爱可可AI前沿推介(10.27)

[LG] Rewriting Scaling Laws for Synthetic Data with Compute and Data Constraints
[Princeton University & Stanford University]
https://arxiv.org/abs/2610.31867
---
[CL] LazyFlash: Forward-Dense, Backward-Sparse Attention for Accelerated Long-Context Fine-Tuning
[Microsoft]
https://arxiv.org/abs/2610.31911
---
[LG] Learning Positive Linear Combinations of ReLUs in Fully Polynomial Time
[UW Madison & MIT]
https://arxiv.org/abs/2610.32940
---
[AI] Don't Forget What You Just Learned: Self-Directed Repair Turns Agent Mistakes Into Lasting Lessons
[Princeton University & Meta FAIR]
https://arxiv.org/abs/2610.31846
---
[LG] BiLoCo: Binary Low-Rank Corrections for Post-Training Quantization
[MIT]
https://arxiv.org/abs/2610.32243