[CL] OMEGA: Can LLMs Reason Outside the Box in Math? Evaluating Exploratory, Compositional, and Transformative Generalization [dmodel.ai & UC Berkeley] arxiv.org

[CL] LongWriter-Zero: Mastering Ultra-Long Text Generation via Reinforcement Learning [Singapore University of Technology and Design & Tsinghua University] arxiv.org

[LG] Thought Anchors: Which LLM Reasoning Steps Matter? [Duke University & Aiphabet] arxiv.org

[LG] Hierarchical Reasoning Model [Sapient Intelligence, Singapore] https://arxiv.org/abs/2506.21734 --- [CL] Sequential Diagnosis with Language Models [Microsoft AI] https://arxiv.org/abs/2506.22405 --- [LG] Performance Prediction for Large Systems via Text-to-Text Regression [Google Research] https://arxiv.org/abs/2506.21718 --- [LG] Why Neural Network Can Discover Symbolic Structures with Gradient-based Training: An Algebraic and Geometric Foundation for Neurosymbolic Reasoning [University of Texas at Austin] https://arxiv.org/abs/2506.217 --- [LG] Transformers are Graph Neural Networks [University of Cambridge] https://arxiv.org/abs/2506.22084

成年人最高级的活法：在自己能掌控的领域里，拼尽全力；在自己无法掌控的领域里，保持内心的安宁。

[LG] Gaussian Invariant Markov Chain Monte Carlo [Google DeepMind & UCL] arxiv.org

[LG] Distilling Normalizing Flows [University of Oregon & HSE University & Picsart AI Research] arxiv.org

[LG] Overtuning in Hyperparameter Optimization [LMU Munich] arxiv.org

[CL] OctoThinker: Mid-training Incentivizes Reinforcement Learning Scaling [Shanghai Jiao Tong University] arxiv.org

[LG] Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales [ETH Zürich & DisneyResearch|Studios] https://arxiv.org/abs/2506.19713 --- [LG] Thought Anchors: Which LLM Reasoning Steps Matter? [Duke University & Aiphabet] https://arxiv.org/abs/2506.19143 --- [CL] LongWriter-Zero: Mastering Ultra-Long Text Generation via Reinforcement Learning [Singapore University of Technology and Design & Tsinghua University] https://arxiv.org/abs/2506.18841 --- [CL] OMEGA: Can LLMs Reason Outside the Box in Math? Evaluating Exploratory, Compositional, and Transformative Generalization [dmodel.ai & UC Berkeley] https://arxiv.org/abs/2506.18880

我们真正安身立命的核心竞争力，往往就藏在那些看起来最“没用”、最“务虚”的地方。

[CL] Data Efficacy for Language Model Training [Microsoft Research] arxiv.org