Auto-Rubric as Reward: From Implicit Preferences to Explicit Multimodal Generative Criteria
Paper ā¢ 2605.08354 ā¢ Published ā¢ 21
AI & ML interests
None defined yet.
Recent Activity
View all activity
Papers
View all Papers
FerrieĀ
authored a
paper about 20 hours ago
FerrieĀ
updated a
model 1 day ago
JuanxiĀ
authored a
paper 2 days ago
JuanxiĀ
submitted a
paper to Daily Papers 2 days ago
JuanxiĀ
updated a
model 2 days ago
JuanxiĀ
published a
model 2 days ago
iChubaiĀ
published a
dataset 20 days ago
iChubaiĀ
updated a
dataset 25 days ago
JuanxiĀ
published a
dataset 28 days ago
Post
4426
š¢ Awesome Multimodal Modeling
We introduce Awesome Multimodal Modeling, a curated repository tracing the architectural evolution of multimodal intelligenceāfrom foundational fusion to native omni-models.
š¹ Taxonomy & Evolution:
Traditional Multimodal Learning ā Foundational work on representation, fusion, and alignment.
Multimodal LLMs (MLLMs) ā Architectures connecting vision encoders to LLMs for understanding.
Unified Multimodal Models (UMMs) ā Models unifying Understanding + Generation via Diffusion, Autoregressive, or Hybrid paradigms.
Native Multimodal Models (NMMs) ā Models trained from scratch on all modalities; contrasts early vs. late fusion under scaling laws.
š” Key Distinction:
UMMs unify tasks via generation heads; NMMs enforce interleaving through joint pre-training.
š Explore & Contribute: https://github.com/OpenEnvision/Awesome-Multimodal-Modeling
We introduce Awesome Multimodal Modeling, a curated repository tracing the architectural evolution of multimodal intelligenceāfrom foundational fusion to native omni-models.
š¹ Taxonomy & Evolution:
Traditional Multimodal Learning ā Foundational work on representation, fusion, and alignment.
Multimodal LLMs (MLLMs) ā Architectures connecting vision encoders to LLMs for understanding.
Unified Multimodal Models (UMMs) ā Models unifying Understanding + Generation via Diffusion, Autoregressive, or Hybrid paradigms.
Native Multimodal Models (NMMs) ā Models trained from scratch on all modalities; contrasts early vs. late fusion under scaling laws.
š” Key Distinction:
UMMs unify tasks via generation heads; NMMs enforce interleaving through joint pre-training.
š Explore & Contribute: https://github.com/OpenEnvision/Awesome-Multimodal-Modeling
- 3 replies
iChubaiĀ
updated a
dataset about 1 month ago
iChubaiĀ
published 2
datasets about 1 month ago
iChubaiĀ
updated a
dataset about 2 months ago
FerrieĀ
updated a
dataset about 2 months ago
JuanxiĀ
published a
dataset about 2 months ago
JuanxiĀ
authored a
paper 2 months ago
JuanxiĀ
authored a
paper 3 months ago
Post
2166
Recent Updates on ScalingOpt | Your Stars are Appreciated
We are pleased to announce several key updates to the ScalingOpt project:
Pyramid Visualization Structure
Following a suggestion from Yufei, we have introduced a pyramid-based visualization framework to systematically outline the layered architecture of Foundation Modelsāfrom foundational principles to infrastructure-level details. This addition is designed to assist teams in organizing and presenting related materials more clearly.
Integration of Optimizer Summaries by Yifeng
We extend a warm welcome to Yifeng (author of MARS), who has joined the project. He has contributed a comprehensive summary of over 100 optimizers, now available in ScalingOpt. This resource can be accessed via the āOptimization Summary Sheetā on the homepage or under the Optimizers page, featuring a reader-friendly interface that supports easy viewing, downloading, and citation.
Growing Community of Members
We continue to update and expand the list of active members. Researchers interested in Optimization & Efficient AI are encouraged to join and participate in discussions. Feedback and suggestions are also highly welcomed and will be reviewed and incorporated on an ongoing basis.
Tutorials in Progress
The tutorial development is actively underway. Currently, we have prepared over 300 slides and are refining and expanding the content in collaboration with contributors.
This community is driven purely by passion and a commitment to open knowledge sharing. Your support through starring the repository is greatly appreciated!
We are pleased to announce several key updates to the ScalingOpt project:
Pyramid Visualization Structure
Following a suggestion from Yufei, we have introduced a pyramid-based visualization framework to systematically outline the layered architecture of Foundation Modelsāfrom foundational principles to infrastructure-level details. This addition is designed to assist teams in organizing and presenting related materials more clearly.
Integration of Optimizer Summaries by Yifeng
We extend a warm welcome to Yifeng (author of MARS), who has joined the project. He has contributed a comprehensive summary of over 100 optimizers, now available in ScalingOpt. This resource can be accessed via the āOptimization Summary Sheetā on the homepage or under the Optimizers page, featuring a reader-friendly interface that supports easy viewing, downloading, and citation.
Growing Community of Members
We continue to update and expand the list of active members. Researchers interested in Optimization & Efficient AI are encouraged to join and participate in discussions. Feedback and suggestions are also highly welcomed and will be reviewed and incorporated on an ongoing basis.
Tutorials in Progress
The tutorial development is actively underway. Currently, we have prepared over 300 slides and are refining and expanding the content in collaboration with contributors.
This community is driven purely by passion and a commitment to open knowledge sharing. Your support through starring the repository is greatly appreciated!
- 1 reply
Post
2625
ScalingOpt is continuously evolving! We are steadily expanding the Community section with new content. For our Blog, we've launched by featuring work from Jianlin Su and are actively translating insightful posts from scientific communities into English to share on ScalingOpt (we'll keep curating excellent community blogs and providing English versions alongside the originals).
We operate under the Creative Commons Attribution-NonCommercial principle, sharing knowledge freely and openly. We welcome your ideas, suggestions, and feedback to help shape ScalingOpt's future.
If you find this initiative valuable, please consider following and starring the project to show your support. Thank you!
We operate under the Creative Commons Attribution-NonCommercial principle, sharing knowledge freely and openly. We welcome your ideas, suggestions, and feedback to help shape ScalingOpt's future.
If you find this initiative valuable, please consider following and starring the project to show your support. Thank you!
- 2 replies