本期播客精华汇总: [CV] LLMs can see and hear without any training (大语言模型无需任何训练就能看会听): Meta AI 提出 Multimodal Iterative LLM Solver (MILS) 框架,无需训练即可赋予 LLM 多模态能力,利用 LLM 的推理能力和现成的多模态模型,实现零样本多模态学习,并在多种任务上取得 SOTA 结果。核心创新在于无需训练和利用 LLM 涌现能力。 [CL] Beyond Turn-taking:Introducing Text-based Overlap into Human-LLM Interactions (超越轮流发言:在人与大模型交互中引入文本重叠): Sungkyunkwan University & Google DeepMind 研究人员提出在人机文本交互中引入文本重叠机制,模仿自然人际对话。开发 OverlapBot 原型,用户研究表明重叠机制提升了沟通性、沉浸感和互动速度。核心创新在于突破传统轮流模式,提升人机对话自然性。 [LG] Joint Learning of Energy-based Models and their Partition Function (基于能量的模型及其配分函数的联合学习): Google DeepMind 提出联合学习能量模型及其配分函数的新框架,解决 EBMs 配分函数难以计算的问题。提出 Min-Min 优化公式和双重随机梯度下降算法,无需 MCMC 即可训练,并在多标签分类和标签排序任务上验证有效性。核心创新在于联合学习配分函数和无需 MCMC。 [LG] Diverse Preference Optimization (多样化偏好优化): Meta 提出 Diverse Preference Optimization (DivPO) 方法,解决 LLM 后训练阶段多样性坍缩问题。DivPO 在偏好优化中引入多样性考量,选择高质量但不太常见的回复作为优选样本,显著提升生成内容多样性,同时保持质量。核心创新在于偏好优化中引入多样性考量。 [LG] Think Smarter not Harder:Adaptive Reasoning with Inference Aware Optimization (更聪明而不是更努力地思考:基于推算感知优化的自适应推理): MetaAI 提出 Inference Budget-Constrained Policy Optimization (IBPO) 算法,让 LLM 具备推理预算意识,根据问题难度自适应调整推理长度。IBPO 在 MATH500 数据集上显著提升了受控推理预算下的性能,核心创新在于推理预算约束和自适应推理长度。完整推介:https://mp.weixin.qq.com/s/1JVAk0_nICg6QOhjvQGPJA
No persons identified in this episode.
This episode hasn't been transcribed yet
Help us prioritize this episode for transcription by upvoting it.
Popular episodes get transcribed faster
Other recent transcribed episodes
Transcribed and ready to explore now
SpaceX Said to Pursue 2026 IPO
10 Dec 2025
Bloomberg Tech
Don’t Call It a Comeback
10 Dec 2025
Motley Fool Money
Japan Claims AGI, Pentagon Adopts Gemini, and MIT Designs New Medicines
10 Dec 2025
The Daily AI Show
Eric Larsen on the emergence and potential of AI in healthcare
10 Dec 2025
McKinsey on Healthcare
What it will take for AI to scale (energy, compute, talent)
10 Dec 2025
Azeem Azhar's Exponential View
Reducing Burnout and Boosting Revenue in ASCs
10 Dec 2025
Becker’s Healthcare -- Spine and Orthopedic Podcast