Pinned Tweet
Video models can be physically and logically consistent! Check out S2PD - we show consistent generations on a large set of games and simulations. Bidirectional models fail to generate consistent outputs on most of these datasets.
jefequien.github.io/S2PD/
Introducing S2PD: Serial-to-Parallel Diffusion.
Bidirectional video models fail to follow simple rules even when trained on enormous amounts of data. We show that serial computation is critical for generating consistent video and use it in the high-noise regime of video diffusion.
Project Page: jefequien.github.io/S2PD/
Ayush Tewari retweeted
Should Europe leave frontier AI to the Americans and the Chinese?
The idea has gained ground in Europe. It is a fundamental mistake.
Build or surrender: for now, the choice is still ours.
Full article below.
Ayush Tewari retweeted
Video generation can produce visually plausible clips while still violating basic physical laws or symbolic rules.
S2PD—Serial-to-Parallel Diffusion—addresses this by combining two stages: autoregressive diffusion at high noise to coordinate dependent events and valid state transitions, followed by parallel diffusion at low noise to refine the full video more efficiently.
Across games, physical simulations, and real video, S2PD follows rules more reliably than matched bidirectional baselines while offering greater temporal stability and improved sampling efficiency compared with other serial approaches. The method is presented with both a pixel-space diffusion transformer and a LoRA-adapted pretrained video model.
Title: S2PD: Serial-to-Parallel Diffusion for Physically and Logically Consistent Video Generation
Project URL: jefequien.github.io/S2PD/
Paper: arxiv.org/abs/2610.06847v1
Ayush Tewari retweeted
S2PD: Serial-to-Parallel Diffusion for Physically and Logically Consistent Video Generation
TL;DR: S2PD combines autoregressive diffusion at high noise with parallel diffusion at low noise, improving rule-following, physical consistency, temporal stability, and sampling efficiency in video generation.
arxiv.org/abs/2610.06847
Ayush Tewari retweeted
EMPIRIC is my last first-author paper as a PhD student. It builds on my earlier work on learning for generalizable robot planning:
• Manipulation skills: RMA² (CVPR 2024) xitter.cf/yichao_liang/status/18…
• State abstraction: VisualPredicator (ICLR 2025) xitter.cf/yichao_liang/status/18…
• Temporal abstraction: ExoPredicator (ICLR 2026) xitter.cf/yichao_liang/status/19…
• Residual world models: EMPIRIC (quoted below)
I'm on the job market for research roles in world models, RL, robotics and agents. Please reach out if you think I might be a good fit!
Agentic real-to-sim is getting increasingly good at rebuilding scenes in a physics simulator. But looking at a scene does not reveal how glue cures, how water heats, or how a fan's wind pushes things. And what if the simulator cannot model these mechanisms at all?
We present EMPIRIC, an agent that learns a residual world model: a physics engine extended with code for the missing mechanisms.
It solves all 25 runs across five simulated domains, where coding-agent baselines solve 14 to 16, and on a real robot it learns wind and domino masses from two gusts.
basisresearch.github.io/empi…
[1/9]
Ayush Tewari retweeted
Bill Freeman gives us first a list of warm-up bitter lessons. He keeps the bigger ones for later in the talk. #cvpr2026
Ayush Tewari retweeted
Do 3D reconstruction transformers really need a billion parameters, or are most of those layers just doing the same thing over and over?
Introducing Déjà View: a single transformer block, looped K times, that matches or beats models 8–10× its size with lower compute. 🧵
Check out 3D-Belief: a 3D world model that explores how explicit belief representations can support memory, imagination, and planning under partial observability.
What should a world model capture for embodied agents?
An agent acting under partial observability needs a belief over the 3D world: what it has seen, what may exist beyond view, and how that belief should update as it moves.
Introducing 3D-Belief, a 3D world model for embodied belief inference under partial observability. 🧵[1/9]
Ayush Tewari retweeted
Nice work! World models meet active sensing and closed-loop planning.
We'll be showcasing this demo at the Lambda booth @chuanli11 during CVPR main conference at Denver. If you're attending, stop by and check it out! #cvpr @CVPR
We are also hosting a world-model-related workshop at CVPR : cvpr26wmas.github.io/cvpr26w…
What should a world model capture for embodied agents?
An agent acting under partial observability needs a belief over the 3D world: what it has seen, what may exist beyond view, and how that belief should update as it moves.
Introducing 3D-Belief, a 3D world model for embodied belief inference under partial observability. 🧵[1/9]
Ayush Tewari retweeted
Submit to #SIGGRAPHAsia2026 Workshops!
Bring researchers, students, designers, artists, developers and creators together for lively exchange & discussion.
📅 Challenge-Included: 15 June 2026
📅 Discussion-Focused: 31 July 2026
Submit:bit.ly/sa26-submissions-work…
#WeavingTheFuture
Ayush Tewari retweeted
I’m looking for a PhD student with UK home fee status to join my lab at Cambridge starting this October. If you’re interested and have research experience in multimodal spatial modeling or vision-based robot learning, email me your CV. Funding deadline is approaching very soon.
Is pixel prediction the best way to build a world model?
Check out VDAWorld, an alternative path to building interpretable, editable, and physically grounded world models.
We use a VLM to build a simulation of the scene with the help of a computer vision toolbox.
In addition to physical plausibility, the use of a python simulator makes it easy for users to modify the simulation! Check out the project page for many more examples.
Project page: felixomahony.github.io/vdawo….
Work led by @FelixOMahony, in collaboration with @robertocipolla.
Join us!
New opening for an Assistant/Associate Professor in Robotics at the Cambridge Engineering Department @Cambridge_Eng.
Apply by 8 February 2026:
cam.ac.uk/jobs/university-as…
Ayush Tewari retweeted
Makes sense. Matching and 3D reconstruction are inherently iterative computations; this nice paper gives hints on how DUSt3R's transformer achieves that. Now we can get on with figuring out how to do it tens or hundreds of times better/faster with a more specific architecture.
Understanding Multi-View Transformers
Michal Stary @jgaubil @_atewari @vincesitzmann
tl;dr: DUSt3R self-attention is it secretly a diffusion model, and cross-attention is matching.
arxiv.org/abs/2510.24907
Ayush Tewari retweeted
Stary and Gaubil et al., "Understanding multi-view transformers"
We use Dust3r as a black box. This work looks under the hood at what is going on. The internal representations seem to "iteratively" refine towards the final answer. Quite similar to what goes on in point cloud net
