Research map / Image generation / Autoregressive and latent generation
Group Critical-token Policy Optimization for Autoregressive Image Generation
Guohui Zhang, Hu Yu, Xiaoxiao Ma, JingHao Zhang, Yaning Pan, Mingde Yao, Jie Xiao, Linjiang Huang, Jie Huang, Feng Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| optimize only 30% critical tokens with GCPO instead of all tokens with GRPO | Janus-Pro-7B · GenEval | overall: 0.87 -> 0.90 | Improved |
| GCPO instead of GRPO | Janus-Pro-1B · GenEval | overall: 0.84 -> 0.85 | Improved |
| GCPO instead of GRPO | LlamaGen (HPS reward) · GenEval | overall: 0.39 -> 0.42 | Improved |
| GCPO instead of GRPO on out-of-distribution T2I-CompBench | Janus-Pro-7B · T2I-CompBench Shape | score: 0.3999 -> 0.5173 | Improved |
| GCPO instead of GRPO on Janus-Pro-1B, T2I-CompBench Complex | Janus-Pro-1B · T2I-CompBench Complex | score: 0.3159 -> 0.3059 | Hurt |
| GCPO instead of GRPO with HPS reward | Janus-Pro-7B · DrawBench PickScore | PickScore: 22.12 -> 22.10 | No clear effect |
| add dynamic advantage weight on top of three critical-token types (ablation) | GCPO (Janus-Pro-1B) · GenEval | overall: 0.83 -> 0.85 | Improved |
| combine initial, high-entropy-gradient and low-similarity tokens instead of a single type (ablation) | GCPO (Janus-Pro-1B) · GenEval | overall: 0.82 -> 0.83 (best single type vs all three) | Improved |
| select critical tokens instead of random tokens at the same budget (ablation) | GRPO · GenEval | score | Improved |
| train GRPO on the remaining 70% of tokens rather than critical 30% (ablation) | GRPO · GenEval | score | Hurt |
| raise critical token selection above 10% of tokens per type (ablation) | GRPO · GenEval | score: +2.81 within 10%, +0.95 beyond | No clear effect |
Closest papers by meaning
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.89)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.88)
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.88)
- MaskSense: Confronting the Visual Exploration Trap in Masked Image Generation (NeurIPS 2026 · similarity 0.87)
- Beyond the Dirac Delta: Mitigating Diversity Collapse in Reinforcement Fine-Tuning for Image Generation (NeurIPS 2026 · similarity 0.87)
- GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
