Fine-tunes diffusion generative models with online reinforcement learning via Aggregation-First Reward Tilting (DiffusionART): it aggregates per-prompt rewards into a shared reward-improving direction and adds a trust-region restoration update to replace EMA, improving training stability and GPU efficiency for image and video generation.
AI Agent Papers2026
Jiahao Wang, Hongjin Qian +3Beijing Academy of Artificial Intelligence, Institute of Automation, Chinese Academy of Sciences +1
Measures whether coding agents can build agents that truly satisfy natural-language requirements via AgentSWE: a 25-task benchmark spanning Creation, Editing, and Optimization. Builders develop against few illustrative cases; acceptance runs frozen submissions on held-out cases and requirement-derived rubrics, exposing overfitting, hollow runtimes, and unsupported completion claims.
Xiaohan Qin, Xiangdong Zhang +8School of AI, Shanghai Jiao Tong University, Shanghai Innovation Institute +3
Enables multi-head, depth-wise residual retrieval in Transformers by expanding a single residual stream into multiple context-aware streams using causal depthwise convolutions. Improves pretraining convergence and downstream benchmark scores with minimal overhead, yielding notable iso-loss compute speedups in large MoE models.