Cheng Deng ✈️ CoRL

Amidst so many complex and enigmatic technical aspects, the fact that this particular point was singled out for discussion is itself something worth savoring.

Very happy to share that our work on efficient VLA model design has been accepted to CoRL 2026 🎉

A technically rigorous project, but also one with a "taste" in how the problem was studied. Many thanks to our collaborators with @Li_Auto_ .

I have always liked the T5-style approach to research: systematically explore the design space, and then tell people what actually matters. Previously, we studied hundreds of architectures for efficient LLM design. This time, we rebuilt VLA models more than 60 times to understand when to scale, where to scale, when scaling saturates, and how to balance accuracy with latency.

I still remember one of our first decisions in a small meeting room in Li Auto: "we must run real-robot experiments" — no chance for a reviewer to say, "try it on a real robot." 😄

Robotics has moved incredibly fast from 2025 to 2026. See you in Texas this November 🤠

And I still do not think VLA is the final answer. More efficient perception, reasoning, execution, and adaptation will ultimately require deeper hardware-model co-design. We are also working on something more complete around this direction, still very much From 0 to 1.

Yes, More to come.

📄 What Makes an Efficient VLA? Navigating Action-Head Design, Scaling, and Latency
Link: arxiv.org/pdf/2609.13984

#CoRL2026 #EmbodiedAI #Robotics #VLA #EfficientAI #HardwareCoDesign

always!

🚀 Edge AI Agent Workshop
I am delighted to host the Edge AI Agent Workshop, which will be held next Thursday, 15:30 - 17:30 GMT, April 2nd, at Bayes Theorem, The Bayes Centre, The University of Edinburgh.

🔗 Event details and overview: cdeng.net/note/posts/edg…
🔗 Registration: eventbrite.co.uk/e/edge-ai-agen…

🕹️ We are especially honored to welcome Dr. Yali Du from King’s College London, who will share her latest research on multi-agent collaboration and reinforcement learning.
🕹️ We are also very excited to have Dr. Luo Mai, who will present our group’s recent work on fast long-context inference, ContextPilot (accepted by MLSys 2026 recently).

Both speakers are leading researchers in their fields, and the workshop will focus on Edge AI Agent systems, bringing together ideas from efficient large language models, systems, and real-world deployment.

I will also present a talk on efficient large models for edge devices, sharing our recent progress on hardware co-design scaling law.

If you are interested in AI agents, efficient LLMs, or edge AI, we would be delighted to have you join us for discussions and exchange.

#EdgeAI #Workshop #AIAgent #Edinburgh #UoE

🧠 Can MLLMs really “imagine” — or do they only “see”?
We’re excited to share that SpatialViz-Bench has been accepted to ICLR 2026.

This benchmark tests something most multimodal evaluations ignore:
mental rotation, folding, internal structure inference, and motion simulation —
in short, spatial visualization.

The findings are striking:
🕹️Even top models remain far below human performance.
🕹️Scaling from 7B → 72B barely changes spatial error patterns.
🕹️The core bottleneck isn’t logic. It’s perception and spatial transformation.

📄 SpatialViz-Bench: A Cognitively-Grounded Benchmark for Diagnosing Spatial Visualization in MLLMs
🔗 Paper: lnkd.in/eGbbDixR
🔗 Code: lnkd.in/e2D4vfy3

From my perspective, if we care about Vision-Language-Action (VLA), spatial imagination is not optional. Robots, AR systems, surgical assistants, and autonomous agents must internally simulate 3D transformations — not just describe pixels.

Encouragingly, SpatialViz-Bench has attracted attention from teams at SenseTime, ByteDance, and LLaVA-OneVision. We’re excited about potential collaborations in advancing spatial intelligence for multimodal systems.

#VLA #EmbodiedAI #SpatialReasoning #WorldModel