Update README.md
Browse files
README.md
CHANGED
|
@@ -7,7 +7,7 @@ library_name: ThinkMorph-7B
|
|
| 7 |
---
|
| 8 |
|
| 9 |
<p align="center">
|
| 10 |
-
<img src="https://github.com/ThinkMorph/ThinkMorph/
|
| 11 |
</p>
|
| 12 |
|
| 13 |
|
|
@@ -58,11 +58,11 @@ For installation, usage instructions, and further documentation, please visit ou
|
|
| 58 |
Multimodal reasoning demands synergistic coordination of language and vision. However, determining what constitutes meaningful interleaved reasoning is non-trivial, and current approaches lack a generalizable recipe.
|
| 59 |
We present **ThinkMorph**, a unified model that enables such generalization through a principled approach: treating text and images as complementary modalities that mutually advance reasoning.
|
| 60 |
<p align="center">
|
| 61 |
-
<img src="https://github.com/ThinkMorph/ThinkMorph/
|
| 62 |
</p>
|
| 63 |
Guided by this principle, we identify tasks requiring concrete, verifiable visual engagement and design a high-quality data pipeline that trains models to generate interleaved images and text as progressive reasoning traces.
|
| 64 |
<p align="center">
|
| 65 |
-
<img src="https://github.com/ThinkMorph/ThinkMorph/
|
| 66 |
</p>
|
| 67 |
|
| 68 |
ThinkMorph delivers substantial gains on **vision-centric** tasks, achieving an average improvement of 34.74% over the base model while consistently surpassing text-only and image-only modes.
|
|
@@ -70,7 +70,7 @@ By fine-tuning with **merely ~24K** samples, it achieves out-of-domain performan
|
|
| 70 |
|
| 71 |
Intriguingly, ThinkMorph unlocks emergent properties that represent a *hallmark of multimodal intelligence*: the elicitation of unseen visual manipulation skills, the self-adaptive switching between reasoning modes according to task complexity, and better test-time scaling via diversified thoughts.
|
| 72 |
<p align="center">
|
| 73 |
-
<img src="https://github.com/ThinkMorph/ThinkMorph/
|
| 74 |
</p>
|
| 75 |
These findings suggest promising directions for future work to characterize the emergent capabilities of unified models for multimodal reasoning.
|
| 76 |
|
|
|
|
| 7 |
---
|
| 8 |
|
| 9 |
<p align="center">
|
| 10 |
+
<img src="https://github.com/ThinkMorph/ThinkMorph/raw/main/assets/logo.png" width="40%"> <br>
|
| 11 |
</p>
|
| 12 |
|
| 13 |
|
|
|
|
| 58 |
Multimodal reasoning demands synergistic coordination of language and vision. However, determining what constitutes meaningful interleaved reasoning is non-trivial, and current approaches lack a generalizable recipe.
|
| 59 |
We present **ThinkMorph**, a unified model that enables such generalization through a principled approach: treating text and images as complementary modalities that mutually advance reasoning.
|
| 60 |
<p align="center">
|
| 61 |
+
<img src="https://github.com/ThinkMorph/ThinkMorph/raw/main/assets/interleaved_design.jpg" width="100%"> <br>
|
| 62 |
</p>
|
| 63 |
Guided by this principle, we identify tasks requiring concrete, verifiable visual engagement and design a high-quality data pipeline that trains models to generate interleaved images and text as progressive reasoning traces.
|
| 64 |
<p align="center">
|
| 65 |
+
<img src="https://github.com/ThinkMorph/ThinkMorph/raw/main/assets/thinkmorph_main.jpg" width="100%"> <br>
|
| 66 |
</p>
|
| 67 |
|
| 68 |
ThinkMorph delivers substantial gains on **vision-centric** tasks, achieving an average improvement of 34.74% over the base model while consistently surpassing text-only and image-only modes.
|
|
|
|
| 70 |
|
| 71 |
Intriguingly, ThinkMorph unlocks emergent properties that represent a *hallmark of multimodal intelligence*: the elicitation of unseen visual manipulation skills, the self-adaptive switching between reasoning modes according to task complexity, and better test-time scaling via diversified thoughts.
|
| 72 |
<p align="center">
|
| 73 |
+
<img src="https://github.com/ThinkMorph/ThinkMorph/raw/main/assets/emrging_prop.jpg" width="100%"> <br>
|
| 74 |
</p>
|
| 75 |
These findings suggest promising directions for future work to characterize the emergent capabilities of unified models for multimodal reasoning.
|
| 76 |
|