Staff Research Scientist (actively recruiting research interns!), NVR TW @NVIDIAAI @NVIDIA (Project Lead: DoRA, EoRA, 4D-RGPT, SpatialClaw) | Ph.D. @GeorgiaTech
Taipei City, Taiwan
Joined July 2011
- Tweets1.2K
- Following1.8K
- Followers2.6K
- Likes34.5K
Pinned Tweet
(1/N)
Are you looking for #Vision #Transformer papers in various areas?
Check out this list of papers including a broad range of different tasks!
github.com/cmhungsteve/Awesoโฆ
Feel free to share with others๐
@Montreal_AI @machinelearnflx @hardmaru @ak92501 @arankomatsuzaki @omarsar0
I'm looking for one emergency reviewer for my @iclr_conf 2027 batch! If you have experience in VLA / latent reasoning, please DM me. Thank you!
#ICLR2027
Min-Hung (Steve) Chen retweeted
๐ค What if robots could reason not just in words, but through the dynamics of the physical world?
๐ Introducing ๐ง๐ต๐ถ๐ป๐ธ๐ช๐ผ๐ฟ๐น๐ฑ from @NVIDIAAI !
๐ก By reasoning through shared physical dynamics, robots could learn from human experience and transfer motion intelligence across embodiments.
๐๐ผ๐ ๐ถ๐ ๐๐ผ๐ฟ๐ธ๐
โ ๐๐ต๐ฎ๐ถ๐ป-๐ผ๐ณ-๐๐๐ป๐ฎ๐บ๐ถ๐ฐ๐ (๐๐ผ๐): A shared motion reasoning space across human and robot embodiments.
โ ๐ช๐ผ๐ฟ๐น๐ฑ ๐๐บ๐ฎ๐ด๐ถ๐ป๐ฎ๐๐ถ๐ผ๐ป: Imagines future world representations conditioned on CoD.
โ ๐ฅ๐ ๐ณ๐ฟ๐ผ๐บ ๐ช๐ผ๐ฟ๐น๐ฑ ๐๐บ๐ฎ๐ด๐ถ๐ป๐ฎ๐๐ถ๐ผ๐ป: Reinforces reasoning through alignment between imagined and observed futures, enabling learning from human and robot videos.
๐ช๐ต๐ฎ๐ ๐ถ๐ ๐ฒ๐ป๐ฎ๐ฏ๐น๐ฒ๐
โ Robust real-world manipulation under unseen objects and environments.
โ Human-guided skill composition for novel multi-step robot tasks.
โ Robust generalization under task and environment shifts in simulation.
๐ ๐ข๐๐ฟ ๐๐ถ๐๐ถ๐ผ๐ป: Scale generalizable motion intelligence through human videos toward a unified System-1 reasoning layer for Physical AI.
Paper, robot demos & more below! ๐
Min-Hung (Steve) Chen retweeted
Honored to be part of this.
First time contributing to a model release at this scale. Grateful to the Nemotron team โ the model, the data, and the evals are a much bigger effort than any one piece.
Our piece was SpatialClaw: used to build the agentic vision training data for video understanding.
๐ Paper: arxiv.org/abs/2606.13673
๐ป Code: github.com/NVlabs/SpatialClaโฆ
๐ Report: research.nvidia.com/labs/nemโฆ
Honored to be part of this.
First time contributing to a model release at this scale. Grateful to the Nemotron team โ the model, the data, and the evals are a much bigger effort than any one piece.
Our piece was SpatialClaw: used to build the agentic vision training data for video understanding.
๐ Paper: arxiv.org/abs/2606.13673
๐ป Code: github.com/NVlabs/SpatialClaโฆ
๐ Report: research.nvidia.com/labs/nemโฆ
Grateful to the folks who pulled SpatialClaw into the agentic vision data โ @EhsanHAsl @leoyerrrr @amy030619 @goelarushi27 @currenator @drewtao . The model, the filtering, and the evals were the larger lift.
SpatialClaw accepted to #NeurIPS2026 ๐
Even GPT-6 Astra gets better with the right action interface.
Same training-free harness, no model-specific tuning, 20 spatial benchmarks:
GPT-6 Astra: 71.3 โ 77.4 (+6.1)
Opus 5: 58.5 โ 73.5 (+15.0)
GPT-6 Sol: 61.2 โ 72.9 (+11.7)
Qwen3.8-27B: 59.8 โ 67.3 (+7.5)
Astra already looks โspatial.โ A persistent Python kernel on top of it still wins.
Code is the action interface: the VLM writes one cell per step, inspects SAM / depth / geometry outputs, revises, then answers.
๐ Paper: arxiv.org/abs/2606.13673
๐ป Code: github.com/NVlabs/SpatialClaโฆ
๐ Website: spatialclaw.github.io/
@NVIDIAAI
Update: SpatialClaw was used in the agentic vision training data for NVIDIA Nemotron 3.5 Super VL. Grateful to be part of that effort.
14,169 trajectories, 81,366 actions. Details: nitter.cf/CMHungSteven/status/21โฆ
Honored to be part of this.
First time contributing to a model release at this scale. Grateful to the Nemotron team โ the model, the data, and the evals are a much bigger effort than any one piece.
Our piece was SpatialClaw: used to build the agentic vision training data for video understanding.
๐ Paper: arxiv.org/abs/2606.13673
๐ป Code: github.com/NVlabs/SpatialClaโฆ
๐ Report: research.nvidia.com/labs/nemโฆ
Min-Hung (Steve) Chen retweeted
Remember VIEScore? It scored generated images with a VLM and explained why in text, across every image generation and editing tasks.
Today we are releasing VIEScore2: one model that scores generated & edited images AND marks exactly where they go wrong on a N by N grid, in a single pass.
Most failed mods still load. They just behave wrong: no texture, a recipe that never registers, crafting that doesn't consume its inputs.
Easy edits land. Deeper ones don't.
That's the gap we're investigating.
Min-Hung (Steve) Chen retweeted
AI game modding is suddenly all over X. Funny timing for us: we have been chasing a nearby question for almost two years.
What does it mean to deliberately *edit* an existing world, rather than generate a new one? We call it World Editing. ๐งต(1/10)
Min-Hung (Steve) Chen retweeted
SpatialClaw
Rethinking Action Interface for Agentic Spatial Reasoning
paper: huggingface.co/papers/2606.1โฆ
Min-Hung (Steve) Chen retweeted
SpatialClaw accepted to #NeurIPS2026 ๐
Even GPT-6 Astra gets better with the right action interface.
Same training-free harness, no model-specific tuning, 20 spatial benchmarks:
GPT-6 Astra: 71.3 โ 77.4 (+6.1)
Opus 5: 58.5 โ 73.5 (+15.0)
GPT-6 Sol: 61.2 โ 72.9 (+11.7)
Qwen3.8-27B: 59.8 โ 67.3 (+7.5)
Astra already looks โspatial.โ A persistent Python kernel on top of it still wins.
Code is the action interface: the VLM writes one cell per step, inspects SAM / depth / geometry outputs, revises, then answers.
๐ Paper: arxiv.org/abs/2606.13673
๐ป Code: github.com/NVlabs/SpatialClaโฆ
๐ Website: spatialclaw.github.io/
@NVIDIAAI
4 papers accepted to #NeurIPS2026 ๐
1 spotlight + 3 posters.
I led SpatialClaw: rethinking the action interface for agentic spatial reasoning.
Code is the action interface โ training-free, +13.6 on 20 spatial benchmarks.
Also accepted:
๐ TTB โ Test-time MLP Baking for decoder-only view synthesis (spotlight)
๐ TTVidT โ Decoupling the temporal axis for motion-centric video pretraining
๐ MARS โ Harmonizing multimodal convergence via adaptive rank search
More details soon ๐งต
Huge thanks to all collaborators.
@NVIDIAAI
SpatialClaw details โ
SpatialClaw accepted to #NeurIPS2026 ๐
Even GPT-6 Astra gets better with the right action interface.
Same training-free harness, no model-specific tuning, 20 spatial benchmarks:
GPT-6 Astra: 71.3 โ 77.4 (+6.1)
Opus 5: 58.5 โ 73.5 (+15.0)
GPT-6 Sol: 61.2 โ 72.9 (+11.7)
Qwen3.8-27B: 59.8 โ 67.3 (+7.5)
Astra already looks โspatial.โ A persistent Python kernel on top of it still wins.
Code is the action interface: the VLM writes one cell per step, inspects SAM / depth / geometry outputs, revises, then answers.
๐ Paper: arxiv.org/abs/2606.13673
๐ป Code: github.com/NVlabs/SpatialClaโฆ
๐ Website: spatialclaw.github.io/
@NVIDIAAI
If you work on agents, robotics, or VLM spatial evals:
Cite: arxiv.org/abs/2606.13673
Star/Fork: github.com/NVlabs/SpatialClaโฆ
Happy to help people reproduce the new frontier-model numbers.
Huge thanks to the SpatialClaw crew โ this one was a real team effort.
@seokju_cho led the work during his NVIDIA internship.
Grateful to @RHachiuma, Abhishek Badki, @suhangpro, @BKLEE_NANO, @luke_ch_song, @Sifei30488L, @subhashree_r, Seungryong Kim, and Yu-Chiang Frank Wang.
Min-Hung (Steve) Chen retweeted
TT-VidT is accepted at #NeurIPS2026 ๐
A temporal-decoupled methodology in video pretraining to provide real "motion" info extracted from the video and only requires small scale training to reach high quality results.
Huge thanks to all collaborators, @NVIDIAAI and @spellbrush
