来自 @爱可可-爱生活 的第一手AI快报，用最通俗的语言，聊最前沿的人工智能科研进展～ #人工智能# #科技前沿#
[CL] Reasoning with Exploration: An Entropy Perspe
[LG] Less is More: Undertraining Experts Improves
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
[LG] Wanting to Be Understood Explains the Meta-Pr
[CL] Multipole Attention for Efficient Long Contex
[CL] Direct Reasoning Optimization：LLMs Can Reward
[LG] Verifying the Verifiers：Unveiling Pitfalls an
[CL] Refract ICL：Rethinking Example Selection in t
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。
[LG] Spectral Estimation with Free Decompression[U
[LG] TreeRL: LLM Reinforcement Learning with On-Po
[CL] Large Language Models and Emergence: A Comple
[LG] Tversky Neural Networks: Psychologically Plau
[CL] You Only Fine-tune Once：Many-Shot In-Context
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
[CL] Draft-based Approximate Inference for LLMs[Fu
[LG] On the Similarities of Embeddings in Contrast
[LG] Sequential-Parallel Duality in Prefix Scannab
[LG] Self-Adapting Language Models[MIT]https://arx
[LG] CoRT: Code-integrated Reasoning within Thinki
