Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Plug-and-Play Context Feature Reuse for Efficient Masked Generation
Authors: Xuejie Liu, Anji Liu, Guy Van den Broeck, Yitao Liang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate whether the use of Local-FEs can effectively lead to efficiency gains and, more importantly, whether it can achieve better trade-offs between generation quality and inference speed. To this end, we apply Re CAP to three representative MGM baselines and conduct a thorough evaluation. Our experiments span a diverse set of settings, varying in task (class-conditional vs. unconditional generation) and model architecture (decoder-only vs. encoder-decoder). The selected baselines are: i) Mask GIT [5], ii) MAR [29], and iii) MAGE [27]. |
| Researcher Affiliation | Academia | 1Institute for Artificial Intelligence, Peking University 2School of Computing, National University of Singapore 3Computer Science Department, University of California, Los Angeles |
| Pseudocode | No | The paper describes the method and procedure in narrative text and diagrams, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is publicly available at https://github.com/liebenxj/Re CAP. |
| Open Datasets | Yes | In particular, on Image Net256 [7] class-conditional generation, Re CAP achieves up to 2.4 faster inference than the base model with minimal performance drop, and consistently delivers better efficiency fidelity trade-offs under various generation settings. ... To validate this hypothesis, we analyze the representations computed by a pretrained Mask GIT [5] model on 50K samples from the Image Net256 validation set. |
| Dataset Splits | Yes | To validate this hypothesis, we analyze the representations computed by a pretrained Mask GIT [5] model on 50K samples from the Image Net256 validation set. |
| Hardware Specification | Yes | All inference times are re-evaluated using the official implementations on a single NVIDIA A800 GPU with a default batch size of 200 and reported as time per image. ... All inference times reported below are measured on an NVIDIA RTX 4090 GPU with a batch size of 32. |
| Software Dependencies | No | The paper mentions using pretrained models and official implementations but does not specify particular software or library versions (e.g., PyTorch version, Python version, CUDA version) used for the experiments. |
| Experiment Setup | Yes | Mask GIT adopts a cosine decoding schedule and a confidence-based token sampler. When adapting Re CAP, we use the same token sampler to obtain St after each Full-FE step (see Appendix B). Following the decoding schedule used in Mask GIT, early decoding steps reveal only a small number of tokens, while later steps decode progressively more. ... (Appendix A) Increased initial choice temperature to τ2(1) = 5.5 (still with linear decay) Temperature scaling for token sampling (τ1(t)): ... Polynomial unmasking schedule (replacing cosine): |