2025.07.15 | 数据集支持虚拟人生成;强化学习需防数据污染。

HuggingFace 每日AI论文速递

本期的 12 篇论文如下:

[00:24] 🗣 SpeakerVid-5M: A Large-Scale High-Quality Dataset for Audio-Visual Dyadic Interactive Human Generation(SpeakerVid-5M:用于视听二元交互式虚拟人生成的大规模高质量数据集)

[01:12] 🤔 Reasoning or Memorization? Unreliable Results of Reinforcement Learning Due to Data Contamination(推理还是记忆?数据污染导致强化学习结果不可靠)

[02:03] 🤖 EmbRACE-3K: Embodied Reasoning and Action in Complex Environments(EmbRACE-3K:复杂环境中的具身推理与行动)

[03:02] 🤔 REST: Stress Testing Large Reasoning Models by Asking Multiple Problems at Once(REST:通过同时提问多个问题来压力测试大型推理模型)

[03:56] 🧮 Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation(递归混合:学习动态递归深度以实现自适应Token级别计算)

[04:46] 🧠 LayerCake: Token-Aware Contrastive Decoding within Large Language Model Layers(LayerCake:大语言模型层内的Token感知对比解码)

[05:39] ⚖ CompassJudger-2: Towards Generalist Judge Model via Verifiable Rewards(CompassJudger-2:基于可验证奖励的通用判别模型)

[06:27] 🎬 MoVieS: Motion-Aware 4D Dynamic View Synthesis in One Second(MoVieS:一秒内实现运动感知的四维动态视角合成)

[07:18] 🧮 A Practical Two-Stage Recipe for Mathematical LLMs: Maximizing Accuracy with SFT and Efficiency with Reinforcement Learning(数学大型语言模型的实用两阶段方案:通过监督微调最大化准确率,通过强化学习优化效率)

[08:05] 🇰 From KMMLU-Redux to KMMLU-Pro: A Professional Korean Benchmark Suite for LLM Evaluation(从KMMLU-Redux到KMMLU-Pro:用于LLM评估的专业韩国基准套件)

[09:08] 🖼 DreamPoster: A Unified Framework for Image-Conditioned Generative Poster Design(DreamPoster:一个用于图像条件生成海报设计的统一框架)

[09:54] 🖼 Favicon Trojans: Executable Steganography Via Ico Alpha Channel Exploitation(Favicon木马:通过ICO Alpha通道利用实现的可执行隐写术)

【关注我们】

您还可以在以下平台找到我们,获得播客内容以外更多信息

小红书: AI速递

To listen to explicit episodes, sign in.

Stay up to date with this show

Sign in or sign up to follow shows, save episodes and get the latest updates.

Select a country or region

Africa, Middle East, and India

Asia Pacific

Europe

Latin America and the Caribbean

The United States and Canada