lmsdss commited on
Commit
ecffa6c
路
verified 路
1 Parent(s): e7db1c4

Add paper abstract, resource links, citation, and model card layout

Browse files
Files changed (1) hide show
  1. README.md +68 -2
README.md CHANGED
@@ -1,5 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  # RegionReasoner-7B
2
 
3
- This repository contains model files for RegionReasoner.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4
 
5
- - Code: https://github.com/lmsdss/RegionReasoner
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: transformers
3
+ pipeline_tag: image-text-to-text
4
+ tags:
5
+ - regionreasoner
6
+ - visual-reasoning
7
+ - visual-grounding
8
+ - multi-round-reasoning
9
+ - reinforcement-learning
10
+ - arxiv:2602.03733
11
+ ---
12
+
13
+ <div align="center">
14
+
15
  # RegionReasoner-7B
16
 
17
+ ### Region-Grounded Multi-Round Visual Reasoning
18
+
19
+ **ICLR 2026**
20
+
21
+ [![Paper](https://img.shields.io/badge/arXiv-2602.03733-b31b1b.svg)](https://arxiv.org/abs/2602.03733)
22
+ [![Code](https://img.shields.io/badge/GitHub-Code-181717.svg?logo=github)](https://github.com/lmsdss/RegionReasoner)
23
+ [![Training Data](https://img.shields.io/badge/Hugging_Face-Training_Data-FFD21E.svg)](https://huggingface.co/datasets/lmsdss/regionreasoner_data)
24
+ [![Test Data](https://img.shields.io/badge/Hugging_Face-Test_Data-FFD21E.svg)](https://huggingface.co/datasets/lmsdss/regionreasoner_test_data)
25
+
26
+ Wenfang Sun 路 Hao Chen 路 Yingjun Du 路 Yefeng Zheng 路 Cees G. M. Snoek
27
+
28
+ </div>
29
+
30
+ ---
31
+
32
+ RegionReasoner-7B is the released model for **RegionReasoner**, a reinforcement
33
+ learning framework for region-grounded, multi-round visual reasoning.
34
+ The framework combines explicit bounding-box references with global鈥搇ocal
35
+ semantic consistency, and is evaluated on detection and segmentation tasks
36
+ using **RegionDial-Bench**.
37
+
38
+ ## Abstract
39
+
40
+ Large vision-language models have achieved remarkable progress in visual reasoning, yet most existing systems rely on single-step or text-only reasoning, limiting their ability to iteratively refine understanding across multiple visual contexts. To address this limitation, we introduce a new multi-round visual reasoning benchmark with training and test sets spanning both detection and segmentation tasks, enabling systematic evaluation under iterative reasoning scenarios.
41
+
42
+ We further propose RegionReasoner, a reinforcement learning framework that enforces grounded reasoning by requiring each reasoning trace to explicitly cite the corresponding reference bounding boxes, while maintaining semantic coherence via a global-local consistency reward. This reward extracts key objects and nouns from both global scene captions and region-level captions, aligning them with the reasoning trace to ensure consistency across reasoning steps.
43
+
44
+ RegionReasoner is optimized with structured rewards combining grounding fidelity and global-local semantic alignment. Experiments on detection and segmentation tasks show that RegionReasoner-7B, together with our newly introduced benchmark RegionDial-Bench, considerably improves multi-round reasoning accuracy, spatial grounding precision, and global-local consistency, establishing a strong baseline for this emerging research direction.
45
+
46
+ *Abstract reproduced from [the paper](https://arxiv.org/abs/2602.03733)
47
+ by Wenfang Sun, Hao Chen, Yingjun Du, Yefeng Zheng, and Cees G. M. Snoek,
48
+ under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
49
+ Only paragraph formatting has been adjusted.*
50
+
51
+ ## Resources
52
+
53
+ | Resource | Link |
54
+ | --- | --- |
55
+ | Paper | [arXiv:2602.03733](https://arxiv.org/abs/2602.03733) |
56
+ | PDF | [Read the paper](https://arxiv.org/pdf/2602.03733) |
57
+ | Code and evaluation instructions | [lmsdss/RegionReasoner](https://github.com/lmsdss/RegionReasoner) |
58
+ | Training dataset | [lmsdss/regionreasoner_data](https://huggingface.co/datasets/lmsdss/regionreasoner_data) |
59
+ | Test dataset | [lmsdss/regionreasoner_test_data](https://huggingface.co/datasets/lmsdss/regionreasoner_test_data) |
60
+
61
+ ## Citation
62
 
63
+ ```bibtex
64
+ @inproceedings{sun2026regionreasoner,
65
+ title = {RegionReasoner: Region-Grounded Multi-Round Visual Reasoning},
66
+ author = {Sun, Wenfang and Chen, Hao and Du, Yingjun and Zheng, Yefeng and Snoek, Cees G. M.},
67
+ booktitle = {International Conference on Learning Representations},
68
+ year = {2026},
69
+ url = {https://arxiv.org/abs/2602.03733}
70
+ }
71
+ ```