Future Blog Post

less than 1 minute read

Published: January 01, 2199

This post will show up by default. To disable scheduling of future posts, edit config.yml and set future: false.

Daily AI Papers — September 29, 2026

10 minute read

Published: September 29, 2026

1. YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality

Authors: Ruibin Yuan, Jiahao Pan, Junyan Jiang, Zhiyue Wu, Ziya Zhou, Jiankai Sun, Yizhi Li, Ge Zhang, Yicheng Gu, Zeyue Tian, Junyu Dai, Hanfeng Lin, Kai Li, Shangda Wu, Xuanjie Liu, Jiaming Wang, Zihan Liu, Yue Wang, Yinghao Ma, Hanzhi Yin, Kangrui Chen, Xinyue Zhang, Ziyang Ma, Mengqi Liao, Hejia Zhao, Guowei Huang, Chao Yan, Lei Ke, Jianwei Yu, Bei Liu, Joe Guo, Liumeng Xue, Gus Xia, Wei Xue, Yike Guo arXiv: arxiv.org/abs/2609.33757 Summary: YuE2 uses a single autoregressive/non-autoregressive Mixture-of-Transformers to plan a readable score, expand it into semantic music tokens, and render full-song audio. It beats evaluated public baselines on WildSongBench, approaches proprietary systems in expert listening, and supports score-guided edits and zero-shot covers from the same checkpoint.

Daily AI Papers — September 28, 2026

11 minute read

Published: September 28, 2026

1. FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders

Authors: Hongyang Du, Yunfei Xie, Junjie Ye, Jiawei Yang, Xiaoyan Cong, Haodong Zhang, Yongchao Huang, Haiyu Wu, Zongxia Li, Shihang Gui, Dawei Liu, Runhao Li, Jingcheng Ni, Chen Wei, Randall Balestriero, Yue Wang arXiv: arxiv.org/abs/2609.31620 Summary: FuseReg trains representation autoencoders over random subsets of visual-encoder layers, penalizing sensitivity to cross-layer disagreement and allowing one decoder to reconstruct from full, sparse, or single-layer fusions. On ImageNet-256, it improves reconstruction flexibility and cuts unguided gFID by 27% with decoder replacement alone and 29% when regularizing both decoder and diffusion training.

Daily AI Papers — September 27, 2026

9 minute read

Published: September 27, 2026

1. Single-stream Policy Optimization

Authors: Zhongwen Xu, Zihan Ding arXiv: arxiv.org/abs/2509.13232 Summary: Single-stream Policy Optimization replaces group-based baselines with a persistent KL-adaptive value tracker and globally normalized advantages, avoiding degenerate groups and synchronization barriers in LLM reinforcement learning. On five hard mathematics benchmarks with Qwen3-8B, it improves average maj@32 by 3.4 percentage points over GRPO while converging more smoothly and wasting less computation.