RL训到图像生成，Gemma 4开源

重点关注

• OmniOpt: Taxonomy, Geometry, and Benchmarking of Modern Optimizers score 10
  入选 HF Daily Papers；HF 热度: 72 upvotes (+4)；有代码实现；关键词(1): pretraining
• MANCE: Manifold Aware Concept Erasure score 9
  入选 HF Daily Papers；HF 热度: 41 upvotes (+4)；有代码实现
• Perceptual Flow Matching for Few-Step Generative Modeling score 9
  入选 HF Daily Papers；HF 热度: 13 upvotes (+3)；有代码实现；关键词(1): distillation
• AlayaWorld: Long-Horizon and Playable Video World Generation score 10
  入选 HF Daily Papers；HF 热度: 78 upvotes (+4)；有代码实现；关键词(5): deployment, production, real-time, embodied, open-source
• PraMem: Practice-derived Experiential Memory for Long-horizon Behavior Prediction score 8
  入选 HF Daily Papers；HF 热度: 7 upvotes (+2)；有代码实现；关键词(2): compression, serving
• Bridging Interleaved Multi-Modal Reasoning as a Unified Decision Process score 10
  入选 HF Daily Papers；HF 热度: 34 upvotes (+4)；有代码实现；关键词(2): reasoning, vision-language
• TREK: Distill to Explore, Reinforce to Refine score 6
  入选 HF Daily Papers；HF 热度: 7 upvotes (+2)；关键词(4): distillation, GRPO, agentic, reasoning
• Parallelized Autoregressive Decoding for Omni-Modal Dense Video Captioning score 8
  入选 HF Daily Papers；HF 热度: 18 upvotes (+3)；有代码实现
• PixCon: Clean-Positive Contrastive Learning for Foundation-Model Semi-Supervised Segmentation score 6
  入选 HF Daily Papers；HF 热度: 2 upvotes (+1)；有代码实现
• CONFLUX: A Latent Diffusion Model for 3D Chest-CT Synthesis with RL Post-Training score 5
  入选 HF Daily Papers；HF 热度: 4 upvotes (+1)；关键词(1): post-training

也值得关注

• Gemma 4 Technical Report score 8
  入选 HF Daily Papers；HF 热度: 38 upvotes (+4)；关键词(1): reasoning
• CSympNet-ID: conformal-symplectic map learning for linearly damped Hamiltonian systems score 4
  机构: Peking University；关键词(1): scaling
• SAMPLe: SAM-based Optimizer for Prompt Learning in VLMs score 4
  关键词(1): vision-language；顶会接收: ECCV
• Breaking Spurious Correlations via Generative Randomization and Cross-Variant Self-Supervised Learning score 4
  关键词(1): fine-tune；顶会接收: CVPR
• Revisiting Scene Graph Generation from the Perspective of Detector-Conditioned Reachability score 4
  关键词(1): reasoning；顶会接收: ECCV
• Canopy: A Heterograph Foundation Model for Metabolic Engineering score 4
  关键词(2): lightweight, edge；顶会接收: ICML
• Flex-Forcing: Towards a Unified Autoregressive and Bidirectional Video Diffusion Model score 6
  入选 HF Daily Papers；HF 热度: 10 upvotes (+3)