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..
Transition Matching: Scalable and Flexible Generative Modeling
Authors: Neta Shaul, Uriel Singer, Itai Gat, Yaron Lipman
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments We evaluate the performance of our Transition Matching (TM) variants Difference TM (DTM), with T = 32 TM steps, Autoregressive TM (ARTM-2,3) with T = 2, 3 (resp.), and Full History TM (FHTM-2,3) with T = 2, 3 (resp.) on the text-to-image generation task. ... Our main evaluation results are reported in Tables 1 and 8 (in Appendix) on the Di T architecture. We find that DTM outperforms all baselines, and yields the best results across all metrics... |
| Researcher Affiliation | Collaboration | Neta Shaul ,1, Uriel Singer ,2 Itai Gat2 Yaron Lipman2 1Weizmann Institute of Science, 2FAIR at Meta |
| Pseudocode | Yes | Algorithm 1 Transition Matching Training Require: p T Data Require: qt,Y |T Process Require: T Number of TM steps 1: while not converged do 2: Sample t U([T 1]), XT p T 3: Sample (Xt, Y ) qt,Y |T ( |XT ) 4: L(θ) ˆD(Y, pθ Y |t( |Xt)) 5: θ θ γ θL Optimization step 6: end while 7: return θ |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: We use licensed data, code will be potentially released at a later date, all implementation details are provided in the main paper and appendix. |
| Open Datasets | Yes | Evaluation datasets are Parti Prompts [54] and MS-COCO [26] text/image benchmarks. |
| Dataset Splits | Yes | Evaluation datasets are Parti Prompts [54] and MS-COCO [26] text/image benchmarks. |
| Hardware Specification | Yes | Generation time for a single image on a single H100 GPU is provided in Table 11. |
| Software Dependencies | No | The paper mentions Python code for training in Figures 25, 26, and 27, and references a Di T backbone and flow matching loss, but it does not specify version numbers for Python, PyTorch, or any other software libraries or frameworks used. |
| Experiment Setup | Yes | All experiments are performed with the same 1.7B parameters Di T backbone (f θ) [32]... The models are trained for 500K iterations, with a 2048 total batch size, 1 e 4 constant learning rate and 2K iterations warmup. |