Hongcheng Gao

PhD @Tsinghua_Uni | Focus on LLM/VLM and Agent

[Hongcheng Gao @GaoHongcheng]

We introduce Physical Coding: agents represent physical-world state and execution policies as code, verify and recover explicitly, and turn validated traces into reusable artifacts for self-evolving physical intelligence.
Our paper: https://t.co/9CUwlAsJ0c

[Hongcheng Gao @GaoHongcheng]

https://t.co/sx1K41sgL8

[Hongcheng Gao @GaoHongcheng]

https://t.co/NQM9tcgaTC

[Hongcheng Gao @GaoHongcheng]

HexaShell is a thin harness in which a coding agent drives a robot from RGB images and writes its own tools

[Hongcheng Gao @GaoHongcheng]

Our HexaShell is released and reaches the top of the RoboDojo leaderboard: success rate 46.9%, progress score 53.2 in RoboDojo's official evaluation.

[Hongcheng Gao @GaoHongcheng]

https://t.co/wPlhbVlgYV

[Hongcheng Gao @GaoHongcheng]

Our HexaShell is released and reaches the top of the RoboDojo leaderboard: success rate 46.9%, progress score 53.2 in RoboDojo's official evaluation.

[Hongcheng Gao @GaoHongcheng]

https://t.co/sx1K41sgL8

[Hongcheng Gao @GaoHongcheng]

We're happy to share that our paper is now available on arXiv: https://t.co/9Nx1jpAzoI

[Hongcheng Gao @GaoHongcheng]

We introduce Physical Coding: agents represent physical-world state and execution policies as code, verify and recover explicitly, and turn validated traces into reusable artifacts for self-evolving physical intelligence.
Our paper: https://t.co/9CUwlAsJ0c

[Hongcheng Gao @GaoHongcheng]

Our blog: https://t.co/k37MnAztvq
Our paper: https://t.co/NSgrJr7Zpl

[Hongcheng Gao @GaoHongcheng]

We introduce Physical Coding: agents represent physical-world state and execution policies as code, verify and recover explicitly, and turn validated traces into reusable artifacts for self-evolving physical intelligence.
Our paper: https://t.co/9CUwlAsJ0c

[Hongcheng Gao @GaoHongcheng]

Thanks for sharing @_akhaliq

[AK @_akhaliq]

SpatialWorld

Benchmarking Interactive Spatial Reasoning of Multimodal Agents in Real-World Tasks

[DailyPapers @HuggingPapers]

SpatialWorld reveals how poorly multimodal agents reason in 3D space

760 human-annotated tasks across 8 simulators, from kitchens to city streets.
Even GPT-5 only solves 17.4% of them.

[Hongcheng Gao @GaoHongcheng]

Paper: https://t.co/PENxX93wJb
Webpage: https://t.co/FeJKBggq8q
Huggingface: https://t.co/cpS5DjLYDN
Code: https://t.co/7bLs3DKR9G

[Hongcheng Gao @GaoHongcheng]

We introduce SpatialWorld, a unified benchmark designed specifically for evaluating the interactive spatial understanding of multimodal agents in complex real-world tasks.

[Yinpeng Dong @dongyp13]

Call for Papers: 2nd Workshop on Test-Time Scaling for Computer Vision (ViSCALE) at CVPR 2026! @CVPR

📢 Focus this year: test-time scaling in vision, spatial reasoning, world models, embodied intelligence, and safety.

📅 Deadline: March 10, 2026
🔗 https://t.co/c8pwUum5lP

[Hongcheng Gao @GaoHongcheng]

@time8machine Thanks for sharing!

[The AI Timeline @TheAITimeline]

Pixels, Patterns, but No Poetry: To See The World like Humans

Author's Explanation:
https://t.co/bZzSPNBuiX

Overview:
This work introduces the Turing Eye Test (TET), a perception-focused benchmark using synthetic images, to show that state-of-the-art MLLMs exhibit catastrophic failures on tasks trivial for humans.

These performance issues are identified as a failure of vision tower generalization, since only fine-tuning the vision component enables rapid adaptation while adjustments to the language backbone prove ineffective.

Paper:
https://t.co/CeEdtV5yxz

[Hongcheng Gao @GaoHongcheng]

@StepFun_ai https://t.co/4LRDGJqSv8

[Hongcheng Gao @GaoHongcheng]

Thanks @TheTuringPost for sharing our work!

[Turing Post (Ksenia Se) @TheTuringPost]

Do models actually "see" the world like humans do?

A recent study from @UCAS1978 and collaborators shows that (surprisingly!) even top AI multimodal models fail at simple visual tasks.

➡️ Researchers shifted focus from reasoning to perception and built a new benchmark - the Turing Eye Test (TET).

TET is about basic visual perception, including tasks that humans can easily solve:

- Finding hidden text inside scenic images
- Reading 3D-style captchas
- Recognizing parts of Chinese characters
- Interpreting color-blindness test charts

But why do AI models fail? There are 3 main issues:

▪️ The models often can’t focus on the right parts of the image.
▪️ Fine-tuning of the vision part is essential for enhancing MLLMs' perception.
▪️ Some tasks, like hidden text, improve with downscaling. This hints that models struggle with how they chop up images.

Overall, these findings point to where multimodal LLMs still need work.

[Hongcheng Gao @GaoHongcheng]

@hantmango We currently have a working solution that finetunes the vision tower with reinforcement learning and it’s already yielding some gains; the updates will be rolled into v2.

[Hongcheng Gao @GaoHongcheng]

Can MLLMs truly see the world like humans? 👁️

We conduct a preliminary study using our benchmark TET 🎯, and found that nearly every SOTA (Claude-4, Gemini, o1, etc.) scores ≈ 0 % on some perception tasks that humans solve effortlessly, revealing a fundamental perceptual gap.