# DeltaV: Thinking with Visual State Updates in Unified Large Multimodal Models **A unified large multimodal model that thinks with visual state updates—modeling only the sparse, reasoning-critical changes across reasoning steps instead of regenerating full images.** [![arXiv](https://img.shields.io/badge/Arxiv-DeltaV-b31b1b.svg?logo=arXiv)](https://arxiv.org/abs/2607.08434) [![HuggingFace](https://img.shields.io/badge/HuggingFace-DeltaV--2B-yellow.svg?logo=HuggingFace)](https://huggingface.co/wpj20000/DeltaV-2B/tree/main) [![ModelScope](https://img.shields.io/badge/ModelScope-DeltaV--2B-green.svg)](https://www.modelscope.cn/models/wpj2003/DeltaV-2B) [![Dataset](https://img.shields.io/badge/Dataset-StructCoT-orange.svg)](https://www.modelscope.cn/datasets/wpj2003/StructCoT) [![Website](https://img.shields.io/badge/Website-DeltaV-blue.svg)](https://pengjie-w.github.io/DeltaV/) [![Demo](https://img.shields.io/badge/Demo-Live-purple.svg)](http://vlrlabmonkey.xyz:10088/)
--- ## News * ```2026.07.09 ``` 🚀 We release [DeltaV-2B](https://huggingface.co/wpj20000/DeltaV-2B/tree/main), a unified large multimodal model for interleaved multimodal reasoning. ## Introduction DeltaV is a unified large multimodal model (ULMM) designed to think with visual state updates during interleaved multimodal reasoning. Conditioned on historical visual states, it incrementally predicts compact **visual update tokens** that capture sparse but reasoning-critical changes across reasoning steps, avoiding repeated modeling of unchanged content. Token budgets are dynamically allocated by the **TSIM Router** according to temporal visual variation, and visual states are encoded by the **TSIM-Tok** tokenizer. This repository releases the **DeltaV-2B ULMM** together with the **TSIM-Tok tokenizer**, inference scripts, and tiny samples. ## TODO / Roadmap This release focuses on **inference** with the pretrained DeltaV-2B checkpoint. The following components are not included yet and will be released in a future update: - [ ] **StructCoT dataset** — the full StructCoT dataset and training data. (A small inference example, `data/struct_infer_sample.json`, is included so inference is runnable; the benchmark numbers in the tables below are kept as a record.) - [ ] **Zebra-CoT / StructCoT evaluation (scoring) tools** — the LLM-API-based scorers and their guide, pending the public StructCoT release. - [ ] **DeltaV training** — the two-stage DeltaV ULMM training scripts and configs. - [ ] **TSIM-Tok training and testing** — the TSIM-Tok tokenizer training and reconstruction-evaluation code (the tokenizer *model* is retained as DeltaV's visual backbone). ## DeltaV Workflow https://github.com/user-attachments/assets/724b2ff7-279a-4139-a9a0-de734014431d ## Repository Layout ```text deltav/ DeltaV model code: modeling, processing, configuration, backbone tsim_tok/ TSIM-Tok visual tokenizer and TSIM Router inference/ Inference scripts/ Ready-to-run scripts for DeltaV inference and data utilities configs/ Model and acceleration configs data/ Tiny samples docs/ Extended tutorials and README media assets tools/ Data processing and inference post-processing ``` ## Installation See [INSTALL.md](INSTALL.md) for the full setup guide. Quick version: ```bash conda create -n deltav python=3.10 -y && conda activate deltav pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121 pip install -r requirements.txt ``` ### Download Checkpoints Download our models from Huggingface. ```bash pip install huggingface_hub python tools/download_model.py -n DeltaV-2B # DeltaV-2B (legacy CLI identifier) ``` You can also download our models from ModelScope. ```bash pip install modelscope python tools/download_model.py -t modelscope -n DeltaV-2B # DeltaV-2B (legacy CLI identifier) ``` The released checkpoint is placed under `weights/`: ```text weights/ deltav_2b/ ``` ## Inference > **DeltaV training is not included in this release.** The two-stage DeltaV ULMM training recipe (on top of a frozen TSIM-Tok) will be added back later — see [TODO / Roadmap](#todo--roadmap). ### DeltaV Inference ```bash # Pure inference for evaluation. Outputs text-only .json, then merge + extract answers. MODEL_PATH=weights/deltav_2b \ JSON_PATH=data/zebra_infer_sample.json \ bash scripts/deltav/infer_deltav.sh ``` To also decode and save generated images, set `VIS_ARGS`. This streams results to `.jsonl` and skips merge/extract steps. Add `--concat_gt_images` to dump a ground-truth montage alongside each prediction. ```bash MODEL_PATH=weights/deltav_2b \ JSON_PATH=data/zebra_infer_sample.json \ VIS_ARGS="--decode_and_save_image --concat_gt_images" \ bash scripts/deltav/infer_deltav.sh ``` ### Input format & token budgets Each inference sample carries the prompt, its input/output image paths, and (optionally) the per-image incremental token budget: ```json { "config": "Visual Logic & Strategic Games - Tetris", "input_prompt": "Fill the entire grid EXCEPT ...", "input_image": ["/abs/path/problem.jpg"], "output_image": ["/abs/path/reasoning_01.jpg", "..."], "num_tokens": [144, 100, 81, ...] } ``` `num_tokens` is the per-image incremental token budget: the first image always uses the base budget `n_base`; each subsequent image uses a routed budget. There are two ways to supply it: - **Precomputed (default)** — if samples already carry `num_tokens`, pass `--use_json_num_tokens` (this is the default `TOKEN_ARGS` in `infer_deltav.sh`). - **TSIM Router (on the fly)** — otherwise set `TOKEN_ARGS="--use_tsim_router --visual_extractor_repo --visual_extractor_ckpt "`. The router (`deltav/tsim_tok/tsim_router.py`) measures temporal visual change with a frozen **DINOv2 ViT-B/14** and maps it to budgets via `tools/data_processing/tsim_intervals.json`. DINOv2 is fetched automatically by `torch.hub` on first use; to run offline, clone it and pass the local paths (see [INSTALL.md](INSTALL.md)). > **TSIM-Tok training and reconstruction evaluation are not included in this release.** The tokenizer *model* is retained as DeltaV's visual backbone, but its training/testing code will be added back later — see [TODO / Roadmap](#todo--roadmap). ## Documentation - [docs/eval_vlmevalkit.md](docs/eval_vlmevalkit.md): understanding benchmarks via VLMEvalKit. This guide is still being refined. The following guides will be published alongside the training / evaluation code (see [TODO / Roadmap](#todo--roadmap)): - [ ] `docs/data_and_token.md`: dataset format and the offline TSIM Router pipeline that turns image similarity into per-image token budgets. - [ ] `docs/eval_zebra_struct.md`: Zebra-CoT and StructCoT scoring. - [ ] `docs/advanced_zero3_gc.md`: ZeRO-3 and gradient checkpointing. - [ ] `docs/packing.md`: sequence packing, length computation, and packing training. ## Qualitative Examples

Qualitative comparison of multimodal reasoning

Qualitative comparison of multimodal reasoning. Full-image modeling (Base) exhibits inconsistent intermediate visual states, while DeltaV maintains consistent visual representations through visual state updates.

## Benchmark ### External Multimodal Reasoning and Understanding Evaluation
Model #Param VStar EMMA M3CoT MathVista VisuLogic MMBench MME‑P MMVP
General ULMMs
Chameleon7B32.58.616.121.74.56.05304.7
Anole7B34.06.615.822.53.76.25086.7
Janus-pro1B43.518.945.937.625.060.2139839.3
Janus-pro7B39.321.549.142.717.566.7150934.7
OmniGen27B41.414.750.360.20.176.1158835.3
Bagel7B70.128.731.472.528.983.7166569.3
EMU3.534B---28.311.413.779116.7
Understanding-centric MLLMs
Qwen3-VL2B71.722.253.061.111.577.1148245.0
Qwen3-VL8B83.730.661.277.622.585.2172959.3
InternVL3.52B68.112.751.360.826.078.2155248.7
InternVL3.58B69.116.659.974.129.782.7168857.3
Latent Interleaved Reasoning Models
Monet7B79.122.144.262.510.675.3163648.7
Mirage8B13.613.91.0829.90.412.35490.0
VPT-Det2B43.520.144.441.825.673.3151634.0
Explicit Interleaved Reasoning ULMMs
Bagel-Zebra-CoT7B64.920.662.672.1055.6164722.0
ThinkMorph7B64.422.448.867.86.578.214788.6
DeltaV [Weight]2B75.928.654.569.323.582.3155551.3
VStar, EMMA, M3CoT, MathVista, and VisuLogic are grouped as multimodal reasoning benchmarks, while MMBench, MME‑P, and MMVP are grouped as multimodal understanding benchmarks. ### In-domain Multimodal Reasoning Evaluation
Model #Param Zebra-CoT StructCoT
2D 3D Science Strategy Overall Strategy Planning Spatial Planning Logic Math Science Visual Search Jigsaw Restoration Overall
Understanding-centric MLLMs
GPT-5.2-67.619.373.354.453.743.133.842.176.350.487.057.155.7
Gemini-3.1 Pro-68.719.083.360.457.971.628.250.278.355.079.465.361.1
Gemini 3.0 Flash-66.519.478.454.554.755.033.344.874.848.483.664.957.8
Qwen3-VL2B44.313.230.39.224.33.431.44.641.429.480.839.332.9
Qwen3-VL8B50.716.956.022.736.621.625.413.159.339.383.846.541.3
InternVL3.58B29.711.448.919.827.56.936.317.536.132.075.841.035.1
Qwen2.5-VL72B43.217.350.125.834.114.834.431.448.036.584.947.042.4
General ULMMs
Chameleon7B13.33.05.29.97.95.612.54.19.113.123.514.411.8
Anole7B10.82.84.88.56.75.40.13.88.912.816.811.49.9
Janus-pro7B31.77.711.518.017.24.324.413.416.612.074.633.925.6
OmniGen27B26.51.39.69.711.80.625.31.58.410.178.128.521.8
Bagel7B43.314.744.516.329.716.424.912.849.035.584.649.038.9
EMU3.534B10.13.68.611.88.52.829.14.619.315.621.118.815.9
Latent Interleaved Reasoning Models
Monet7B37.512.015.123.021.92.319.921.933.825.859.633.828.1
Mirage8B2.22.510.712.47.00.914.312.435.822.511.030.418.2
VPT-Det2B32.33.56.518.715.37.526.58.814.515.173.135.925.9
Explicit Interleaved Reasoning ULMMs
Bagel-Zebra-CoT7B-----7.024.622.833.327.381.041.934.0
ThinkMorph7B43.011.631.422.927.221.419.526.443.426.084.149.938.7
DeltaV [Weight]2B78.920.041.138.344.616.453.066.030.145.684.362.651.1
The StructCoT test set excludes all samples originating from the Zebra-CoT dataset. ## Acknowledgements We would like to thank [Qwen3-VL](https://github.com/QwenLM/Qwen3-VL) and [VFMTok](https://github.com/CVMI-Lab/VFMTok) for providing base models and code, as well as their contributions to this field. We also thank [Zebra-CoT](https://huggingface.co/datasets/multimodal-reasoning-lab/Zebra-CoT) for providing a valuable interleaved multimodal reasoning dataset. We also thank everyone who contributed to this open-source effort. ## Copyright Please do not hesitate to share your valuable feedback—it is a key motivation that drives us to continuously improve our framework. **Note:** Our model is intended for academic research and non-commercial use only. If you are interested in a faster (smaller) or stronger model, please contact us at [xbai@hust.edu.cn](mailto:xbai@hust.edu.cn) or [ylliu@hust.edu.cn](mailto:ylliu@hust.edu.cn).