent harness and navigation tools to perform general-purpose robot navigation without fine-tuning, using a visual-point interface to link decisions to motion.
Wenqing Tian, Zeyu Zhang, Zhaocheng Liu +3 authors
PhysEvo enables a frozen model to recursively improve robot manipulation by diagnosing failures, revising skills, and testing corrections without weight updates.
Zhaochong An, Fei Zhang, Menglin Jia +10 authors
NAVA-WAM directly pretrains robot action policies from observation-only videos via native action-prior learning, improving scalability and control performance without requiring extensive action-labeled trajectories.
Shufan Shen, Zhongni Hou, Junshu Sun +5 authors
On-policy parameter update directions can be transferred to supervised fine-tuning via direction constraints to improve generalization while preserving training efficiency and high-quality trajectory usage.
Jinyuan Li, Chengsong Huang, Langlin Huang +4 authors
R-Quest sustains self-evolving reasoning by filtering invalid questions and enforcing mathematical novelty via validity and novelty feedback, preventing performance collapse over repeated training rounds.
Bo Yin, Xiaobin Hu, Jiaqi Zhao +1 authors
FrameMorrow improves long-horizon video generation by selecting historical frames based on predicted future information needs rather than current content.
Jiayi Chen, Wenxuan Song, Jingbo Wang +13 authors
UniWAM unifies physical reasoning, world generation, and action prediction through joint training on cleaned human and robot data with complementary supervision and flow-matching optimization, achieving strong embodied performance and scaling benefits.
Liancheng Fang, Zhuowei Li, Youngeun Kim +10 authors
ALoDLM improves diffusion language models by adaptively allocating computation per token based on difficulty, closing the quality gap with autoregressive models while preserving fast parallel generation.
Dahyun Chung, Siyoon Jin, Hyunwook Choi +5 authors
ME-World synchronizes first-person video generation for multiple interacting agents by jointly denoising shared token sequences with cross-view action conditioning and shared environment memory.
Hang He, Li Wang, Hao Chen +9 authors
CheckerBench evaluates coding agents on synthesizing static-analysis checkers from defect specifications across diverse repositories, revealing substantial challenges in reliable checker development.
Shukai Gong, Xuanran Zhai, Yintianrun Zhang +16 authors
ViGAR is a hierarchical framework that decomposes long-horizon robot manipulation into visual subgoal planning and subgoal-conditioned execution, improving success rates through shared world-model representations and in-context learning.
Zhengyu Fang, Seoyeon Hong, Jie Yang +4 authors
Latent-MOPD integrates multiple specialist LLMs into a single student via hidden-state and prediction-level supervision, improving performance across math, code, and logic benchmarks.
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