Lightnews — Scholar-powered news

Juan Carlos Niebles @jcniebles.bsky.social · 15d

📢📢 Exciting news!

Our paper, "Exploring Diffusion Transformer Designs via Grafting," has been accepted as an Oral at #NeurIPS2025, with only 77 out of 21k submissions receiving this honor.

📄Paper: arxiv.org/abs/2506.05340
🌎Website: grafting.stanford.edu
🧑🏻‍💻Code: github.com/keshik6/graf...

Exploring Diffusion Transformer Designs via Grafting

grafting.stanford.edu

7

Juan Carlos Niebles @jcniebles.bsky.social · Sep 4

Strefer: our new work for auto-generating instruction data on space–time–focused video tasks: spatiotemporal reasoning, space-time reference understanding, etc. for Video LLMs

✅ Auto & scalable
✅ Fine-grained, space–time–grounded queries
✅ Effective

📄: arxiv.org/abs/2509.03501
🌐: strefer.github.io

Strefer: Empowering Video LLMs with Space-Time Referring and Reasoning via Synthetic Instruction Data

Next-generation AI companions must go beyond general video understanding to resolve spatial and temporal references in dynamic, real-world environments. Existing Video Large Language Models (Video LLM...

arxiv.org

1 1 2

Juan Carlos Niebles @jcniebles.bsky.social · Jun 16

Check out a new episode of The AI Research Lab - Explained on Multimodal AI.

Had a blast creating this with the @salesforce.com team!

youtu.be/r98jGdLtO6Q

What is Multimodal AI? | The AI Research Lab - Explained

YouTube video by Salesforce

youtu.be

2

Juan Carlos Niebles @jcniebles.bsky.social · Jun 12

Congrats Chaitanya on winning the BEST PAPER AWARD 🥇 🏆

Check out details of our work:

arxiv.org/abs/2504.12513

3

Juan Carlos Niebles @jcniebles.bsky.social · Jun 12

Our first #cvpr2025 poster is up!

🕐Come check it out right now until 13:00

“AdaVid: Adaptive Video-Language Pretraining”

🪧ExHall D Poster # 203

📝 arxiv.org/abs/2504.12513

2

Juan Carlos Niebles @jcniebles.bsky.social · Jun 11

Just finished a day at the #CVPR2025 Area Chair workshop. Lots of interesting discussions and ideas, reconnection with colleagues and friends.

Had the chance to present our ViUnit poster to fellow ACs. If you missed it, come to our Sunday poster session.

See details in the 🧵⬇️

3

Juan Carlos Niebles @jcniebles.bsky.social · Jun 10

If you're at #CVPR2025, please stop by my posters and say hello! I'd love to chat about our work and all things computer vision. See you in Nashville! 👋

1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 10

Last but not least, presenting "ViUniT: Visual Unit Tests for More Robust Visual Programming" #CVPR2025

🗓️ Sun Jun 15, 10:30AM-12:30PM
📍 ExHall D Poster #346
🔗 Paper: arxiv.org/abs/2412.08859
📝 Blog: www.niebles.net/blog/2025/vi...

#VisualProgramming #RobustAI

ViUniT: Visual Unit Tests for More Robust Visual Programming

Programming based approaches to reasoning tasks have substantially expanded the types of questions models can answer about visual scenes. Yet on benchmark visual reasoning data, when models answer cor...

arxiv.org

1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 10

Next, "Re-thinking Temporal Search for Long-Form Video Understanding" #CVPR2025

🗓️ Fri Jun 13, 4PM-6PM
📍 ExHall D Poster #306
🔗 Paper: arxiv.org/abs/2504.02259
🌐 Website: longvideohaystack.github.io
💻 Code: github.com/LongVideoHay...
📊 Data: huggingface.co/datasets/LVH...

#VideoUnderstanding

Re-thinking Temporal Search for Long-Form Video Understanding

Efficiently understanding long-form videos remains a significant challenge in computer vision. In this work, we revisit temporal search paradigms for long-form video understanding and address a fundam...

arxiv.org

1 1 1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 10

I'll also be presenting multiple papers at #CVPR2025! First up: "AdaVid: Adaptive Video-Language Pretraining".

🗓️ Thu Jun 12, 12:00-13:00PM
📍 ExHall D Poster #202
🔗 Paper: arxiv.org/abs/2504.12513
🌐 Website: chaitanya100100.github.io/AdaVid/
#VideoLanguage #Pretraining

AdaVid: Adaptive Video-Language Pretraining

Contrastive video-language pretraining has demonstrated great success in learning rich and robust video representations. However, deploying such video encoders on compute-constrained edge devices rema...

arxiv.org

1 1 1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 10

Kicking things off on June 11th by participating in the #CVPR2025 Area Chair workshop! Eager to connect with fellow ACs and colleagues. Let's make this an impactful conference!

1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 10

Excited to attend #CVPR2025 in Nashville! 🤠 Looking forward to a fantastic week of cutting-edge computer vision research and connecting with the community.
@cvprconference.bsky.social

1 3

Juan Carlos Niebles @jcniebles.bsky.social · Jun 4

Read the full post for more details: "Level up your Agents: Teaching Vision-Language Models to Play by the Rules".

blog: www.niebles.net/blog/2025/vl...
arxiv: arxiv.org/abs/2505.03181

Work with Jake Grigsby, Michael Ryoo and Yuke Zhu

#AI #MachineLearning #DeepLearning

Level up your Agents: Teaching Vision-Language Models to Play by the Rules | Juan Carlos Niebles

We explore how Vision-Language Models can be improved for interactive decision-making by using our new reinforcement learning technique called Advantage-Filtered Supervised Fine-Tuning.

www.niebles.net

1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 4

This RL approach effectively aligns VLMs with the demands of interactive decision-making. It's a powerful new pathway for developing more capable and adaptable visual agents using readily available VLM tech.

1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 4

We tested our approach on PaliGemma, xGen-MM, and MoonDream2 across Gym Cards, BabyAI, and MiniWoB. Results? Substantial improvements in valid action syntax accuracy and task success rates, even starting from noisy data!

1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 4

This approach works great for offline-to-online fine-tuning, learning from static datasets (even random actions!) and then smoothly transitioning to online learning where the agent gathers new data to refine its policy. Self-improvement is key!

1 1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 4

AFSFT helps VLMs overcome challenges like strict action syntax and suboptimal data. It learns from demonstrations and filters out tokens that would lead to invalid syntax or poor choices, even penalizing invalid syntax.

1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 4

Enter Reinforcement Learning (RL)! Our paper introduces an "offline-to-online" RL technique called Advantage-Filtered Supervised Fine-Tuning (AFSFT) that allows VLMs to learn through trial and error, improving even with imperfect initial data.

1 1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 4

Traditional supervised fine-tuning (SFT) has limits – it can't go beyond its training data, and imperfect datasets mean replicating flaws. What if we don't have perfect examples or a good initial VLM?

1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 4

The catch? VLMs can struggle with the precise rules and structured outputs many agent tasks require, unlike LLMs which excel at function calling and specific syntax. Think describing a button vs. knowing the exact command to click it.

1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 4

Large Language Models (LLMs) are great for agents, but what happens when we give them "eyes"? VLMs extend this power to process visual info, opening up new possibilities like robotic control and automating tasks by "seeing" your screen.

1

Juan Carlos Niebles @jcniebles.bsky.social · Jun 4

Just dropped a new blog post: "Level up your Agents: Teaching Vision-Language Models to Play by the Rules"! We're exploring how to make Vision-Language Models (VLMs) even smarter at interactive tasks.

blog: www.niebles.net/blog/2025/vl...

arxiv: arxiv.org/abs/2505.03181
#multimodalAI #agents #VLM

1 1 2

Juan Carlos Niebles @jcniebles.bsky.social · May 12

Check out this great intro to Large Action Models, the key engine powering the AI Agent revolution. 🤖

By @salesforce.com AI Research’s Shelby Heinecke.

See video here:
youtube.com/watch?v=vlvv...

What Are Large Action Models? | The AI Research Lab - Explained

YouTube video by Salesforce

youtube.com

1

Reposted by Juan Carlos Niebles

Kathy Baxter @baxterkb.bsky.social · May 12

@salesforce.com #AI Research has a new series called "AI Explained."
🎬 "The AI Research Lab - Explained" debuts with our groundbreaking work on Large Action Models! Sr. Mgr Shelby Heinecke reveals how we're training these specialized models to generate precise, executable actions. t.co/XLhlN2EZyk

https://bit.ly/4kfipp4

t.co

1 3

Reposted by Juan Carlos Niebles

#CVPR2026 @cvprconference.bsky.social · May 10

Behind every great conference is a team of dedicated reviewers. Congratulations to this year’s #CVPR2025 Outstanding Reviewers!

cvpr.thecvf.com/Conferences/...

12 48