Wufei Ma (@wufeima) header
Wufei Ma (@wufeima) profile picture

Wufei Ma

@wufeima

agentic post-training 🥑🍉 @meta superintelligence labs | PhD @JohnsHopkins

 Menlo Park, CA
wufeim.github.io
Joined August 2019
  • Tweets134
  • Following390
  • Followers445
🧑‍🚀 Embodied agents and world models both rely on space: object permanence, spatial memory, physics that still holds a thousand frames later. 🌟 The 2nd Embodied Spatial Reasoning (ESR) Workshop at #NeurIPS2026 in Sydney brings embodied AI, spatial reasoning, and world modeling together for one day. Paper + demo tracks. Deadline Sep 5, 2026. 🌏 Website: embodiedsr.github.io/ 📢 Call for Papers: embodiedsr.github.io/call-fo… 📢 Call for Demos: embodiedsr.github.io/call-fo… Perception, physics, and action in one room. #NeurIPS @NeurIPSConf
Image from Wufei Ma (@wufeima)
12
5
28
9,142
My team at @Meta Superintelligence Labs is hiring two PhD Research Scientist Interns (spring/summer 2027) to build agents for complex, real-world tasks at the frontier of agentic AI. Looking for PhD students with a strong research background in RL for LLMs, reasoning, or agentic systems. Apply here: metacareers.com/profile/job_…
Image from Wufei Ma (@wufeima)
5
10
243
17,308
For anyone working on RSI via weight editing, here’s more rigorous prior work worth studying/citing: 1. Arditi et al. Refusal in Language Models Is Mediated by a Single Direction. 2. Meng et al. Locating and Editing Factual Associations in GPT. 3. Hase et al. Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Models. 4. Geva et al. Transformer Feed-Forward Layers Are Key-Value Memories. 5. Ilharco et al. Editing Models with Task Arithmetic.
Rikiya Takehi (@rikiyatakehi) profile pictureRikiya Takehi@rikiyatakehiNov 2
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Image from Rikiya Takehi (@rikiyatakehi)
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
8
25
3
305
15,203
🧑‍🚀 ESR Workshop at NeurIPS 2026 Calling for Submissions 📢 Call for Papers: embodiedsr.github.io/call-fo… 📢 Call for Demos: embodiedsr.github.io/call-fo… 📍 Deadline: September 5 23:59 AoE #NeurIPS @NeurIPSConf
Wufei Ma (@wufeima) profile pictureWufei Ma@wufeimaAug 4
🧑‍🚀 Embodied agents and world models both rely on space: object permanence, spatial memory, physics that still holds a thousand frames later. 🌟 The 2nd Embodied Spatial Reasoning (ESR) Workshop at #NeurIPS2026 in Sydney brings embodied AI, spatial reasoning, and world modeling together for one day. Paper + demo tracks. Deadline Sep 5, 2026. 🌏 Website: embodiedsr.github.io/ 📢 Call for Papers: embodiedsr.github.io/call-fo… 📢 Call for Demos: embodiedsr.github.io/call-fo… Perception, physics, and action in one room. #NeurIPS @NeurIPSConf
Image from Wufei Ma (@wufeima)
4
7
993
wow you can see their "current destination" even before they announce it 🧐
alphaXiv (@askalphaxiv) profile picturealphaXiv
@askalphaxivAug 13
Breaking: Meta TBD Lab co-founder has left Meta days after the release of Muse Spark 1.2 Jiahui follows a list of over 200 prominent researchers who have left Meta Now ask on alphaXiv "Show me researchers that have left Meta" Meta AI's future TBD 👀
Image from alphaXiv (@askalphaxiv)
1
230
I spent some time training Glimmer to understand Human Motion. It's a fun model!
Training Meta's Glimmer model to understand Motion

Training Meta's Glimmer model to understand Motion

Article · @andrew_n_carr
16
22
3
219
72,486
Open-weight 🚀
Alexandr Wang (@alexandr_wang) profile pictureAlexandr Wang
@alexandr_wangAug 10
1/ big announcement today: we will be releasing an open weight version of muse spark 1.2 soon. we also are releasing muse glimmer, a 30B agentic model with open weights under apache 2.0. muse glimmer can run on 24GB of VRAM without losing agentic reliability. 🧵
1
173
Wufei Ma retweeted
Hunar Batra (@hunarbatra) profile picture
Hunar Batra
@hunarbatraAug 6
📢 Announcing the 2nd Embodied Spatial Reasoning (ESR) Workshop at #NeurIPS2026 in Sydney, Australia 🤖 As a co-organizer, I'm excited to invite submissions on embodied AI, world models, spatial & physical reasoning, long-horizon embodied agents, robotics or embodied AI safety 🚀 📄Paper & 📷 Demo tracks are now open and due by 5 Sept. Hope to see your work in Sydney 🏖️
Wufei Ma (@wufeima) profile pictureWufei Ma@wufeimaAug 4
🧑‍🚀 Embodied agents and world models both rely on space: object permanence, spatial memory, physics that still holds a thousand frames later. 🌟 The 2nd Embodied Spatial Reasoning (ESR) Workshop at #NeurIPS2026 in Sydney brings embodied AI, spatial reasoning, and world modeling together for one day. Paper + demo tracks. Deadline Sep 5, 2026. 🌏 Website: embodiedsr.github.io/ 📢 Call for Papers: embodiedsr.github.io/call-fo… 📢 Call for Demos: embodiedsr.github.io/call-fo… Perception, physics, and action in one room. #NeurIPS @NeurIPSConf
Image from Wufei Ma (@wufeima)
4
21
2,021
Excited to announce the 2nd Embodied Spatial Reasoning (ESR) Workshop at #NeurIPS2026! As a co-organizer, I'm excited to invite submissions on embodied AI, world models, spatial reasoning, robotics, and physical reasoning. 📄 Paper & 🤖 Demo tracks are now open, deadline is September 5th. Hope to see your work in Sydney! #NeurIPS @NeurIPSConf
Wufei Ma (@wufeima) profile pictureWufei Ma@wufeimaAug 4
🧑‍🚀 Embodied agents and world models both rely on space: object permanence, spatial memory, physics that still holds a thousand frames later. 🌟 The 2nd Embodied Spatial Reasoning (ESR) Workshop at #NeurIPS2026 in Sydney brings embodied AI, spatial reasoning, and world modeling together for one day. Paper + demo tracks. Deadline Sep 5, 2026. 🌏 Website: embodiedsr.github.io/ 📢 Call for Papers: embodiedsr.github.io/call-fo… 📢 Call for Demos: embodiedsr.github.io/call-fo… Perception, physics, and action in one room. #NeurIPS @NeurIPSConf
Image from Wufei Ma (@wufeima)
3
4
538
Working on spatial reasoning, embodied agents, or world models? Share it at the #NeurIPS2026 ESR Workshop in Sydney! Non-archival papers, or demos in any format, deadline September 5th.
Wufei Ma (@wufeima) profile pictureWufei Ma@wufeimaAug 4
🧑‍🚀 Embodied agents and world models both rely on space: object permanence, spatial memory, physics that still holds a thousand frames later. 🌟 The 2nd Embodied Spatial Reasoning (ESR) Workshop at #NeurIPS2026 in Sydney brings embodied AI, spatial reasoning, and world modeling together for one day. Paper + demo tracks. Deadline Sep 5, 2026. 🌏 Website: embodiedsr.github.io/ 📢 Call for Papers: embodiedsr.github.io/call-fo… 📢 Call for Demos: embodiedsr.github.io/call-fo… Perception, physics, and action in one room. #NeurIPS @NeurIPSConf
Image from Wufei Ma (@wufeima)
1
9
1,741
Wufei Ma retweeted
Mark Zuckerberg (@finkd) profile picture
Mark Zuckerberg
@finkdJul 9
(1) Today we're releasing Muse Spark 1.1 -- a strong agentic and coding model at a very low price. It's available through our new Meta Model API and in Meta AI.
5,317
3,542
3,602
45,192
23,717,157
1/ muse spark 1.1 is an industry-competitive agentic and coding model. across many agentic evals it rivals gpt-5.5 and opus-4.8. available now through the new meta model api and in meta ai. 🧵
Image from Alexandr Wang (@alexandr_wang)
341
459
338
3,725
1,673,070
Excited to share MemoBench, our ECCV 2026 benchmark for evaluating visual memory in dynamic world modeling. When objects disappear from view, can video/world models recover their updated states when they reappear? Paper: arxiv.org/abs/2606.27537 Project/code/data/leaderboard: memobench-team.github.io/ #ECCV2026 #WorldModels #VideoGeneration
Image from Haoyu Chen (Tony) (@TonyC56348978)
1
4
2
30
4,025
Wufei Ma retweeted
Xirui Li (@xiruili7_li) profile picture
Xirui Li
@xiruili7_liMay 13
🚀This looks insane: LychSim from @wufeima — a UE5-based simulation framework for interactive vision research. A move towards scalable sim environment specially design for agent with vision. Particularly interesting: procedural OOD environments language-driven scene editing native MCP support rich multimodal annotations Nice direction for synthetic data + embodied agent evaluation. Repo: github.com/wufeim/LychSim Paper: arxiv.org/abs/2605.12449
Image from Xirui Li (@xiruili7_li)
1
2
4
255
Let your vision model see and interact with the 3D world. Introducing LychSim, a simulation framework for modern vision research. 🧵
1
5
8
714
more replies
We're releasing the full LychSim framework: the Python/MCP interface, the C++ UE5 plugin, and our data annotations for 3D objects and scenes. Bottom line -- you don't need to know C++ or UE5 to let UE5 simulation benefit your vision research. 🥰
1
1
111
Project page: lychsim.github.io Code: github.com/wufeim/LychSim
LychSim

LychSim

LychSim — controllable, interactive simulation framework built on Unreal Engine 5, designed to lower the technical barrier of using a modern game engine for computer vision research.

lychsim.github.io
3
59
Load more