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。
50 篇论文 · 按点赞排序
Jie Liu, Zilyu Ye, Linxiao Yuan +8 authors
A unified reinforcement learning framework is proposed for interleaved text and image generation, using GRPO and FlowGRPO with modifications to enable scalable multi-round generation.
Hyomin Lee, Sangwoo Park, Yumin Choi +3 authors
T-MAP, a trajectory-aware evolutionary search method, discovers adversarial prompts that bypass safety measures and achieve harmful outcomes through tool interactions in LLM agents.
Amartya Roy, Rasul Tutunov, Xiaotong Ji +2 authors
λ-RLM replaces unbounded recursive code generation with typed functional runtime based on λ-calculus, providing formal guarantees and improved efficiency for long-context reasoning tasks.
Zhengxian Wu, Kai Shi, Chuanrui Zhang +10 authors
A self-evolution training framework for multimodal reasoning uses unsupervised learning with self-consistency signals and group-relative policy optimization to improve performance without labeled data.
Mingzhe Zheng, Weijie Kong, Yue Wu +9 authors
Video generation methods face reliability challenges due to complex solution spaces and noise injection during ODE-to-SDE conversion, but a new approach constrains exploration within a learned data manifold to improve stability and quality.
Zhekai Chen, Yuqing Wang, Manyuan Zhang +1 authors
A large-scale dataset and benchmark are introduced to address limitations in multi-reference image generation by providing structured long-context supervision and standardized evaluation protocols.
Matan Ben-Yosef, Tavi Halperin, Naomi Ken Korem +6 authors
AVControl enables efficient, modular audio-visual generation by training control modalities as separate LoRA adapters on a parallel canvas within LTX-2, achieving superior performance on diverse control tasks while requiring minimal computational resources.
Zhifei Yang, Guangyao Zhai, Keyang Lu +6 authors
FlowScene is a tri-branch generative model that combines multimodal graph conditioning with rectified flow modeling to produce realistic scenes with controlled geometry, appearance, and stylistic coherence.
Gabriel Orlanski, Devjeet Roy, Alexander Yun +6 authors
Software development is iterative, yet agentic coding benchmarks overwhelmingly evaluate single-shot solutions against complete specifications. Code can pass the test suite but become progressively harder to extend. Recent iterative benchmarks attempt to close this gap, but constrain the agent's design decisions too tightly to faithfully measure how code quality shapes future extensions. We introduce SlopCodeBench, a language-agnostic benchmark comprising 20 problems and 93 checkpoints, in which agents repeatedly extend their own prior solutions under evolving specifications that force architectural decisions without prescribing internal structure. We track two trajectory-level quality signals: verbosity, the fraction of redundant or duplicated code, and structural erosion, the share of complexity mass concentrated in high-complexity functions. No agent solves any problem end-to-end across 11 models; the highest checkpoint solve rate is 17.2%. Quality degrades steadily: erosion rises in 80% of trajectories and verbosity in 89.8%. Against 48 open-source Python repositories, agent code is 2.2x more verbose and markedly more eroded. Tracking 20 of those repositories over time shows that human code stays flat, while agent code deteriorates with each iteration. A prompt-intervention study shows that initial quality can be improved, but it does not halt degradation. These results demonstrate that pass-rate benchmarks systematically undermeasure extension robustness, and that current agents lack the design discipline iterative software development demands.
Allen Nie, Xavier Daull, Zhiyi Kuang +10 authors
Generative optimization using large language models faces challenges due to implicit design decisions about artifact modification and learning evidence that significantly impact success across different applications.
Kexin Huang, Haoming Meng, Junkang Wu +10 authors
Reinforcement learning with verifiable rewards improves language model reasoning by focusing on the direction of parameter updates rather than their magnitude, enabling better test-time extrapolation and training-time reweighting methods.
Yunzhe Wang, Runhui Xu, Kexin Zheng +4 authors
GameplayQA presents a framework for evaluating multimodal large language models' perception and reasoning capabilities in 3D environments through annotated multiplayer gameplay videos.
Youbin Kim, Jinho Park, Hogun Park +1 authors
Group3D is a multi-view open-vocabulary 3D detection framework that integrates semantic constraints into instance construction through semantic compatibility groups, improving accuracy in pose-known and pose-free settings.
Yunheng Li, Hangyi Kuang, Hengrui Zhang +4 authors
Researchers developed a token-level reinforcement learning method called PEPO that improves multimodal chain-of-thought reasoning by distinguishing visual grounding from inference through perception-exploration policy optimization.
Hwasik Jeong, Seungryong Lee, Gyeongjin Kang +4 authors
A two-expert architecture for pose-free 3D Gaussian Splatting separates geometry estimation from appearance synthesis, achieving superior performance compared to unified monolithic approaches.
Dongyoung Kim, Sumin Park, Woomin Song +6 authors
A systematic training framework called RoboAlign is proposed to enhance embodied reasoning in multimodal large language models by using zero-shot natural language reasoning and reinforcement learning to improve action accuracy and bridge the gap between language and low-level actions in vision-language-action models.
Baifeng Shi, Stephanie Fu, Long Lian +10 authors
AutoGaze is a lightweight module that reduces redundant video patches before processing by vision transformers or multi-modal large language models, enabling efficient processing of long, high-resolution videos while maintaining performance.
Fabio Valerio Massoli, Andrey Kuzmin, Arash Behboodi
Efficient reasoning in language models is reformulated as a lossy compression problem using conditional information bottleneck to reduce cognitive overhead while maintaining performance.
Haichao Zhang, Yijiang Li, Shwai He +5 authors
A VLM-guided JEPA-style latent world modeling framework combines dense-frame dynamics modeling with long-horizon semantic guidance through dual-temporal pathways to improve hand-manipulation trajectory prediction.
Jiazheng Xing, Fei Du, Hangjie Yuan +7 authors
LumosX framework enhances text-to-video generation through relational attention mechanisms and structured data pipelines for improved face-attribute alignment and subject consistency.
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