本期播客精华汇总: [CL] When More is Less: Understanding Chain-of-Thought Length in LLMs 论文颠覆了思维链推理中“越长越好”的传统认知,揭示了思维链长度与准确率之间存在U型关系,并指出存在最优CoT长度,该长度受模型能力和任务难度影响。论文提出了“长度过滤投票”方法,并强调训练数据中CoT长度分布对模型性能至关重要。 [LG] Curvature Tuning: Provable Training-free Model Steering From a Single Parameter 论文提出了曲率调整 (CT) 方法,一种无需训练的模型操控技术,仅通过单参数 β 调整模型决策边界曲率,即可显著提升模型的泛化性和鲁棒性。CT方法高效、可解释,为模型操控提供了新思路。 [LG] LLMs Can Easily Learn to Reason from Demonstrations Structure, not content, is what matters! 论文发现大语言模型从演示中学习推理能力,关键在于学习长链思考 (Long CoT) 的结构而非内容。即使使用内容错误的演示数据,只要结构正确,模型仍能有效学习推理。论文验证了少量Long CoT数据和参数高效微调方法 (LoRA) 在推理学习中的有效性。 [LG] Monte Carlo Tree Diffusion for System 2 Planning 论文提出了蒙特卡洛树扩散 (MCTD) 框架,将扩散模型的生成能力与蒙特卡洛树搜索 (MCTS) 的迭代搜索能力相结合,实现了AI的“系统2”式规划能力。MCTD通过“降噪作为树形展开”、“指导级别作为元动作”等技术,显著提升了长时程规划任务的性能和测试时计算可扩展性。 [LG] Towards Efficient Optimizer Design for LLM via Structured Fisher Approximation with a Low-Rank Extension 论文提出了基于结构化 Fisher 信息矩阵 (FIM) 近似的优化器设计框架,并基于此框架设计了内存高效优化器 RACS 和 Alice。实验表明,Alice 优化器在LLM预训练中实现了超过 2 倍的加速,RACS 具有极高的内存效率,为LLM高效优化器设计提供了新方法。完整推介:https://mp.weixin.qq.com/s/MYSHyANXeRws59QchSAiuA
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