AI可可AI生活

[CL] From Bytes to Ideas: Language Modeling with Autoregressive U-Nets [FAIR at Meta] https://arxiv.org/abs/2506.14761 在小宇宙查看该单集文稿

[CL] Reasoning with Exploration: An Entropy Perspective [RUC & MSRA & SJTU] https://arxiv.org/abs/2506.14758 在小宇宙查看该单集文稿

[LG] Less is More: Undertraining Experts Improves Model Upcycling [Université de Montréal & Concordia University] https://arxiv.org/abs/2506.14126 在小宇宙查看该单集文稿

工作量会像气体一样膨胀填满给定的时间容器，因此主动设定紧迫的截止日期能激发最高效率，让你成为时间的主人。 在小宇宙查看该单集文稿

[LG] Wanting to Be Understood Explains the Meta-Problem of Consciousness [Google DeepMind] https://arxiv.org/abs/2506.12086 在小宇宙查看该单集文稿

[CL] Multipole Attention for Efficient Long Context Reasoning [UC Berkeley] https://arxiv.org/abs/2506.13059 在小宇宙查看该单集文稿

[CL] Direct Reasoning Optimization：LLMs Can Reward And Refine Their Own Reasoning for Open-Ended Tasks [Microsoft] https://arxiv.org/abs/2506.13351 在小宇宙查看该单集文稿

[LG] Verifying the Verifiers：Unveiling Pitfalls and Potentials in Fact Verifiers [Yonsei University & Stanford University & University of Washington] https://arxiv.org/abs/2506.13342 在小宇宙查看该单集文稿

[CL] Refract ICL：Rethinking Example Selection in the Era of Million-Token Models [Google DeepMind] https://arxiv.org/abs/2506.12346 在小宇宙查看该单集文稿

人生的智慧不在于追求一步到位的完美答案，而在于始终为自己保留更多的选择权和可能性。 在小宇宙查看该单集文稿

[LG] Spectral Estimation with Free Decompression [UC Berkeley & University of Melbourne] https://arxiv.org/abs/2506.11994 在小宇宙查看该单集文稿

[LG] TreeRL: LLM Reinforcement Learning with On-Policy Tree Search [Tsinghua University & California Institute of Technology] https://arxiv.org/abs/2506.11902 在小宇宙查看该单集文稿

[CL] Large Language Models and Emergence: A Complex Systems Perspective [Santa Fe Institute] https://arxiv.org/abs/2506.11135 在小宇宙查看该单集文稿

[LG] Tversky Neural Networks: Psychologically Plausible Deep Learning with Differentiable Tversky Similarity [Stanford University] https://arxiv.org/abs/2506.11035 在小宇宙查看该单集文稿

[CL] You Only Fine-tune Once：Many-Shot In-Context Fine-Tuning for Large Language Model [Google & University of Florida] https://arxiv.org/abs/2506.11103 在小宇宙查看该单集文稿

现代人的焦虑很大程度源于不断的比较。我们习惯用别人的成就来衡量自己，这种"比较"就像快乐的小偷，偷走了本该属于我们的幸福感。 在小宇宙查看该单集文稿

[CL] Draft-based Approximate Inference for LLMs [FuriosaAI & UW-Madison] https://arxiv.org/abs/2506.08373 在小宇宙查看该单集文稿

[LG] On the Similarities of Embeddings in Contrastive Learning [Yonsei University] https://arxiv.org/abs/2506.09781 在小宇宙查看该单集文稿

[LG] Sequential-Parallel Duality in Prefix Scannable Models [MIT CSAIL & Technical University of Munich] https://arxiv.org/abs/2506.10918 在小宇宙查看该单集文稿

[LG] Self-Adapting Language Models [MIT] https://arxiv.org/abs/2506.10943 在小宇宙查看该单集文稿

[LG] CoRT: Code-integrated Reasoning within Thinking [University of Science and Technology of China & Qwen Team & The Chinese University of Hong Kong] https://arxiv.org/abs/2506.09820 在小宇宙查看该单集文稿

[LG] Intention-Conditioned Flow Occupancy Models   C Zheng, S Park, S Levine, B Eysenbach   [Princeton University & UC Berkeley]   本文提出的Intention-Conditioned Flow Occupancy Models (InFOM)通过创新性地结合潜在意图推断与基于流匹配的未来状态占有率建模，并在预训练中优化ELBO、在微调中使用隐式广义策略改进，成功地从未标记的异构离线数据中学习到了能够显著提升下游任务性能的RL基础模型...

[LG] Solving Inequality Proofs with Large Language Models   J Sheng, L Lyu, J Jin, T Xia...   [Stanford University & UC Berkeley]   本文通过构建一个包含奥林匹克级别不等式的新数据集IneqMath，并设计了一套包含最终答案和详细步骤审查的LLM即评判者评估框架，揭示了当前顶尖大语言模型在解决不等式问题时普遍存在的“答案可能正确但推理过程往往不严谨”的巨大鸿沟，并指出模型规模和计算量扩展对此改...

[LG] Reinforcement Learning Teachers of Test Time Scaling   E Cetin, T Zhao, Y Tang   [Sakana AI]   本文通过提出强化学习教师（RLTs）框架，创新性地将RL教师模型的任务设定为在已知问题和答案的前提下生成优质解释，并利用基于学生理解度的密集奖励进行训练，从而高效地生成了无需后处理的高质量蒸馏数据，不仅显著提升了下游学生模型在复杂推理任务上的性能，甚至在零样本跨领域迁移和RL冷启动方面取得了...

[LG] Branched Schrödinger Bridge Matching   S Tang, Y Zhang, A Tong, P Chatterjee   [Duke-NUS Medical School & Quebec AI Institute]   本文提出了分支薛定谔桥匹配（BranchSBM）框架，通过创新地将分支的广义薛定谔桥问题分解为多个可解的非平衡条件随机最优控制问题，并引入参数化的漂移场和分支特定的生长过程以及多阶段训练算法，成功实现了从单一初始分布到多个不同目标分布的动态、能量感知的分支...