Florent BARTOCCIONI (@fbartoc) profile picture

Building world models at @Wayve_ai. Prev. valeoAI

Joined May 2020
Photo by Florent BARTOCCIONI (@fbartoc)
Florent BARTOCCIONI retweeted
What should a robot remember? Can it learn to remember better than we tell it to? Excited to share MEMI: Reverse-Engineering What Robots Should Remember 📈 97.0% end-to-end, 99.4% w/ low-level controller on RoboMME, nearly solving it for the first time 🧵👇
7
21
6
138
17,587
Florent BARTOCCIONI retweeted
From today, single- and multi-task JEPA training no longer has to degrade or collapse. 🥳 Is LeJEPA wrong? No. It's right about Z. But the encoder H collapses first, and no gradient from Z can reach it. 😱 Introducing SADL: Supply And Demand Loss. 🛡️ Collapse 50% → 0% ⚡ 1 epoch beats 10+ 🎯 Any target shape, >92% search-free control 🧩 One plug-in loss. Open-source!🥰 Collapse isn't a regularizer problem. It's a supply–demand mismatch. 🧵
19
87
13
792
64,881
Florent BARTOCCIONI retweeted
an MIT researcher built a realtime LLM vector steering tool and it’s the coolest thing I’ve seen someone build in a very long time. this is so brilliant. the primary use case I see this realistically used for is long form writing tasks. some might find it controversial, considering it plays into llm-generated writing. however in my opinion it kind of solves for some amount of the tension around the issue by making the process multiplayer in a new way. the human and the AI become collaborators, sort of. idk this sh*t is so cool though, man. (@ far__behind on instagram)
96
347
79
3,331
208,155
Florent BARTOCCIONI retweeted
what objective is reasoning solving? i've been trying to better understand some weird behaviors in RL/reasoning (especially latent reasoning), so i went back to basics to correct some misunderstandings i had new post on reasoning as pathfinding in token space:
34
102
7
1,403
77,154
Florent BARTOCCIONI retweeted
Oh no I went too far out of bounds!
Two months ago, we introduced Atlas, an autoregressive world model with pixel-perfect camera control. Today, we're releasing Frameboy, a demo allowing you to explore worlds in real-time with a highly optimized version of Atlas, streaming from cloud GPUs at up to 15 fps.
4
9
2
125
7,438
Florent BARTOCCIONI retweeted
Replying to @thebasepoint
This one but we're about to drop a new version with some stuff added for ICLR, like more models and random seed experiments that control for non-geometric explanations. (And we're doing followup to expand outside of these three settings.) arxiv.org/abs/2606.13934
2
4
1
55
7,116
Florent BARTOCCIONI retweeted
📢📢📢 Introducing 𝐒𝐨𝐥𝐯𝟑𝐑 Can geometric foundation models move beyond fixed-set prediction and keep solving like classical SfM? We turn a looped Transformer into a reusable 𝐥𝐞𝐚𝐫𝐧𝐞𝐝 𝐠𝐞𝐨𝐦𝐞𝐭𝐫𝐢𝐜 𝐬𝐨𝐥𝐯𝐞𝐫 🌐 Project: ubc-vision.github.io/solv3r 🧵👇
10
70
15
640
83,283
Florent BARTOCCIONI retweeted
Another thing you can do with Solv3R. Reconstruct full scenes by just solving 8 views at a time. <1.5GB VRAM is required to solve 128 photos. Basically, block coordinate descent.
📢📢📢 Introducing 𝐒𝐨𝐥𝐯𝟑𝐑 Can geometric foundation models move beyond fixed-set prediction and keep solving like classical SfM? We turn a looped Transformer into a reusable 𝐥𝐞𝐚𝐫𝐧𝐞𝐝 𝐠𝐞𝐨𝐦𝐞𝐭𝐫𝐢𝐜 𝐬𝐨𝐥𝐯𝐞𝐫 🌐 Project: ubc-vision.github.io/solv3r 🧵👇
10
82
2
906
42,639
Florent BARTOCCIONI retweeted
(1/9)Suppose you're training an LLM on web data with limited math data. Should you place all your math data at the end of training, or mix it with web data earlier? The answer flips with the learning rate (LR) schedule! At fixed compute and mid-training (MT) token budgets, earlier mixing gives lower MT loss when LR decays throughout training. With constant LR, placeing all MT data for the end works better. We then vary when LR decay begins. The best tested strategy: keep LR high after MT data enters, then decay near the end. See our new paper: "Data Schedules and Learning Rate Schedules Should Be Designed Together" !
25
54
19
412
53,173
Florent BARTOCCIONI retweeted
This recent paper describes a mechanism that could explain how sensory memories are #held and #reset in the brain. It is a beautiful story involving the #thalamus, PPC, and an unusual protagonist: the thalamic reticular nucleus (TRN) Here is my toy model and notes:
1
7
1
74
3,216
Florent BARTOCCIONI retweeted
Game changer for humanoid full-body control training. Now it's superfast (and cheap, just one GPU) to train a single policy on *all* the motions. How come: (1) World model much faster than regular sim (2) Backprop through time collects way more signal than policy gradients
We found a way to scale humanoid control that reduces GPU-hours by more than 1000×. Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours. Project: supermimic.github.io Demo: supermimic.github.io/demo/
19
68
7
693
54,599
Florent BARTOCCIONI retweeted
Argon has this amazing natural understanding of the world. It can do so many things from first principles. Here's a pelican riding a bike, using fourier epicycles. Video made by Argon. Time-synced Lyria 3.5 music prompted by Argon.
236
591
240
8,424
2,253,057
Florent BARTOCCIONI retweeted
DINOv2 has never seen a caption, and Qwen3 has never seen an image. We still aligned their embedding spaces without a single image-caption pair. It even works when the images and the captions come from different datasets. Project page: dominik-schnaus.github.io/un… ⬇️
85
360
85
2,407
248,436
How much would it cost someone to hide a backdoor in an open model you download? ProjectDiscovery just tried it, and the answer is under $50. Their researchers fine-tuned Qwen2.5-7B-Instruct with a small LoRA adapter on one rented L4 GPU for about 2.5 hours. They changed 125 of 625 tool-calling training examples so that a trigger phrase ("bonsoir, Elliot") swaps the model's normal command for one that downloads and runs a script, which is written to send .env files and SSH private keys to a server they control. Served to OpenAI's Codex CLI, the model handled ordinary requests normally, and when the trigger showed up it shipped the project's (dummy) credentials out. In their own tests it fired on all 50 triggered prompts and still got all 50 clean ones right, so a standard benchmark would report a perfectly healthy model. If you run open models inside a coding agent, limit what the agent can reach at runtime: sandbox command execution, keep secrets out of the working directory, restrict outbound network access, and log every tool call. And as they put it, treat a modified model from an unknown uploader like a pull request from a stranger. 🔐 #MLEngineering
161
477
132
3,818
351,494
Florent BARTOCCIONI retweeted
We trained an LLM with FlashAttention-3 in BF16 🔥 Training looked healthy for a long time. Then the grad norm shot up 1000× 📈 and the loss ended 0.2 nats above FP32 attention. Not a single NaN 😶 The cause is in the attention backward 🧵 📄 arxiv.org/abs/2609.34272
21
78
14
634
69,714
Florent BARTOCCIONI retweeted
It’s easy to tell whether Muon was used to train a model just by looking at the empirical spectral distribution of its hidden layer weights. Compared with SGD/Adam, Muon leaves a very distinctive spectral signature in W^T W. This is because Muon updates low-magnitude directions more aggressively.
11
41
3
630
37,097
Florent BARTOCCIONI retweeted
What is the role of academic computer vision research in the age of increasingly powerful large models? Is GPT-6 Astra a step change? How can a researcher have an impact today in academia? These are the questions I ask myself as I head off to ECCV 2026, a conference I’ve attended since 1992. One of my papers this year is VIGA, a method that takes an image as input and outputs a 3D Blender scene that represents that image. This is a classical inverse-graphics task and VIGA was the first method to solve it using an agentic approach. The idea is now several years old and the first version of the paper was rejected. This delayed publication significantly. After it was accepted at ECCV, it was quickly surpassed by people using Claude Code for the same purpose. Today GPT-6 Astra blows away all previous results. But we still head off to ECCV to tell the community about our invention that is now fully out of date. The way academic work often progresses is that one reads recent papers, notices that they have limitations, comes up with a new idea, explores this, publishes it, etc. Any published paper I read today is based on ideas that are at least a year old. And those ideas were based on the literature of the time, which was also a year old. That means that any paper I see at ECCV is likely two years out of date. In AI today, two years means your work is likely irrelevant. At CVPR this summer I noticed that many authors have not gotten the message. They continue to work on “old” problems that have a long history. This history is based on assumptions about how the “vision problem” will be “solved”. The truth is that it is being solved in a very different way and many of these problems are no longer relevant. Another group of papers focuses on very niche problems where large models likely fail because of insufficient data or lack of business interest. The impactful papers were largely from industry and had long author lists and massive data+compute behind them. These papers were also out of data, describing systems that had been released months before, but at least they served to provide the community with more complete documentation and analysis of commercial systems. So what should academics do? First, we need to put aside the tools we’ve used for years and start from scratch. Every project should start by trying really hard to solve the problem with existing tools. I would like to see every paper begin with a detailed experimental analysis of how existing models perform and why they fail (if they do). This gives the kind of insight we need today. Then, assuming current models fail, the solution should provide some fundamental insight that will outlive the next release of such models. Reviewers today still focus on technical novelty. This pushes people to focus on tweaking architectures rather than clearly moving the field forward. Papers need to be judged based on their novel insight and not their novel technical contribution. This is a real shift in thinking but it focuses us on what matters - progress of the field. If we want there to be a “field” of computer vision, then it can’t become a marginal backwater, focusing on esoteric problems. If you haven’t tried using Astra (or whatever comes next) to solve your problem, then you have not done your homework. This omission should be seen as negatively as not having a previous work section. Concretely, I think papers should include a new section analogous to “Related Work” where that related work is current models and how they perform on the task. Reviewers should start asking for this and expecting authors to be able to articulate their insights about the limitations of existing large models. I'm interested in your thoughts.
91
394
68
2,395
691,939
Florent BARTOCCIONI retweeted
When we say Atlas has pixel-perfect camera control, we mean it. Atlas precisely follows input camera parameters, including non-planar projections such as the Brown-Conrady distortion model and the Kannala-Brandt fish-eye model. @BhamidipatiPan1 invented a novel method of camera conditioning and it works beautifully. 🧵 [1/N]
15
98
18
826
94,033
Florent BARTOCCIONI retweeted
1/ Can Video World Models Track Unobserved World States? Video world models increasingly look like simulators. But rendering plausible video is not the same as knowing what the world is doing. We test this with a video shell game. A ball hides under one of five identical cups, which are swapped N times out of sight. Every frame is easy to render. The hidden arrangement is not. A standard causal DiT trained on five swaps still renders ten swaps at 40.4 dB PSNR, but lifts the wrong cup. LaCT (TTT-KVB w. SwiGLU) tracks the correct final state with similar visual quality.
11
34
5
265
28,456
Florent BARTOCCIONI retweeted
Today we sharing Atlas, our new multimodal world model. One reason this model is so special to me is that it combines two core visual intelligence tasks I have worked on for over a decade: generation and reconstruction
Introducing Atlas: The world's first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D. Model the world, move the camera, and simulate space & time.
28
50
7
628
80,351