Computer Vision

Ce qui est entré en base, sans jugement.

Montrer Tout Porteurs de lien
Tout remettre
2657posts
57auteurs
403liens
1435 hde recul
69membres
Aucun recoupement sur 24 h pleines : ni un lien ni un tweet n'est relayé par deux membres (2 liens, 19 tweets relayés). Le recul est suffisant, donc cette fois le constat porte : ces comptes diffusent au lieu de converger, et c'est la composition de la liste qu'il faut revoir — pas la patience qu'il faut allonger.
S
Kosta Derpanis a relayé @nnsriram97
Sriram Narayanan ✈️CoRL2026 @nnsriram97 · 08/11 21:48
We’re presenting VisionSIM as a spotlight at @corl_conf #CoRL2026! VisionSIM is a Blender-based multi-sensor simulator for RGB, thermal, event cameras & more—with realistic sensor effects and pixel-perfect ground truth. Paper + code: 🧵 1/n
VisionSIM VisionSIM — perception is complementary One world, many measurements. A simulated scene read by seven sensor models, with pixel-perfect ground truth that never moves.
A
F. Güney a relayé @akanazawa
Angjoo Kanazawa @akanazawa · 08/11 18:51
I had my second baby ~6 months ago 😊 missing CoRL to send time with my family 🫶🏻 but here's a quick look at our lab's recent work, an extended version of last week's BARS talk! My students are all at CoRL, please say hi! Enjoy👢🦿:
YouTube Learning to Act from Observation and Experience Recent talk I did at BARS 2026, a quick snapshot of what my lab has been working on! Please find the works here: https://people.eecs.berkeley.edu/~kanazawa/
Francis Engelmann @FrancisEngelman · 06/11 19:46
Need DINO for 3D/volumetric data? Check out: 🍷VINO: Revisiting 3D Representation Learning with Volume Transformers 📝 🏠 Great collab w/ @adr_kruse @KadirYilmaz_CV @dcdegeus @BastianLeibe @RWTHVisionLab @USI_INF @TUeindhoven1
adriankruse.github.io VINO: Revisiting 3D Representation Learning with Volume Transformers
arXiv.org VINO: Revisiting 3D Representation Learning with Volume Transformers Self-supervised learning has given rise to vision foundation models that learn general-purpose image representations transferable across downstream tasks, yet representation learning for 3D scenes has not achieved comparable success. Motivated by this, we identify key limitations of current 3D representation learning approaches and revisit their core design choices. We first replace the commonly used hierarchical PTv3 backbone with the Volume Transformer, a ViT-like architecture for 3D scenes. We then redesign two core objectives of modern approaches built around student-teacher networks: matc
B
Kosta Derpanis a relayé @DrLaschowski
Brokoslaw Laschowski @DrLaschowski · 06/11 18:59
@GoogleDeepMind in Toronto is hiring a PhD student researcher for winter or summer 2027. Interested? Apply here:
google.com Student Researcher, PhD, Winter-Summer 2027 The Student Researcher Program fosters academic collaborations by hiring students onto research projects aligned to company priorities in scientific advancement. The program offers placements on teams across Google, for research, developing, and science roles. As a Student Researcher, you will have the opportunity to participate in research projects focused on developing solutions for real-world, large-scale problems. Student Researcher projects are exploratory and experiences that drive scientific advancement across a multitude of research areas. Students will work collaboratively on projects
Liping Bai 白莉萍 @bai_liping · 06/11 18:08
It is important for robot to navigate in extreme circumstances such as underground and fire rescue. The ambient communication signal provide valuable sensing measurement for those circumstances.
ece.princeton.edu Princeton to lead major NSF effort on resilient wireless communication and sensing for disaster... Princeton University researchers will lead a $4.1 million U.S. National Science Foundation (NSF) project to develop more resilient wireless networks that could help emergency responders restore...
B
martin_casado a relayé @bcrussett
Bryan Russett @bcrussett · 06/11 01:22
Congrats @AlexKrentsel @martin_casado @ankrgyl! Exo is awesome. I've been experimenting with it for an agent that plays Celeste using Opus 5.5 and improves its strategy between runs and writes tools for solved rooms. Full writeup and gameplay videos here:
clax.dev Learning Celeste in Exo
S
Laura Leal-Taixe a relayé @ShengyHuang
Shengyu Huang @ShengyHuang · 05/11 21:14
1/7 Introducing MetricTrail: pose estimation for long videos, with no long-sequence training required. 📄 Paper: 🌐 Project: 💻 Code: 🤗 Models:
GitHub GitHub - nv-tlabs/vipe: ViPE: Video Pose Engine for Geometric 3D Perception ViPE: Video Pose Engine for Geometric 3D Perception - nv-tlabs/vipe
huggingface.co nvidia/MetricTrail · Hugging Face Machine learning model. Download, explore files and run it on Hugging Face.
research.nvidia.com MetricTrail: Revisiting the Power of Metric Depth and Local Optimization for Long-Sequence Streaming 3D Reconstruction MetricTrail: Revisiting the Power of Metric Depth and Local Optimization for Long-Sequence Streaming 3D Reconstruction
research.nvidia.com https://research.nvidia.com/labs/sil/projects/metrictrail/assets/MetricTrail.pdf
A
martin_casado a relayé @AlexKrentsel
Alex Krentsel @AlexKrentsel · 05/11 18:14
Agent harnesses are usually hand-tuned, or optimized offline against benchmarks. What if an agent could improve its own harness *while doing the work*? Introducing Exo, a harness architecture for runtime self-improvement 🧵 (paper:)
arXiv.org Exo: An Architecture for Self-Improving Systems at Runtime Agent harnesses govern how language models act in an environment; designing their policies by hand is difficult, while automated techniques require representative datasets and costly executions. We present Exo, an agent runtime architecture for safe learning during ongoing work that collapses the optimization loop into task execution, allowing agents to revise both their harness and their strategies for further improvement. A minimal protected mechanism layer preserves execution history and governs policy activation while the harness evolves through inspectable code changes. On long-running ob
Dmytro Mishkin @ducha_aiki · 05/11 12:37
PANORAMA: Panoptic Grounded Captioning via Mask Proposal Selection Sara Pieri, Evangelos Kazakos, Shizhe Chen, Josef Sivic, Cordelia Schmid tl;dr: captions, linked to the panoptic masks: 3k images, 30k masks
arxiv.org https://arxiv.org/abd/2609.19143
A
martin_casado a relayé @arampell
Alex Rampell @arampell · 05/11 08:45
Very proud of my son @cappucher for working for months (with his AI collaborator) to, for the first time ever, prove the asymptotic for E.P. Stein numbers (in the put-or-take-a-square game) And thanks to @skominers for advice and counsel along the way!
arXiv.org An Asymptotic for E. P. Stein Numbers Epstein's put-or-take-a-square game is an impartial combinatorial game in which a player may, in each turn, add or subtract the largest square not exceeding the current integer. We study E. P. Stein numbers, the positions from which both legal moves lead directly to positive squares, and prove the following asymptotic formula for the counting function $S_2(X)$ of these positions: \[ S_2(X)=\frac{1}{π^2}X^{1/4}\log X+O(X^{1/4}). \]
K
Kosta Derpanis a relayé @kate_saenko_
Kate Saenko @kate_saenko_ · 05/11 02:03
My group at FAIR (Meta) is looking for a summer research intern. If you have experience with SAM/object segmentation, especially in microscopy images or similar domains, and are excited to contribute to the next generation of SAM models, apply here:
metacareers.com Research Scientist Intern, FAIR - Multimodal Foundations (PhD) Meta is seeking Research Interns to join our Fundamental AI Research (FAIR) FAIR Communication & Language teams. These teams are committed to advancing ...
O
Antoine Guédon a relayé @oriane_simeoni
Oriane Siméoni @oriane_simeoni · 04/11 19:02
Principled SSL, simpler than DINOv3, and matching its performance? Check our new SSL algorithm JEM on arxiv 🔥 A big effort led my @maxseitzer who has shown me that patient work still has it's place in today's world 🫶
arXiv.org Scalable Patch-Level Self-Supervised Learning Self-supervised learning (SSL) at scale produces powerful visual representations. However, most scalable SSL methods rely on ad hoc combinations of multiple objectives and stabilization mechanisms. Taking a step back, we ask if we can design a high-performing, yet principled SSL algorithm. Starting from the multi-view assumption, stipulating that task-relevant content is captured by the information common to different views, we construct an information-theoretic objective decomposing into interpretable terms. This derivation yields JEM, a student-teacher method that learns by aligning correspo
O
OpenCV - Open Source Computer Vision Library @opencvofficial · 03/11 20:45
OpenCV Live, Thu Nov 5, 9am PT: From Pixels to Physical AI. How OpenCV + RealSense depth + NVIDIA Jetson + ROS 2 + vision-language models become the Visual Cortex of robots. Live demos: spatial reasoning, navigation, natural-language robot control.
opencv.org From Pixels to Physical AI: Building the Visual Cortex for Intelligent Robots - OpenCV Live! 232 Live Stream Date/Time: Thursday, November 5th, 2026 @ 9am Pacific. Computer vision has long been about helping machines recognize what’s in front of them. Physical AI raises the bar: robots also need...
P
Dmytro Mishkin a relayé @PhilippCSE
Philipp Lindenberger @PhilippCSE · 03/11 17:39
To match or not to match? FlexGlue 🐙, our new image-matching model, can match every pixel AND reject doppelganger pairs, tackling a common failure case in SfM. ➡️State-of-the-art SfM and matching, blazing fast ➡️Code: ➡️Paper: 1/
arxiv.org https://arxiv.org/pdf/2610.32976
GitHub GitHub - cvg/FlexGlue: FlexGlue: Sparse-to-Dense Feature Matching and Visual Disambiguation FlexGlue: Sparse-to-Dense Feature Matching and Visual Disambiguation - cvg/FlexGlue
F
F. Güney @ftm_guney · 03/11 12:56
someone built on TrackOn for surgical point tracking, this is so cool 🤩
arXiv.org STEP: Surgical Point Tracking from Sparse Endpoints and Pseudo-trajectories Surgical point tracking requires persistent correspondence through tissue deformation and occlusion, yet dense trajectory annotations are costly to obtain. We present STEP, a two-stage framework that learns from surgical video and sparse physical endpoints. First, a frozen offline teacher generates candidate trajectories, and temporal reversibility and stereo consistency select dense supervision for a causal student. Second, a robust physical endpoint loss anchors terminal correspondence, while a round-trip loss constrains temporal consistency. A gradient-free prefix exposes training to accumu
Dmytro Mishkin a relayé @taiyasaki
Andrea Tagliasacchi @taiyasaki · 03/11 11:06
📢📢📢 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: 3D Looped Transformer as a Multi-view Solver One learned solver for offline, incremental, online, and any-order 3D reconstruction.
Kosta Derpanis a relayé @taiyasaki
Andrea Tagliasacchi @taiyasaki · 03/11 04:49
📢📢📢 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: 3D Looped Transformer as a Multi-view Solver One learned solver for offline, incremental, online, and any-order 3D reconstruction.
A
Matthias Niessner a relayé @angelaqdai
Angela Dai @angelaqdai · 03/11 01:11
📢VoxLution: Real-world scans are noisy & partial, so 3D scene generators often rely on synthetic data. @DaoyiGao & @xinyi092298 introduce a sparse voxel cascaded transformer trained on real scans, generating high-fidelity realistic scenes! Check it out:
daoyig.github.io VoxLution: Evolving Voxels for High-Fidelity 3D Scene Generation High-fidelity 3D indoor scene generation with a sparse autoregressive cascade, from 25 cm down to 1.5625 cm.
Kosta Derpanis a relayé @davidbau
David Bau @davidbau · 02/11 18:29
I have been teaching myself formal code verification, but as the practice becomes more widespread, we must remember that mathematical perfection is often *not* good enough. Our systems need to be understandable to humans.
davidbau.com davidbau.com
R
Kosta Derpanis a relayé @ziv_ravid
Ravid Shwartz Ziv @ziv_ravid · 02/11 00:43
New episode of The Information Bottleneck with Julian Togelius @togelius about AI and games 🎮🕹️🎲 Julian (NYU professor, co-founder of) has been working on AI for video games since 2004, back when people told him it wasn’t real research. Basically every
modl.ai Home modl.ai uses AI agents and analysts to automate game testing and QA. Find bugs, glitches, and performance issues faster. No SDKs or code needed to get started.
M
Kosta Derpanis a relayé @drmichaellevin
Michael Levin @drmichaellevin · 01/11 23:36
Claude made some visuals for my process of paper-writing, useful for students. In case anyone else wants to use them: (from older blog post)
Forms of life, forms of mind Forms of life, forms of mind | Dr. Michael Levin | How I write papers My process for writing papers centers around a top-down outline and strict separation of the creative process of story-telling from the mechanics of filling in details.
thoughtforms.life https://thoughtforms.life/wp-content/uploads/2026/10/add-to-Substack-post.pdf
U
Kosta Derpanis a relayé @prof_kamilov
Ulugbek S. Kamilov @prof_kamilov · 01/11 20:58
We're hiring in Physical AI at UW–Madison! The ECE department is recruiting faculty at all ranks, with interests spanning multimodal sensing, physical modeling, robotics, and autonomous systems. Full consideration by Dec 14, 2026.
University of Wisconsin–Madison Careers Assistant/Associate/Full Professor in Electrical & Computer Engineering - Physical AI - Madison, Wisconsin, United States Current Employees: If you are currently employed at any of the Universities of Wisconsin, log in to Workday to apply through the internal application process.Job Category:FacultyEmployment Type:RegularJob Profile:ProfessorJob Summary:The Department of Electrical and Computer Engineering seeks to hire a faculty member with expertise in physical artificial intelligence, advancing the foundations and technologies that enable intelligent systems to sense, learn, and act reliably/safely in the physical world. Desired areas include perception and multimodal sensing, learning and modeling of physical
S
Kosta Derpanis a relayé @lamsheeper
Steven Yu @lamsheeper · 31/10 00:09
Can we trace an AI model's answers back to its training data? Surprisingly, even when we know exactly which documents taught a model a fact, attribution methods often fail to find them. Our CATT benchmark investigates why and uncovers two simple fixes🧵
catt.baulab.info Elucidating the Design Space of LM Data Attribution CATT, a controlled benchmark for LM data attribution, and two fixes that recover the ground-truth signal: the negative log-odds loss and top-k token pooling.
F
Kosta Derpanis a relayé @fhahlbohm
Florian Hahlbohm @fhahlbohm · 30/10 17:45
Ever since Thomas’ JIT-fusion update for Instant-NGP, I was wondering how fast a fully fused Zip-NeRF could be. Finally gave it a shot with Opus: 36 min training, 300 ms per megapixel on a 5090, same quality as the official code (~6-7x/10x faster). Code:
GitHub GitHub - fhahlbohm/fused-zipnerf: A fully fused CUDA implementation of Zip-NeRF by Barron et al. (ICCV 2023) for NeRFICG A fully fused CUDA implementation of Zip-NeRF by Barron et al. (ICCV 2023) for NeRFICG - fhahlbohm/fused-zipnerf
E
Kosta Derpanis a relayé @ELLISforEurope
ELLIS @ELLISforEurope · 30/10 02:48
⏰ Last chance! Apply to the #ELLISPhD Program by 31 Oct (23:59 AoE). Join leading AI/ML researchers across Europe with co-supervision, exchanges & networking. Get details & apply now:
ELLIS Society ELLIS PhD Program: Call for Applications 2026 The ELLIS PhD Program is a key pillar of the ELLIS initiative whose goal is to foster and educate the best talent in machine learning and related research areas by pairing outstanding students with leading academic and industrial researchers in Europe. The program also offers a variety of networking and training activi...
I
martin_casado a relayé @ideogram_ai
Ideogram @ideogram_ai · 29/10 23:17
Introducing context window for precise edit in Ideogram. No seams, no artifacts, no Photoshop! Edit any area up to 2K inside very high-resolution images. Everything outside the context window stays untouched. Try it now in the app and the API:
Ideogram The Precise Edit Upgrade: Edit Any Resolution up to 8K | Ideogram The upgraded Precise Edit supports a context window, so it edits images at any resolution up to 8K at their original pixels and renders small text, logos, and fine detail more sharply.
B
Kosta Derpanis a relayé @BreakingNews
Breaking News @BreakingNews · 29/10 21:24
BREAKING: Patrick Stewart, the beloved actor best known for his iconic roles as Jean-Luc Picard in "Star Trek" and Professor X in "X-Men," dies at 86.
NBC News Patrick Stewart, actor known for iconic roles in ‘Star Trek’ and ‘X-Men,’ dies at 86 The beloved star, who played both Capt. Jean-Luc Picard and Professor X, died Wednesday in Los Angeles following diagnoses of Alzheimer’s disease and Lewy body dementia.
W
Stanford Vision and Learning Lab a relayé @wensi_ai
Wensi Ai @wensi_ai · 29/10 21:23
Long-horizon, whole-body robot policies still fail during deployment. How do we best close this last-mile gap? 🤖 🚨 Introducing I3L: Iterative Interactive Imitation Learning for Whole-Body Mobile Manipulation Website: Paper:
arXiv.org I3L: Iterative Interactive Imitation Learning for Whole-Body Mobile Manipulation Policies trained offline from teleoperated demonstrations can degrade outside their training distribution. This gap is particularly acute in long-horizon whole-body mobile manipulation, where long horizons amplify compounding errors and coordinated arm, torso, and base motion makes exhaustive demonstration coverage impractical. We introduce I3L, an integrated hardware and learning system that combines whole-body correction with iterative policy refinement to address this last-mile reliability gap. Its low-cost, open-source JoyLo+ interface uses policy-tracking leader arms and touch-triggered t
i-3l.github.io I3L: Iterative Interactive Imitation Learning for Whole-Body Mobile Manipulation An integrated hardware and learning system that couples targeted whole-body correction with iterative policy refinement.
I
Walter Scheirer a relayé @ComputerSociety
IEEE Computer Society @ComputerSociety · 29/10 00:41
Help honor global leaders and rising talent shaping the future of #AI. Nominate for two new #IEEE Computer Society #awards: AI Innovator Award: AI Rising Stars Award: Deadline: 15 Dec 2026 Open worldwide, no membership required.
IEEE Computer Society IEEE Computer Society AI Innovator Award Honoring contributions to the field of Artificial Intelligence that have transformed the technical state of the art and its impact on society.
IEEE Computer Society IEEE Computer Society Artificial Intelligence Rising Stars Award An annual international recognition program established to honor undergraduate excellence, innovation, & responsible advancement in AI.
Douglas Gray @de3ug · 27/10 23:21
This reminds me of the original Yahoo web directory, except now an LLM does the browsing. Instead of nearest-neighbor retrieval, it navigates a hierarchy of topics, documents, and passages. I really like that the index structure is human readable.
arXiv.org Retriever-Free Retrieval: Navigating Documents with Decision Models While language models have become dramatically more capable, retrieval still largely relies on lexical matching or similarity in fixed embedding spaces. This creates a capability asymmetry: what the retriever fails to find, even the strongest generator cannot recover. We propose DecisionRAG, which brings the capabilities of pretrained language models to retrieval using decision models that jointly evaluate candidate branches in a single forward pass, without generating any tokens, and show that judging candidates in a shared context makes their relevance estimates comparable. Like a person bro
J
Frank Dellaert a relayé @Jsphamigo
Joseph Amigo @Jsphamigo · 27/10 00:20
Great to see first-order gradients through learned world models scale to whole-body tracking. Congrats @qiayuanliao et al.! 👏 We explored a related, decoupled variant in DMO (CoRL '25): forward in the real simulator, backward through a learned model. ⁠
machines-in-motion.github.io DMO: Decoupled Model-Based Policy Optimization for First-Order RL in Robotics Efficient first-order gradient RL using high-fidelity simulators and learned models for decoupled backpropagation, enabling robust sim-to-real transfer in locomotion tasks.
N
Aviad Dahan a relayé @NimrodShabtay
Nimrod Shabtay @NimrodShabtay · 26/10 11:25
Give a VLM a few demos as text: ~100% accuracy. As images, with no task hint: 45%. VLMs don’t learn from what they see in context. We teach them to watch and learn 🧵 Paper: Dataset:
arXiv.org Watch and Learn: Teaching Visual In-Context Learning by Removing Textual Shortcuts Vision-language models (VLMs) struggle with in context learning compared to their textual counterparts. Our preliminary study shows that VLMs rely on textual shortcuts and underutilize the visual demonstrations. We therefore hypothesize that VLMs are insufficiently trained to perform visual in-context learning (vICL). To this end, we construct MIME, a task-agnostic vICL training set without textual shortcuts. We ensure the instruction does not state the rule linking the demonstration images to their answers, so the model must infer it from the visual demonstrations. Fine-tuning on MIME improve
huggingface.co NimrodShabtay1986/MIME · Datasets at Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.
E
Sayan Deb Sarkar a relayé @efedele16
Elisabetta Fedele @efedele16 · 26/10 03:57
Tired of conditioning 3D generative models only using global text or image prompts? 👀 ➡️ SpaceFlow allows you to locally control the geometry and the appearance of your generations, using geometric primitives to inject independent 3D controls 👩‍🎨 🌍:
spaceflow3d.github.io SpaceFlow: Locally Controllable 3D Generation Locally controllable 3D generation with local geometric control and part-specific text or image appearance cues.
H
Kosta Derpanis a relayé @HildeKuehne
Hilde Kuehne @HildeKuehne · 26/10 00:40
If you want to join the TueAI crowd as Phd, we are currently recruiting : ELLIS PhD program () Deadline on October 31 🚨 IMPRS-IS PhD program () Deadline November 15, 2026
ELLIS Society ELLIS PhD Program: Call for Applications 2026 The ELLIS PhD Program is a key pillar of the ELLIS initiative whose goal is to foster and educate the best talent in machine learning and related research areas by pairing outstanding students with leading academic and industrial researchers in Europe. The program also offers a variety of networking and training activi...
imprs.is.mpg.de Application
M
Aviad Dahan a relayé @mikeeisenberg
Michael Eisenberg @mikeeisenberg · 25/10 17:35
The AI physical Olympics will be held in Tel Aviv 🇮🇱!! Details and registration below! Tel Aviv, May 18–19, 2027. 10,000 people in the arena. Forget scripted, pre-recorded demos where the robot succeeds on take 47. This one will be live! The
p.ai Redirecting...
C
Kosta Derpanis a relayé @ccloy
Chen Change Loy @ccloy · 25/10 17:18
What if you could step inside a van Gogh painting? We used Mundo to imagine new viewpoints from a single painting. Take a look at this short demo. Project: Blog:
huggingface.co Mundo: From seeing a place to simulating what happens there A Blog post by Yihang Luo on Hugging Face
mundo-world-model.github.io Mundo: Customized World Model Turn context images into editable 3D worlds for camera exploration, simulation, and robot planning.
L
Frank Dellaert a relayé @ganlumomo
Lu Gan @ganlumomo · 25/10 00:02
Check out 𝗠𝗼𝗿𝗽𝗵𝗠𝗶𝗺𝗶𝗰, a 𝗰𝗿𝗼𝘀𝘀-𝗲𝗺𝗯𝗼𝗱𝗶𝗺𝗲𝗻𝘁 𝘄𝗼𝗿𝗹𝗱 𝗺𝗼𝗱𝗲𝗹 and 𝘄𝗵𝗼𝗹𝗲-𝗯𝗼𝗱𝘆 𝗺𝗼𝘁𝗶𝗼𝗻 𝘁𝗿𝗮𝗰𝗸𝗲𝗿 for Unitree G1, AGIBOT X2 Ultra, and Booster T2 humanoid robots:
MorphMimic MorphMimic: Morphology-Conditioned World Model and Policy Sharing Across Humanoids One World, One Model: one world model and one policy track whole-body motion on the Unitree G1, AgiBot X2 Ultra and Booster T2.
A
Kosta Derpanis a relayé @petrenko_ai
Aleksei Petrenko @petrenko_ai · 24/10 16:58
To channel my inner Schmidhuber, we found this softmax "bug" ~1.5 years ago and published in this paper In reality, it's pretty hard to come up with a training scenario where this matters because gradient w.r.t. highly likely tokens is tiny anyway.
arxiv.org https://arxiv.org/pdf/2603.11682
K
Dmytro Mishkin a relayé @kornia_foss
Kornia @kornia_foss · 24/10 09:51
So, what’s new in kornia 0.9.0? First, the conventions page - what are images, what is the channel order, arrays, coordinates, angles. We also this information in the docstrings, so you or your agent always know.
Kornia Conventions & Pitfalls - Kornia The conventions every Kornia function follows: (B, C, H, W) float images in [0, 1], (x, y) points versus (h, w) sizes, degrees versus radians, homography direction, align_corners defaults, box formats and a pitfall checklist.
K
Dmytro Mishkin a relayé @kornia_foss
Kornia @kornia_foss · 23/10 21:01
Meet kornia==0.9.0 - more correct, much faster, more documented, better half-precision, CUDA and MPS support.
More than 800 issues have been raised and fixed. Some ops are now 2000x faster. Others not crash :) Way better docs. Check the release post
GitHub Release v0.9.0 - more correct, faster, better compile, half-precision, MPS and CUDA · kornia/kornia Overview for humans Meet kornia==0.9.0 - more correct, much faster, more documented, better half-precision, CUDA and MPS support. More than 800 issues have been raised and fixed. For example, now: ...
O
Kosta Derpanis a relayé @orlitany
Or Litany @orlitany · 23/10 20:21
Excited to share our NeurIPS 2026 paper (+code): Learnable Spectral Activations 📢 Prior INR work tunes which frequencies are available. LSA learns how to compose them. Gains on audio, images, NeRFs & more. 📄 Kudos to Tamir Shor who led it like a champ
arXiv.org Learnable Spectral Activations Implicit neural representations (INRs) are shaped by the spectral structure induced by their input encodings and activation functions. Existing methods improve fitting primarily by modifying which frequencies are available to the network, through coordinate encodings or periodic nonlinearities. However, frequency access is not the only bottleneck: signals with localized or spatially varying structure require the network to efficiently compose frequencies into multi-harmonic internal responses. We introduce learnable spectral activations (LSA), which replace fixed neuron-level nonlinearities wi
Ambient.ai @ambientai · 23/10 20:20
By the time a weapon is visible on a campus camera, the incident has already begun. Prevention lives upstream: loitering, tailgating, a person falling. One of 6 criteria in our new AI campus security evaluation guide for higher ed.
ambient.ai AI Campus Security: What Higher Education Should Evaluate Six criteria to evaluate any AI campus security platform, phrased as RFP-ready questions, with a vendor scorecard and one real higher education deployment read against all six.
R
Dmytro Mishkin a relayé @rosinality
Rosinality @rosinality · 23/10 19:36
Fixing numerical precision problem in softmax.
arXiv.org A Fix for Softmax and Cross-Entropy Standard implementations of softmax, log-softmax, and cross-entropy have two numerical bugs. The first bug, Softmax Collapse, rounds the target logit's gradient to zero if that one logit is much larger than the others, resulting in gradients that no longer sum to zero. The second bug, Softmax Cancellation, occurs when student logits are close to teacher logits, and produces non-zero gradients even when the student and teacher logits are equal. We present simple implementations of log-softmax, softmax, and cross-entropy that fix both bugs. Our fix eliminates the previously-observed "Slingshot"
G
Kosta Derpanis a relayé @thegautamkamath
Gautam Kamath @thegautamkamath · 23/10 18:59
How do you tell if authors fabricated their research results? Sometimes: they just tell you. Read, e.g., this caption carefully. Another banger of a post by my friend Nihar B. Shah on his blog, where he found at least 20+ such #ICLR2027 submissions:
Research On Research Self-confessed Fabrications and Rise of TBDs Some submissions to ICLR 2027 and papers on arXiv contain a very interesting instruction: to replace the reported experimental results with actual measurements before submission. The manuscripts th…
C
Stanford Vision and Learning Lab a relayé @GChongkai
Chongkai Gao @GChongkai · 23/10 18:43
🤖Robots need world models for visual planning, but they don't need to imagine every frame. Introducing DreamCan: Sparse Visual Planning for Long-Horizon Robot Manipulation. Project: Code: Paper:
arXiv.org Dream What I Can: Sparse Visual Planning for Long-Horizon Robot Manipulation Video models provide rich priors for generalizable robotic manipulation, but plausible visual futures are not always executable. We introduce DreamCan, a hierarchical policy that grounds video models through sparse visual planning: a high-level planner adapted from a pre-trained video model and a low-level visuomotor policy. Rather than predicting dense trajectories, the high-level planner generates sparse visual subgoals that balance reachability and task utility. From offline, action-free robot videos, we estimate these quantities and use dynamic programming to select subgoal sequences witho
dreamcan.github.io Dream What I Can: Sparse Visual Planning forLong-Horizon Robot Manipulation DreamCan: sparse visual planning for long-horizon robot manipulation with reachable, useful visual subgoals.
GitHub GitHub - HeegerGao/DreamCan: Codebase for DreamCan Codebase for DreamCan. Contribute to HeegerGao/DreamCan development by creating an account on GitHub.
R
Kosta Derpanis a relayé @rosinality
Rosinality @rosinality · 23/10 17:28
Fixing numerical precision problem in softmax.
arXiv.org A Fix for Softmax and Cross-Entropy Standard implementations of softmax, log-softmax, and cross-entropy have two numerical bugs. The first bug, Softmax Collapse, rounds the target logit's gradient to zero if that one logit is much larger than the others, resulting in gradients that no longer sum to zero. The second bug, Softmax Cancellation, occurs when student logits are close to teacher logits, and produces non-zero gradients even when the student and teacher logits are equal. We present simple implementations of log-softmax, softmax, and cross-entropy that fix both bugs. Our fix eliminates the previously-observed "Slingshot"
G
Yoshitomo Matsubara a relayé @thegautamkamath
Gautam Kamath @thegautamkamath · 23/10 17:07
How do you tell if authors fabricated their research results? Sometimes: they just tell you. Read, e.g., this caption carefully. Another banger of a post by my friend Nihar B. Shah on his blog, where he found at least 20+ such #ICLR2027 submissions:
Research On Research Self-confessed Fabrications and Rise of TBDs Some submissions to ICLR 2027 and papers on arXiv contain a very interesting instruction: to replace the reported experimental results with actual measurements before submission. The manuscripts th…
R
Laura Leal-Taixe a relayé @RdeLutio
Riccardo de Lutio @RdeLutio · 23/10 07:40
Excited to share our new Cosmos3-Nano–based ArtiFixer is now out 🚀 Almost matching the quality of the Wan 2.1 14B based one with just 8B parameters in the generation tower. And most importantly 3–4× faster inference while using about half the GPU memory!
research.nvidia.com ArtiFixer: Enhancing and Extending 3D Reconstructionwith Auto-Regressive Diffusion Models ArtiFixer: Enhancing and Extending 3D Reconstruction with Auto-Regressive Diffusion Models
N
Kosta Derpanis a relayé @NeurIPSConf
NeurIPS Conference @NeurIPSConf · 23/10 06:57
To all NeurIPS 2026 authors: please review the NeurIPS poster instructions and see our official poster printing partner here: You may use any poster printing service you like, but we recommend using Poster Sessions Online, so you can purchase your poster
neurips.cc NeurIPS 2026 Poster Instructions
R
Kosta Derpanis a relayé @Robin_Tian
Ran TIAN @Robin_Tian · 23/10 00:53
If you have ever trained a Transformer Language Model and wondered why there are "spikes" in loss -- our recent paper () gives an explanation that arguably gets to the most essence so far, as it replicates loss spikes by simple quadratic models.
arXiv.org Simulating Training Spikes of Deep Neural Networks by Layer-wise Random Feature Models Intermittent ``spikes'' in loss, during the pre-training of Large Language Models (LLMs), remains a mysterious phenomenon both theoretically and empirically. In this paper, we show that simple quadratic models with heavy-tailed, power-law distributed inputs can replicate these spikes. We introduce a theoretical framework and simulation algorithm to model per-layer training dynamics of Deep Neural Networks (DNNs), using such quadratic models. In a case study on an Adam hyperparameter configuration (where $β_2$ is set close to $1$) known to induce spikes in Transformer LMs, our simulation succes
A
Kosta Derpanis a relayé @ArashVahdat
Arash Vahdat @ArashVahdat · 23/10 00:28
🚀 Are you a PhD student pushing the boundaries of AI, robotics, graphics, or accelerated computing? NVIDIA's Graduate Fellowship offers up to $60K in funding, mentorship from NVIDIA researchers, and a summer research internship! ⏳ Apply by October 30!
NVIDIA Blog NVIDIA Opens Applications for 2027–2028 Graduate Fellowships With Awards Up to $60,000 The NVIDIA Graduate Fellowship Program, in its 26th year, is now accepting applications worldwide.
I
martin_casado a relayé @ideogram_ai
Ideogram @ideogram_ai · 22/10 18:21
Today, we’re introducing Ideogram's first vector model. Generate clean, editable SVGs with smooth gradients and sharp typography from a text prompt, or turn any image into a detailed, scalable vector. Live in Ideogram and the API:
Ideogram AI Vector Generator: Ready-to-use SVG Artwork | Ideogram Create logos, print designs, and illustrations from a prompt or an image. Get scalable, editable SVG artwork ready to bring into your next project.
Antoine Guédon a relayé @Cohere_Labs
Cohere Labs @Cohere_Labs · 22/10 15:50
Our Open Science Community is excited to share our events for next week, October 26 - October 31! Check out the upcoming sessions: 📆 Surflo: Consistent 3D Surface Flow Model with Global State with @antoine_guedon: 📆 Building AI That Teaches: The
luma.com Antoine Guedon - Surflo: Consistent 3D Surface Flow Model with Global State · Luma Geometry is invariant to viewpoint, which makes any collection of images a redundant encoding of a single 3D state. Existing feed-forward reconstruction models…
Kosta Derpanis a relayé @YiMaTweets
Yi Ma @YiMaTweets · 22/10 14:39
The lectures of my new course "Principles of Deep Representation Learning" are now available at the Youtube site: with new lectures continuing to be added there. I will open-source all the slides later in the textbook website:
ma-lab-berkeley.github.io Landing page for the book Principles and Practice of Deep Representation Learning. https://ma-lab-berkeley.github.io/deep-representation-learning-book
YouTube Principles of Deep Representation Learning | HKU DATA8014 Par HKU data8014
A
Kosta Derpanis a relayé @ashmrz10
Ashkan Mirzaei @ashmrz10 · 21/10 23:57
Check out Cosmos Simulation: Interactive World Simulation with Cosmos for Physical AI: 🌐Website: 📜Report: 🤖 Models:
huggingface.co Cosmos3-Examples - a nvidia Collection Domain-specialized Post-training Examples of Cosmos3
research.nvidia.com Interactive World Simulation with Cosmos for Physical AI Interactive world simulation with Cosmos: real-time, streaming models for camera navigation, robot manipulation, and autonomous driving.
research.nvidia.com https://research.nvidia.com/labs/cosmos-lab/cosmos-simulation/technical-report.pdf
K
Dmytro Mishkin a relayé @kornia_foss
Kornia @kornia_foss · 21/10 21:00
From now on, instead of question “How does AI help?” we ask “Which part was done by human, if any?”
GitHub docs(agents): require GitHub signature line and ask which part was done by a human by ducha-aiki · Pull Request #6223 · kornia/kornia What does this pull request do? Updates .github/pull_request_template.md, AGENTS.md, and CONTRIBUTING.md: Replaces the PR template section ## How did AI help? with ## Which part was done by a huma...
Ambient.ai @ambientai · 21/10 18:00
AI can watch every school camera at once and flag what matters. The question is whether you can trust what it does next. Oct 28, 1 PM ET: Jody Russell joins a panel on AI-powered school security, smarter detection done responsibly. Register:
app.webinar.net https://app.webinar.net/r/6J5olnjlEBb
D
Yuki a relayé @danilodegoede
Danilo de Goede @danilodegoede · 21/10 16:45
How many tokens does it take to capture the gist of an image? Introducing GIST: an image tokenizer that learns to preserve what matters and chooses how many tokens each image needs, in one forward pass. Now on arXiv 🎉🎉 Demo:
arXiv.org Semantically Steerable Image Tokenization Adaptive image tokenizers compress an image into a variable number of tokens matched to its complexity, and recent work makes this allocation controllable: conditioned on a target reconstruction error, the tokenizer itself decides how many tokens an image needs. However, the fidelity measure guiding this allocation shapes which aspects of the image are preserved under compression. The pixel-wise error used in prior work is blind to the spatial structure that carries an image's meaning, and under heavy compression such tokenizers are prone to abruptly discarding tokens that encode essential con
ddgoede.github.io GIST: Semantically Steerable Image Tokenization An adaptive image tokenizer steered by a semantic target error. Up to 30% lower LPIPS and 37% lower DreamSim than KARL at matched token counts.
Dmytro Mishkin a relayé @taiyasaki
Andrea Tagliasacchi @taiyasaki · 20/10 09:24
📢📢📢 introducing 𝐑𝐨𝐛𝐨𝐂𝐥𝐨𝐭𝐡 Neural materials still learn from 100 BRDFs measured in 2003 at MERL. As of today, that is no longer true (for cloth materials). home: arXiv: 500 cloths × 580 views – 80 billion observations
arxiv.org https://arxiv.org/pdf/2610.21090
RoboCloth RoboCloth: A Large-Scale Real Cloth Material Dataset for Neural Reflectance Reconstruction 500 real cloth materials × ≈580 calibrated HDR frames each, captured by two robot arms — ≈80 billion RGB observations — with a two-stage neural BRDF reconstruction pipeline.
S
Kosta Derpanis a relayé @ShengyHuang
Shengyu Huang @ShengyHuang · 20/10 07:35
1/7 I'm super excited to share LidarVerse! One model that reconstructs, forecasts and generates LiDAR across sensors, in real time. 📄 Paper: 🌐 Project: 💻 Code: 🤗 Models:
arXiv.org Into the LidarVerse: A Real-Time Unified LiDAR World Model Across Tasks and Sensors We present LidarVerse, a real-time unified LiDAR world model for sensor simulation across tasks and sensors. Existing LiDAR simulators largely follow one of two complementary paradigms: reconstruction-based methods preserve the geometry of a captured scene but require lengthy per-scene optimization, whereas generative methods amortize learning across data but lack the consistency needed for long-horizon rollouts. LidarVerse brings both paradigms into a single, unified conditional generative framework. A universal tokenizer maps LiDAR range sequences into a shared latent space, where a bidirect
GitHub GitHub - nv-tlabs/lidarverse: A Real-time LiDAR World Model that generates, forecasts, and resimulates within a single model. A Real-time LiDAR World Model that generates, forecasts, and resimulates within a single model. - nv-tlabs/lidarverse
huggingface.co nvidia/lidarverse · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.
research.nvidia.com Into the LidarVerse LidarVerse unifies generative and reconstruction-grounded LiDAR sensor simulation in one real-time world model.
Kosta Derpanis a relayé @jbhuang0604
Jia-Bin Huang @jbhuang0604 · 20/10 05:20
How to Train Models Too Big for One GPU AI models are getting larger! How do we train them when they can't fit? This video covers how we represent and store real numbers, mixed-precision training, and the Zero Redundancy Optimizer (ZeRO) & FSDP.
YouTube Training Models Too Big for One GPU | ZeRO & FSDP How do you train a model that does not fit on one GPU? In this video, we cover how GPUs represent and store real numbers, mixed-precision training, and the Z...
I
martin_casado a relayé @isaac_mason_
isaacmason @isaac_mason_ · 19/10 18:28
new video! making splats move with #sparkjs + #threejs + @theworldlabs marble covers: splat reveal -> waving leaves -> splat dust -> splat fire -> skinned splavatars hoping this gives some inspiration for more sparkjs demos! a webgpu sparkjs is also in
YouTube Animate Gaussian Splats with Spark Bring Gaussian splat scenes to life with Spark. This tutorial explores how to add motion and atmosphere, from revealing a splat and making tree leaves sway t...
S
Laura Leal-Taixe a relayé @ShengyHuang
Shengyu Huang @ShengyHuang · 19/10 18:19
1/7 I'm super excited to share LidarVerse! One model that reconstructs, forecasts and generates LiDAR across sensors, in real time. 📄 Paper: 🌐 Project: 💻 Code: 🤗 Models:
arXiv.org Into the LidarVerse: A Real-Time Unified LiDAR World Model Across Tasks and Sensors We present LidarVerse, a real-time unified LiDAR world model for sensor simulation across tasks and sensors. Existing LiDAR simulators largely follow one of two complementary paradigms: reconstruction-based methods preserve the geometry of a captured scene but require lengthy per-scene optimization, whereas generative methods amortize learning across data but lack the consistency needed for long-horizon rollouts. LidarVerse brings both paradigms into a single, unified conditional generative framework. A universal tokenizer maps LiDAR range sequences into a shared latent space, where a bidirect
GitHub GitHub - nv-tlabs/lidarverse: A Real-time LiDAR World Model that generates, forecasts, and resimulates within a single model. A Real-time LiDAR World Model that generates, forecasts, and resimulates within a single model. - nv-tlabs/lidarverse
huggingface.co nvidia/lidarverse · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.
research.nvidia.com Into the LidarVerse LidarVerse unifies generative and reconstruction-grounded LiDAR sensor simulation in one real-time world model.
K
Kosta Derpanis a relayé @kooroshroohi
Koorosh Roohi @kooroshroohi · 19/10 03:46
Happy to share that PhotoQuilt got accepted to #WACV2027 🎉 Some bugs are mistakes. Some are ideas in disguise. Training-free photomosaics at any resolution. Zoom in and every tile is its own image. 🔗 🤗 Check out our demo:
huggingface.co PhotoQuilt - a Hugging Face Space by kooroshrh Training-free arbitrary-resolution photomosaics
kooroshrh.github.io PhotoQuilt PhotoQuilt: a training-free framework for generating photomosaics at arbitrary resolution via bootstrapped tiled denoising.
M
Kosta Derpanis a relayé @coopermj_aiml
Michael Cooper @coopermj_aiml · 18/10 19:41
In reading group reviewing flow models, I realized I didn't understand the change-of-variables rule as fully as I probably should. So I put together this blog post and collection of visuals to help me more fully understand the Jacobian-inverse term. 💪
Michael Cooper Visualizing the Change-of-Variables Rule in Probability Building geometric intuition for the Jacobian in the change-of-variables formula, with interactive figures.
O
Visual Geometry Group (VGG) a relayé @DebOishi
Oishi Deb @DebOishi · 18/10 12:51
Thank u G-Research @GRESEARCHjobs for featuring our @icmlconf ICML (Oral) 2026 paper on Benchmarking at the Edge of Comprehension!! Thank u to Samuele Marro & the excellent collaborators from @MSFTResearch @MSFTResearchCam! #ICML2026!
G-Research ICML 2026: Paper reviews | G-Research Explore G-Research researchers’ reviews of standout papers from ICML 2026 and discover the latest ideas shaping machine learning research.
L
Kosta Derpanis a relayé @ldkong1205
Lingdong Kong @ldkong1205 · 17/10 21:37
🤔 Do generated videos form a consistent 3D world? Introducing 𝗢𝗿𝗯𝗶𝗮, our new benchmark for testing how well video models preserve the worlds they generate. As the camera explores a scene, the model needs to preserve the original environment,
orbia-bench.github.io Orbia: Do Generated Videos Form a Consistent 3D World? A benchmark for 3D consistency in world models: explore a scene, return to it, and test what the model remembers.
Kosta Derpanis a relayé @arankomatsuzaki
Aran Komatsuzaki @arankomatsuzaki · 17/10 15:14
PingpongDuo Humanoid table tennis now has doubles, while I'm still single. proj: abs:
Archon Robotics PingpongDuo — Agile Humanoid Table Tennis Doubles Agile humanoid table tennis doubles with safe team coordination.
arXiv.org PingpongDuo: Agile Humanoid Table Tennis Doubles with Safe Team Coordination Unlike singles, humanoid table tennis doubles requires teammates to strike in strict alternation while sharing a confined workspace. Rapid repositioning can compromise striking accuracy, while uncoordinated teammate movements introduce collision risks. We present PingpongDuo, a hierarchical framework for agile humanoid doubles with safe team coordination. At the team level, a planner jointly selects positioning targets and uses learned risk estimates to filter risky proposals, preferentially preserving the current striker's target. At the individual level, separate footwork and striking skills
D
Kosta Derpanis a relayé @serra9lozano
David Serrano-Lozano @serra9lozano · 17/10 14:04
Excited to share SyncLight, accepted at #NeurIPS2026! Click a light in one image, pick its color and intensity, and SyncLight, a repurposed image generative model, relights every other view to match. No camera poses, 3DGS or distillation. 🌐 🧵👇
color.cvc.uab.cat https://color.cvc.uab.cat/synclight
G
Kosta Derpanis a relayé @3blue1brown
Grant Sanderson @3blue1brown · 16/10 23:58
Some reflections on AI in math. I found it helpful to simply list out the various parts of what a mathematician does, why math is valuable, and where the funding comes from. From there, what happens if AI gets more capable?
YouTube There's never been a better time to go into math Sharing a few thoughts on the future of mathThese videos are funded directly by viewers: https://3b1b.co/supportAn equally valuable form of support is to sim...
Frank Dellaert a relayé @zhenjun_zhao
Zhenjun Zhao @zhenjun_zhao · 16/10 23:44
FoundationVIO: Enforcing Metric Scale in Feed-forward Visual-Inertial Odometry Jin Rhee, Yuhao Zhang, @fdellaert, @MauriceFallon tl;dr: IMU measurements + GFM predictions->factor graph
arXiv.org FoundationVIO: Enforcing Metric Scale in Feed-forward Visual-Inertial Odometry Feed-forward geometric foundation models (GFMs) directly predict dense geometry and camera poses from batches of images. Emerging visual SLAM methods chain together consecutive GFM batches to form dense point cloud maps. A core challenge is how to address the ambiguous geometric scale of GFM predictions: independent predictions are locally consistent up-to-scale but not constrained by any direct physical measurements. Models trained on data with known scale improve the consistency of GFM reconstructions but again do not directly measure it. Consequently, trajectories and 3D reconstructions fro
Z
Zhengzhong Tu @_vztu · 16/10 22:07
Machine unlearning erases specific knowledge from a trained foundation model, such as a person's face, private data, or copyrighted content. We study MLLM unlearning and find 3 evaluation pitfalls that make unlearning look better than it is. 🧵1/
arXiv.org Are We Really Measuring Forgetting? Pitfalls in Evaluating Machine Unlearning for Multimodal LLMs As machine unlearning for multimodal large language models (MLLMs) rapidly advances, rigorous evaluation of its effectiveness becomes increasingly important. However, existing benchmarks largely inherit evaluation protocols from text-only unlearning and overlook the unique challenges of unlearning in multimodal settings. In this work, we identify three major pitfalls that may lead to misleading conclusions: the lack of heldout evaluation to distinguish genuine concept-wise forgetting from instance-wise overfitting, the reliance on synthetic, fine-tuned knowledge rather than real-world knowledg
P
Kosta Derpanis a relayé @phillip_isola
Phillip Isola @phillip_isola · 16/10 14:13
For anyone interested in the theory of neural net representations, and representational alignment, I highly recommend Sharut’s PhD defense talk, linked below, which is an amazingly clear and deep presentation!
YouTube Sharut Gupta Thesis Defense: Grounding Neural Representations from Many Views of the World My PhD thesis defense at MIT CSAIL on "Grounding Neural Representations from Many Views of the World."This thesis explores how neural representations can be ...
K
Dmytro Mishkin a relayé @kornia_foss
Kornia @kornia_foss · 16/10 11:34
Our new #rust kornia-slam pipeline is finally ou. Check the site for more details: It’s fully #opensource — we are looking for battle testing users and contributors to improve it. Feel free to reach out to us !
Kornia kornia-slam: visual-inertial SLAM and VIO in Rust | Kornia Open-source visual-inertial SLAM and odometry (VIO) in Rust: ORB features, stereo + IMU, loop closure and Copper tasks. Integration help for your robot.
S
Sukrut Rao a relayé @sidgairo18
Siddhartha Gairola @sidgairo18 · 16/10 09:54
🚨 Checkout our new work on learning self-supervised visual part representations: "DisParQ: Self-Supervised Part Concepts for Interpretable Vision Foundation Models". Very proud and happy with this work. Tremendous effort by all involved! 🤗 🚀 1/10 📄
arXiv.org DisParQ: Self-Supervised Part Concepts for Interpretable Vision Foundation Models Concept-based vision models represent images through an intermediate layer of human-inspectable concepts, so what a model relies on can be traced to those concepts. However, those models are often limited to fixed categories or depend on language to define their concepts. We introduce DisParQ (Discrete Parts with Quantized attributes), a method that learns spatially grounded, discrete concept representations from a powerful frozen vision-only self-supervised backbone. It requires no class labels and no language supervision. Each image patch is assigned to exactly one concept from a learnable p
Y
Andrew Davison a relayé @cheng_yuzh21715
Yuzhou Cheng @cheng_yuzh21715 · 16/10 00:34
Our MLP-Splatting (ECCV 2026) code is out:. We splat MLPs as light-field primitives instead of Gaussians: more expressive, fewer primitives, and better aligned with object semantics. Give it a spin!
GitHub GitHub - kxhit/mlp-splatting: (ECCV2026 MLP Splatting) Each mlp is a light field primitive instead Gaussians. We splat mlps, making the representation more expressive, fewer primitive numbers, and align with object semantics. (ECCV2026 MLP Splatting) Each mlp is a light field primitive instead Gaussians. We splat mlps, making the representation more expressive, fewer primitive numbers, and align with object semantics. -...
Ambient.ai @ambientai · 15/10 21:00
In a colocation data center, security ends at someone else's door, and the risk collects in the shared spaces between. Our latest blog on why a valid badge can't say who belongs there: #DataCenterSecurity #AIinPhysicalSecurity #AgenticPhysicalSecurity
ambient.ai Colocation Data Center Physical Security: Shared Spaces In a colocation data center, shared spaces are the operator's to secure. Why a badge can't say who belongs, and where operators and tenants each stand.
Zheng Tang @tangzhengthomas · 15/10 20:54
🚀 NeurIPS 2026 Smart Buildings Contest! 🏆 $40K in prizes 🎤 Round 1 winners invited to present in Paris 🌍 Build HVAC policies that generalize across unseen buildings to improve energy efficiency. 📅 Papers: Nov 13, 2026 📅 Final: Jan 15, 2027 🔗
smart-buildings-2026.github.io The Smart Buildings ChallengeLearning Foundation Models for Real-World Optimization at Scale A NeurIPS 2026 challenge to build one HVAC control policy that generalizes across unseen commercial buildings.
K
Juan Carlos Niebles a relayé @keshigeyan
Keshigeyan Chandrasegaran @keshigeyan · 15/10 19:04
Excited to share Halcyon, our latest research on multimodal pretraining! Building strong text-to-image generative models is simpler than we think! Blog: Models: Paper:
huggingface.co Halcyon - a radicalnumerics Collection Halcyon family of autoregressive text-to-image and image-to-text models.
radicalnumerics.ai https://radicalnumerics.ai/assets/halcyon_paper.pdf
radicalnumerics.ai Text-to-image generative modeling without language pretraining Halcyon is a family of early-fusion multimodal autoregressive models pretrained from scratch on image-text pairs, with no internet-scale language pretraining.
O
OpenCV - Open Source Computer Vision Library @opencvofficial · 15/10 17:14
Half of all described species on Earth are insects, and almost nobody is counting them. 🦋 The Mothbox changes that: a $450 open-source trap that photographs and IDs insects with YOLOv8 + BioCLIP. Dr. Andrew Quitmeyer joins OpenCV Live at 9: #OpenCV
YouTube Mothbox - Open Source Insect Biodiversity Monitoring System - OpenCV Live! 229 Join OpenCV Live for a conversation with Dr. Andrew James Quitmeyer, founder of Digital Naturalism Laboratories (Dinalab), about the Mothbox, an open-source ...
J
Kosta Derpanis a relayé @rajabi2001
Javad Rajabi @rajabi2001 · 15/10 14:40
📢🚀 Excited to share AReA: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers Training-free, up to 6× the native resolution, up to 3.2× faster. 📄Paper: 🌐Webpage: 🤗HF:
arXiv.org AReA: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers Generating images beyond the training resolution of Diffusion Transformers (DiTs) remains challenging, as the number of tokens grows rapidly with resolution, full attention becomes expensive, and outputs often suffer from structural distortions and degraded detail. We identify uniform full attention as the common source of both problems. Attending to all tokens at every step causes quadratic cost and exposes the model to more tokens and relative distances beyond the training range, pushing the model out of distribution. We observe that attention's effective receptive field narrows during denoi
rajabi2001.github.io AReA: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers Training-free, efficient high-resolution generation: global attention early, local windows late.
E
Francis Engelmann a relayé @ELLISforEurope
ELLIS @ELLISforEurope · 15/10 10:58
🎓 Work with TWO leading European researchers across borders while you achieve your research goals in the #ELLISPhD Program! Check out the value Dominika Długosz sees in joint supervision. Apply now to start your PhD in 2027 at. Deadline is 31 October!
ELLIS Society ELLIS PhD Program: Call for Applications 2026 The ELLIS PhD Program is a key pillar of the ELLIS initiative whose goal is to foster and educate the best talent in machine learning and related research areas by pairing outstanding students with leading academic and industrial researchers in Europe. The program also offers a variety of networking and training activi...
I
Kosta Derpanis a relayé @Ismini_L
Ismini Lourentzou @Ismini_L · 14/10 23:32
🧩 Introducing ELSA3D, a new unified 3D model that generates 3D assets and describes them in language, in one backbone. 🌐 Led by @Tianjiao_Yu and accepted at #NeurIPS2026 👏 Instead of dumping text + 3D tokens into one flat sequence, sparse Anchor
plan-lab.github.io https://plan-lab.github.io/elsa3D
X
Kosta Derpanis a relayé @xyz2maureen
Xueyan Zou @xyz2maureen · 14/10 18:22
I am going to recruit one international PhD (Non-Mainland citizen) student this cycle (2027 Fall), please email me if you are interested in taking a PhD position in Tsinghua University, CollegeAI. Intro:
maureenzou.github.io XYZ
G
Kosta Derpanis a relayé @GerardPonsMoll1
Gerard Pons-Moll @GerardPonsMoll1 · 14/10 13:46
This year we are offering a virtual humans course again: I am starting a repo with basic tools to go deeper into the material and allow students to start building and doing research: Code: The first tool is a viewer for the SMPL
GitHub GitHub - gerardponsmoll/vh-toolbox: Interactive demos for the Virtual Humans lecture (University of Tübingen): SMPL viewer with shape and pose controls Interactive demos for the Virtual Humans lecture (University of Tübingen): SMPL viewer with shape and pose controls - gerardponsmoll/vh-toolbox
virtualhumans.mpi-inf.mpg.de Virtual Humans
N
Sayan Deb Sarkar a relayé @Neiluss_
Neil @Neiluss_ · 13/10 18:43
🚀 Introducing SpaceFlow: Locally Controllable 3D Generation Choose which parts follow your input shape, which are generated freely and how each part should look. Training-free generation with local shape control and part-specific text/image cues. 🧵⬇️
spaceflow3d.github.io SpaceFlow: Locally Controllable 3D Generation Locally controllable 3D generation with local geometric control and part-specific text or image appearance cues.
N
Sayan Deb Sarkar a relayé @Neiluss_
Neil @Neiluss_ · 13/10 17:00
🚀 Introducing SpaceFlow: Locally Controllable 3D Generation Choose which parts follow your input shape, which are generated freely and how each part should look. Training-free generation with local shape control and part-specific text/image cues. 🧵⬇️
spaceflow3d.github.io SpaceFlow: Locally Controllable 3D Generation Locally controllable 3D generation with local geometric control and part-specific text or image appearance cues.
S
Kosta Derpanis a relayé @sinaatalay
Sina Atalay @sinaatalay · 13/10 14:11
Introducing ManimGX, the 3D animation engine for agents. - Blazingly fast: 90× faster than ManimCE - Rust + wgpu core - Compatible with Manim's Python API - Lightweight: one pip install, no system dependencies Open source: Docs:
GitHub GitHub - academa-labs/manimgx: The animation engine for agents: blazingly fast math and 3D videos with Manim's Python API, on a Rust GPU core. The animation engine for agents: blazingly fast math and 3D videos with Manim's Python API, on a Rust GPU core. - academa-labs/manimgx
ManimGX ManimGX Blazingly fast 3D animations, for creators and agents. Open source, with Manim's Python API and a Rust GPU core.
S
Kosta Derpanis a relayé @SophieLWang
Sophie Wang @SophieLWang · 13/10 00:05
I wrote a blog post detailing the writing approach I took while working on my recent paper: I describe the behind the scenes for how this paper came to be, why and how I storyboarded figures before running experiments, and some figure design tips! :)
Sophie Wang behind the scenes of paper writing
I
Kosta Derpanis a relayé @Ismini_L
Ismini Lourentzou @Ismini_L · 12/10 20:06
How can robots understand our intentions, read social cues, and collaborate with us? 🤖 Excited to co-organize the 1st Workshop on Social Embodied AI (SEAI) at #WACV2027 with Yuanzhe Liu, @IrohXu, @RehgJim, @mengdixu_, @SynoRing, and the rest of our
wacv27seai.synoring.ai The 1st Workshop on Social Embodied AI @ WACV 2027 WACV 2027 Workshop on Social Embodied AI (SEAI). Bringing together computer vision, multimodal learning, embodied AI and human-robot interaction researchers.
martin_casado a relayé @a16z
a16z @a16z · 12/10 17:31
co-founders Dhaval Adjodah and Owen He with a16z's Martin Casado on recursive self-improvement, and how to tell when it's real: A lot of RSI pitches are really just people using AI to build better AI. Tech has always worked that way. You needed computers
makermaker.ai MakerMaker.AI
C
Visual Geometry Group (VGG) a relayé @ChuanxiaZ
Chuanxia Zheng @ChuanxiaZ · 12/10 11:31
How can we generate and evaluate physically grounded articulated 3D assets? 🚀 Introducing ARTISAN: a multi-agent generation pipeline and a comprehensive benchmark 🔗 Project: 💻 Code and datasets:
GitHub GitHub - physicalvision/artisan: Towards Physically Grounded Articulated 3D Assets: A Multi-Agent Pipeline and Benchmark Towards Physically Grounded Articulated 3D Assets: A Multi-Agent Pipeline and Benchmark - physicalvision/artisan
physicalvision.github.io Artisan — Towards Physically Grounded Articulated 3D Assets A multi-agent pipeline with physical simulation feedback, and a benchmark that measures how generated objects look, move, and function.
Kosta Derpanis a relayé @hkproj
Umar Jamil @hkproj · 10/10 17:21
I’m in Seoul building Mistral’s founding Korea team. Looking for kind, ambitious researchers and engineers to build and customise LLMs for real problems, from data and evals to RL and training at scale. Here until 16th October. DM me or apply here:
jobs.ashbyhq.com Applied Scientist, South Korea ABOUT MISTRAL Mistral provides full-stack AI solutions: from frontier models to developer tools, applications, and compute. We partner with enterprises tackling the hardest problems across high-stakes industries like finance, manufacturing, defense, healthcare, and the public sector, co-creating customized AI systems that they can run on their terms. We are a dynamic, collaborative team passionate about AI and its potential to transform society. Our diverse workforce thrives in competitive environments and is committed to driving innovation. Our teams are distributed between Europe, North Am
O
Visual Geometry Group (VGG) a relayé @DebOishi
Oishi Deb @DebOishi · 10/10 10:59
Checkout our VGI White Paper! Great 2 collaborate with amazing @HirokatuKataoka @gatheluck @YonglongT @elliottszwu @FragileGoodwill @jianyuan_wang Robert Geirhos @AdtRaghunathan @y_m_asano @RamananDeva @AjdDavison @du_yilun @jiajunwu_cs @liuzhuang1234 1/n
arXiv.org Visual General Intelligence: A White Paper This paper reconsiders intelligence from a vision-centered perspective and examines whether intelligence emerging from visual experience and learning may provide a pathway toward AGI. In the language domain, beginning with the introduction of the Transformer architecture, the GPT series has demonstrated transfer to unseen tasks through autoregressive language modeling on web-scale text combined with aggressive scaling. This raises a natural question, namely, what capabilities and forms of intelligence can emerge from visual modalities such as images, videos, and geometry? In this paper, we dis
O
Visual Geometry Group (VGG) a relayé @DebOishi
Oishi Deb @DebOishi · 09/10 23:13
Checkout our VGI White Paper! Great 2 collaborate with amazing @HirokatuKataoka @gatheluck @YonglongT @elliottszwu @FragileGoodwill @jianyuan_wang Robert Geirhos @AdtRaghunathan @y_m_asano @RamananDeva @AjdDavison @du_yilun @jiajunwu_cs @liuzhuang1234 1/n
arXiv.org Visual General Intelligence: A White Paper This paper reconsiders intelligence from a vision-centered perspective and examines whether intelligence emerging from visual experience and learning may provide a pathway toward AGI. In the language domain, beginning with the introduction of the Transformer architecture, the GPT series has demonstrated transfer to unseen tasks through autoregressive language modeling on web-scale text combined with aggressive scaling. This raises a natural question, namely, what capabilities and forms of intelligence can emerge from visual modalities such as images, videos, and geometry? In this paper, we dis
L
Laura Leal-Taixe @lealtaixe · 09/10 17:11
Come build the future of simulation for AV and robotics at @nvidia! We are looking for interns to work on generative AI, world models, reconstruction, policy learning, VLMs, geometry and scene understanding. Apply here!
nvidia.wd5.myworkdayjobs.com Research Intern, Spatial Intelligence - Summer 2027 NVIDIA is searching for an outstanding computer vision and deep learning intern researcher to join the spatial intelligence lab. We are passionate about research that pushes boundaries but also has impact in the real world! What you’ll be doing: Research, design and implement methods for one of the key research topics of the team: generative AI for videos, reconstruction, simulation, video understanding, policy learning, reinforcement learning, VLMs. Publish original research at the end of the internship. Collaborate with other team members and teams. Collaborate with external researchers. Wha
J
Kosta Derpanis a relayé @JYenphraphai
Jiraphon Yenphraphai @JYenphraphai · 09/10 15:59
[1/4] 🏛️ Building Rome from a Single Image 🏛️ We reconstruct 3D scenes from the input image, including geometry beyond the observed view across indoor, outdoor, and large-scale scenes. 👉 Project:
build-rome.github.io Building Rome from a Single Image Our method reconstructs detailed scene meshes from one photograph, completing geometry beyond the observed view for indoor, outdoor, and large-scale scenes.
Ambient.ai @ambientai · 09/10 00:00
"We didn't get there by hiring more people. We got there by getting smarter." That's Bill Owen, Public Safety Director at Univ. of Northwestern-St. Paul. 15 buildings, one patrol officer, on existing cameras. Our 6-criteria AI campus security guide:
ambient.ai AI Campus Security: What Higher Education Should Evaluate Six criteria to evaluate any AI campus security platform, phrased as RFP-ready questions, with a vendor scorecard and one real higher education deployment read against all six.
Y
Kosta Derpanis a relayé @yxue_yxue
Yuxuan Xue @yxue_yxue · 08/10 15:28
Amazing results, very entertaining to watch the teaser ;) We proudly power the human appearance: A Blender Add-On for unifying 15 HOI datasets: More frontier exciting research coming soon!
yuxuan-xue.com Revisiting Avatar-As-Image:High-Fidelity Registration is All You Need AvaImg: A multi-stage optimization pipeline for high-fidelity SMPL-X+D registration with UV texture mapping, enabling Avatar-As-Image representation for 3D clothed humans.
yuxuan-xue.com HOI-Blender: A Unifying Blender Add-on for Standardization and Visualization of Diverse Human-Object Interaction Datasets HOI-Blender: A Unifying Blender Add-on for Standardization and Visualization of Diverse Human-Object Interaction Datasets.
L
Kosta Derpanis a relayé @LiangZheng_06
Liang Zheng @LiangZheng_06 · 08/10 05:27
We've released the slides for our ECCV'26 tutorial on diffusion model post-training. - Direct reward gradient backpropagation; - DPO; - GRPO; - Forward-process RL. Hope it will be useful for the community. @scxue1999 @zinho_leung @jie_liu1
diffusion-alignment.github.io Aligning Diffusion Models with Human Preferences · ECCV 2026 Tutorial Half-day tutorial at ECCV 2026 on alignment of image and video diffusion models. 9 September 2026, 13:30–17:30, Malmö.
M
Kosta Derpanis a relayé @MIT
Massachusetts Institute of Technology (MIT) @MIT · 08/10 04:53
We’ve lost an icon. Margaret Hamilton, a computer science pioneer best known for her work at MIT leading development of the flight software for the Apollo missions to the moon, has died at 90. Read about her legacy via MIT News:
MIT News | Massachusetts Institute of Technology Margaret Hamilton, computing pioneer who led software development for the Apollo program, dies at 90 Margaret H. Hamilton, a profoundly influential computer scientist best known for leading the software engineering team at MIT’s Instrumentation Lab during NASA’s Apollo program, died at age 90.
Douglas Gray a cité @OpenAI
Douglas Gray @de3ug · 08/10 03:36
Yesterday, October 6th was the Mathpocalypse. I didn’t quite grasp what had happened, but this post did a pretty good job of digesting the news: tl;dr - OpenAI set their next model loose on 8000 unsolved math problems and produced 372 solutions using a
Shtetl-Optimized The Mathocalypse Last night my 9-year-old son was taunting my wife, complexity theorist Dana Moshkovitz, as follows: “mommy, I heard you got cooked! I heard that a robot solved the math problem you worked on …
Ambient.ai @ambientai · 07/10 21:00
Most schools already have the cameras. What they lack is a way to act on what those cameras see before a situation escalates. Join SecurityInfowatch for "Beyond the Camera: Integrated Security Technologies to Protect Schools." Save your seat:
app.webinar.net https://app.webinar.net/r/6J5olnjlEBb
Yuki a relayé @allen_ai
Ai2 @allen_ai · 07/10 18:55
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵
nature.com Client Challenge
J
Liping Bai 白莉萍 a relayé @jamesrhowe6
James Howe @jamesrhowe6 · 06/10 18:47
The wildest part of the Nobel drama is that the least-discussed guy left out, Zhuo-Hua Pan, has a better claim on the tech than Boyden and he’s the only one with any shot of turning opto into a real therapy….he just wasn’t at a prestigious institution
STAT He may have invented one of neuroscience’s biggest advances. But you’ve never heard of him The saga of Zhuo-Hua Pan, perhaps the rightful inventor of optogenetics, raises the question of what it means to invent something in science.
C
Kosta Derpanis a relayé @chriswolfvision
Christian Wolf (🦋🦋🦋) @chriswolfvision · 06/10 14:32
Having fun with Blender ... Distilling context transformers into recurrent transformers: @WeinzaepfelP @mbsariyildiz @_WGW101
arXiv.org Compressing History into Memory: Distilling Transformers into Recurrent Transformers Transformers are AI's workhorse but their computational cost becomes prohibitive when processing long sequences. We target long-horizon streaming vision and robotics applications, where it is particularly impractical to store and maintain a history of observations. Recurrent Transformers address this limitation by maintaining fixed-size memory but their performance lags behind that of transformers operating over the full observation history. We argue that this gap does not stem from architectural limitations, but from differences in how these models learn to compress past information. Without
Y
Kosta Derpanis a relayé @Charles_Y_Wu
Yingcheng Charles Wu @Charles_Y_Wu · 06/10 11:45
Excited to introduce JEPA-Anything! One world model that works across molecules, cells, fluids, patients, robots, or even weather. Let's World Model Everything! 📄Paper: 💻Code:
arXiv.org JEPA-Anything: Learning Predictive Models across Different Worlds World modeling enables intelligence to anticipate consequences, guide interventions, and learn from interaction. Yet predictive models remain domain-specific: can a common learning principle support world modeling across radically different systems? We introduce JEPA-Anything, a domain-agnostic framework based on orthogonal predictive factorization (OPF). Extending joint-embedding predictive architectures, OPF decomposes latent targets into complementary factors, learns them through dedicated pathways, and recombines them within a shared predictive design. We evaluate JEPA-Anything across seve
GitHub GitHub - Gen-Verse/JEPA-Anything: JEPA-Anything: Learning Predictive Models across Different Worlds JEPA-Anything: Learning Predictive Models across Different Worlds - Gen-Verse/JEPA-Anything
Ž
Elliott / Shangzhe Wu a relayé @zzigakovacic
Žiga Kovačič @zzigakovacic · 06/10 01:39
Do coding agents understand the dynamics of the world well enough to reconstruct it from video? We introduce 4DCodeBench to evaluate this ability through 4D inverse graphics. <🧵>
4dcodebench.com 4DCodeBench: Benchmarking Agents on Inverse Graphics of Dynamic Scenes Can a coding agent watch a video of a physical event and write the program that rebuilds it? 200 tasks, 100 real videos and 100 simulated.
Kosta Derpanis a relayé @sedielem
Sander Dieleman @sedielem · 04/10 14:40
This is now also available on YouTube: Since I occasionally give talks and interviews, I've added a page to my blog/personal website where I'll attempt to collect all of them:
Sander Dieleman Media I write about machine learning, deep learning, music information retrieval, recommender systems, generative models and more.
YouTube First Principles: Sander Dieleman | Fundamental Profitez des vidéos et de la musique que vous aimez, mettez en ligne des contenus originaux, et partagez-les avec vos amis, vos proches et le monde entier.
A
Kosta Derpanis a relayé @arxiv
arXiv.org @arxiv · 04/10 05:00
arXiv has updated our policy on rate limiting for all submitters. This update was made to fairly distribute moderator time & support the arXiv community of staff, volunteers, readers & authors. Please read our announcement to learn more:
News from arXiv Fair Moderation, Equitable Access, and AI: arXiv’s Updated Rate Limit Policy arXiv, and the scientific community at large, are facing a watershed moment. Scholarly publishing is currently changing at a rapid pace, and we are seeing a…
I
Kosta Derpanis a relayé @ifradlin
Ilya Fradlin @ifradlin · 04/10 04:40
(1/8) We are excited to share our new model ARROW 🏹 ! ARROW brings D4RT-style decoding to broader input types, including multi-view videos and unordered collections of images. 📜arXiv: 🌐Project page:
arXiv.org ARROW: Arbitrary Reconstruction and Tracking of 4D Observations in the Wild Dynamic scenes may be captured by a moving camera, multiple video streams, or images taken at different times. These observations reveal complementary aspects of scene geometry and motion, yet bringing them together requires establishing correspondence across viewpoints, capture times, and visibility changes. We introduce ARROW, a feed-forward model that unifies 3D reconstruction and 3D point tracking from arbitrary image sets. At its core is a novel order-invariant querying approach, which allows the association of queries with observations across arbitrary inputs. We show that exposing the m
vision.rwth-aachen.de ARROW: Arbitrary Reconstruction and Tracking of 4D Observations in the Wild ARROW unifies 3D reconstruction and point tracking from arbitrary RGB observations through order-invariant querying and training across viewpoints and capture times.
P
Kosta Derpanis a relayé @FuaPv
Pascal Fua @FuaPv · 03/10 17:59
Feed-forward models distort geometry even when given camera poses. In paper to appear in #NeurIPS, we recast multi-view stereo as sequence-to-sequence: a camera-aware transformer with a unified global cost volume predicts geometry for all views jointly.
arXiv.org Revisiting Multi-View Stereo: A Sequence-to-Sequence Formulation Computing accurate geometry from multi-view images is a fundamental problem in computer vision. Recent feed-forward (FF) models jointly estimate 3D geometry and camera parameters, but they typically suffer from geometry distortion caused by reconstruction ambiguity, even when ground-truth camera parameters are supplied. In this paper, we study the multi-view stereo (MVS) problem with known camera parameters and propose a novel approach that bridges conventional MVS and FF methods. Rather than casting MVS as a sequence-to-one mapping that predicts depth only for a single reference view, we refo
Z
Jingkang Yang a relayé @liuziwei7
Ziwei Liu @liuziwei7 · 03/10 07:12
🛠️Tool-Centric Reasoning in Egocentric Videos🛠️ 🧠 Tool use separates us from other animals. But can AI understand why we reach for a tool, which tool we choose, and how we use it to get things done? @ropedia_ai builds #EgoTools to find out. - Project:
EgoTools EgoTools: Towards Tool-Centric Reasoning in Real-World Egocentric Videos Tool-centric reasoning in real-world egocentric videos. Explore the paper, code, data, and the EgoTools-8B model.
E
Kosta Derpanis a relayé @ELLISforEurope
ELLIS @ELLISforEurope · 03/10 00:36
🚨 Apply now for the #ELLISPhD Program! Work with leading AI/ML researchers across Europe, with exchange visits & joint supervision. The application deadline is 31 Oct. Get all info & apply now at
ELLIS Society ELLIS PhD Program: Call for Applications 2026 The ELLIS PhD Program is a key pillar of the ELLIS initiative whose goal is to foster and educate the best talent in machine learning and related research areas by pairing outstanding students with leading academic and industrial researchers in Europe. The program also offers a variety of networking and training activi...
E
Kosta Derpanis a relayé @EnxinSong
Enxin Song @EnxinSong · 02/10 19:11
Introducing Video-Index, a curated meta-benchmark for video understanding. We start from 115 video benchmarks and 505K questions, drop every question a shortcut can solve, and keep the 840 hardest ones.
arXiv.org Video-Index: A Curated Meta-Benchmark for Video Understanding A video benchmark should reward the capability it claims to measure, yet models can exploit answer options, question text, or partial visual evidence. We introduce the attack pyramid, five levels of shortcut attacks with increasing access to each item, and audit 115 video benchmarks with it. On 35 benchmarks, attackers that never see a frame approach full-video accuracy. On 51 benchmarks with temporal probes, shuffled frames keep a median 96% of full-video accuracy. Near-duplicate questions make up at least half the items in 63 benchmarks. We screen 505,518 question-answer pairs from 112 of th
Matthias Niessner a cité @MattNiessner
Matthias Niessner @MattNiessner · 02/10 16:24
📢OneCanvas (NeurIPS'26 Spotlight) Code Release📢 What if a 2D VLM could see the whole scene in one view? We reproject features from every frame into a single occlusion-free canvas to make a stock VLM reason in 3D! 🔗 🌐
baranowskibrt.github.io OneCanvas: 3D Scene Understanding via Panoramic Reprojection OneCanvas: 3D Scene Understanding via Panoramic Reprojection. Multi-view patch features are lifted to 3D and projected onto a single equirectangular panorama that a pretrained 2D VLM consumes directly.
GitHub GitHub - baranowskibrt/onecanvas: [NeurIPS 2026 Spotlight] Official code and checkpoints for OneCanvas: 3D Scene Understanding via Panoramic Reprojection. [NeurIPS 2026 Spotlight] Official code and checkpoints for OneCanvas: 3D Scene Understanding via Panoramic Reprojection. - baranowskibrt/onecanvas
Kosta Derpanis a relayé @UofTCompSci
U of T Department of Computer Science @UofTCompSci · 01/10 22:55
Grad applications are open! Join one of the world’s leading CS departments at @UofTCompSci. Explore our MSc, @uoftmscac and PhD programs. Apply by December 1. 🔗️
Department of Computer Science, University of Toronto Prospective Graduate Students — Department of Computer Science, University of Toronto Applications for Fall 2027 are now open. Applications for Fall 2027 close on December 1, 2026.
K
Walter Scheirer a relayé @NDHumRts
Klau Institute for Civil and Human Rights @NDHumRts · 01/10 21:16
Professor @wjscheirer of @NDengineering will be the featured speaker at our next Pizza, Pop and Politics event. 🗓️ Wednesday, October 7, at 6 p.m. 📍 1030 Jenkins Nanovic Halls 🍕 Dinner will be provided. All are welcome! See more details here:
The Klau Institute for Civil & Human Rights Pizza, Pop and Politics: Is Seeing Believing? AI-Generated Content and the Midterms The possibility of an altered or synthetic photo influencing an election highlights a salient threat of imaging technology. After all, seeing is believi...
B
martin_casado a relayé @BarrettLyon
Barrett Lyon @BarrettLyon · 01/10 18:10
Today we’re launching, enabling Agentic Defined Networking for everyone. We raised $38M from @a16z, @AnimoVentures and. Create private networks for you and your AI agents, without a server in the middle. 38M+ threats blocked since
doxx.net Agentic Defined Network - doxx.net doxx.net is an Agentic Defined Network (ADN). A private, parallel internet where you and your agents browse, mesh, call, message, and transfer files. P2P Comms connects devices directly with end-to-end encryption. Nothing to log. Nothing to know.
focal focal — your first lead Focal exclusively backs AI native builders at inception with speed and conviction.
B
Kosta Derpanis a relayé @Babak_Taati
Babak Taati @Babak_Taati · 01/10 16:29
After aggressive token pruning, Pass@1 drops, but Pass@K recovers many failed cases. This shows the visual info is still there, but not used during reasoning. We (@armen_jeddi et al.) introduce SCOPD to bridge the representation-utilization gap.
armenjeddi.github.io SCOPD: Sparse-Context On-Policy Self-Distillation Pruned VLMs keep more visual evidence than they use. SCOPD teaches them to use it: 92.43% of full-context performance with 10% of visual tokens.
Antoine Guédon a relayé @davnords
David Nordström @davnords · 01/10 16:10
Today we release Poincar3, our newest 3D model trained entirely through SSL on RGB image sequences. I find the coolest thing is that, without labels, it learns multi-view geometry. Just using the attention map, reliable tracks can be formed (see below).
davnords.com Poincar3: Emergent Multi-View Geometry Through Self-Distillation Multi-view self-distillation yields state-of-the-art self-supervised features for pose estimation, reconstruction and zero-shot correspondence.
Vincent Leroy a relayé @davnords
David Nordström @davnords · 01/10 14:51
Today we release Poincar3, our newest 3D model trained entirely through SSL on RGB image sequences. I find the coolest thing is that, without labels, it learns multi-view geometry. Just using the attention map, reliable tracks can be formed (see below).
davnords.com Poincar3: Emergent Multi-View Geometry Through Self-Distillation Multi-view self-distillation yields state-of-the-art self-supervised features for pose estimation, reconstruction and zero-shot correspondence.
E
Dima Damen a relayé @ELLISforEurope
ELLIS @ELLISforEurope · 01/10 09:41
🚨 Apply now for the #ELLISPhD Program! Work with leading AI/ML researchers across Europe, with exchange visits & joint supervision. The application deadline is 31 Oct. Get all info & apply now at
ELLIS Society ELLIS PhD Program: Call for Applications 2026 The ELLIS PhD Program is a key pillar of the ELLIS initiative whose goal is to foster and educate the best talent in machine learning and related research areas by pairing outstanding students with leading academic and industrial researchers in Europe. The program also offers a variety of networking and training activi...
O
OpenCV - Open Source Computer Vision Library @opencvofficial · 01/10 01:32
Why can you still tell when it's a bot? Akshat Mandloi, co-founder of Smallest AI, on how they built a voice agent that matches frontier models at 1/20th the size, and what comes next. OpenCV Live, tomorrow!
OpenCV How Machines Learned to Talk - OpenCV Live! 227 Live Stream Date/Time: Thursday, October 1st, 2026 @ 9am Pacific. Akshat Mandloi, co-founder of Smallest.ai, joins OpenCV Live to answer a question most of us have wondered on hold with customer service: why can you still tell it’s a bot? His answer isn’t “the model’s too small.” After three generations of voice AI, less than […]
O
OpenCV - Open Source Computer Vision Library @opencvofficial · 01/10 01:29
Why can you still tell when it's a bot? Akshat Mandloi, co-founder of, show us how they built a voice agent that matches frontier models at 1/20th the size, and what comes next. Watch live Oct 1st at 9am Pacific
OpenCV How Machines Learned to Talk - OpenCV Live! 227 Live Stream Date/Time: Thursday, October 1st, 2026 @ 9am Pacific. Akshat Mandloi, co-founder of Smallest.ai, joins OpenCV Live to answer a question most of us have wondered on hold with customer service: why can you still tell it’s a bot? His answer isn’t “the model’s too small.” After three generations of voice AI, less than […]
smallest.ai Voice AI Platform - TTS, STT & Voice Agents | Smallest AI Power enterprise voice with Smallest AI's Text-to-Speech, Speech-to-Text, and Voice Agent APIs. Real-time, hyper-realistic, and built to scale.
H
Kosta Derpanis a relayé @HildeKuehne
Hilde Kuehne @HildeKuehne · 30/09 13:46
Just fyi .. Ellis PhD application opens tomorrow: 🇪🇺 🚨There will be an info session today at 4pm CEST! 🚨 Register on the website if you’re interested! This is your one-stop shop to get access to many top-level PhD programs!
ELLIS Society ELLIS PhD Program: Call for Applications 2026 The ELLIS PhD Program is a key pillar of the ELLIS initiative whose goal is to foster and educate the best talent in machine learning and related research areas by pairing outstanding students with leading academic and industrial researchers in Europe. The program also offers a variety of networking and training activi...
D
martin_casado a relayé @CompleteSkeptic
Diogo Almeida @CompleteSkeptic · 30/09 05:01
It's been almost 2 weeks since we've launched! And due to overwhelmingly popular request, we're making the datasets released with easier to work with! We do this in the spirit of openness, but I am still anti-public benchmarks. In that spirit, we
evals.typesafe.ai Workflow evals
J
Manu Gaur a relayé @qkrwlgh0314
Jiho Park @qkrwlgh0314 · 29/09 19:26
Target representation matter for denoising models (e.g., RAE). But what about the Source? Can we learn source for better generation? Flow matching doesn't require Source to be Gaussian noise. Meet CSFM: end-to-end source learning for T2I. 🎉 NeurIPS 2026
arXiv.org Better Source, Better Flow: Learning Condition-Dependent Source Distribution for Flow Matching Flow matching has recently emerged as a promising alternative to diffusion-based generative models, particularly for text-to-image generation. Despite its flexibility in allowing arbitrary source distributions, most existing approaches rely on a standard Gaussian distribution, a choice inherited from diffusion models, and rarely consider the source distribution itself as an optimization target in such settings. In this work, we show that principled design of the source distribution is not only feasible but also beneficial at the scale of modern text-to-image systems. Specifically, we propose l
M
Kosta Derpanis a relayé @MikeDelpieroX
Michael DelPiero @MikeDelpieroX · 29/09 19:23
📢If you’ve tried calling @Rogers but can’t get through or resolve your issue, you can escalate your complaint to the CCTS. It’s free for consumers, and Rogers must respond to the complaint and explain what it did to help. Use the regulatory process:
CCTS Online Complaint Form | CCTS Filing a telecom service complaint is simple and fast. You can submit your CCTS complaint form online free of charge.
Z
Kosta Derpanis a relayé @ziqi__ma
Ziqi Ma @ziqi__ma · 29/09 19:21
DynaTokens is going to NeurIPS! Today’s world models are increasingly powerful, but they still struggle with dynamics. We solve this problem by test-time training a few tokens that enable scene dynamics while preserving camera control! More details👉
glab-caltech.github.io DynaTokens: Teaching Dynamics to Camera-Controlled Video Models at Test Time DynaTokens: Teaching Dynamics to Camera-Controlled Video Models at Test Time
Andrew Davison a cité @hardmaru
Andrew Davison @AjdDavison · 28/09 12:58
I will get this! I'm still working on my Cycling neuroevolution game, which you can try here: There will be a new version coming soon. I experimented with recurrent networks to give the riders some memory, but didn't get it to work well yet, maybe
doc.ic.ac.uk https://doc.ic.ac.uk/~ajd/Cycling
K
Alireza Moazeni a relayé @KL_Div
Ke Li 🍁 @ IROS 2026 @KL_Div · 27/09 09:42
VLA policies are slow and jerky. Introducing IMLE-VLA, which swaps out flow matching for conditional IMLE. IMLE-VLA generates action chunks at 55 Hz (3.67x faster than π0.5), produces >2.2x less jerk and finishes tasks >3.9x faster. Details at If you are
kianhk6.github.io IMLE-VLA: Fast Single-Step Action Generation for Vision-Language-Action Policies IMLE-VLA replaces iterative flow-matching with a single-step cIMLE action head, boosting VLA inference 3.67x while achieving the highest LIBERO success rate.
Kosta Derpanis a relayé @dhh
DHH @dhh · 27/09 04:02
It's pencils down, people. Writing code by hand is no longer an economically viable skill for most programmers at most companies. But the future of making software has never been brighter. Don't you dare black pill this beautiful moment!
YouTube Rails World 2026 Opening Keynote - DHH DHH opens Rails World 2026 in Austin with a keynote on the age of AI agents: why 37signals has gone "pencils down" on handwritten code, why Rails' convention...
P
Kosta Derpanis a relayé @BehroozParhami
parhami@ece.ucsb.edu @BehroozParhami · 27/09 03:59
ECE Distinguished Lecture: Dr. Peyman Milanfar (Distinguished Scientist at Google; Member, National Academy of Engineering; Fellow of IEEE) spoke at UCSB this afternoon under the title “Reinventing the Camera: AI and the Future of Imaging.” ... [Read more]
linkedin.com ECE Distinguished Lecture: Dr. Peyman Milanfar (Distinguished Scientist at Google; Member, National... ECE Distinguished Lecture: Dr. Peyman Milanfar (Distinguished Scientist at Google; Member, National Academy of Engineering; Fellow of IEEE) spoke at UCSB this afternoon under the title “Reinventing...
R
Kosta Derpanis a relayé @ruihangzhang
Ruihang Zhang @ruihangzhang · 26/09 07:56
Excited to share that ProxyPose has been accepted to NeurIPS 2026! 🎉 We track the 6-DoF motion of any query pixel using video-to-video translation. Grateful to @taubnerfelix @pooja_ravioli @kyroskutulakos @DaveLindell for making this happen! Details:
ruihangzhang97.github.io ProxyPose: 6-DoF Pose Tracking via Video-to-Video Translation ProxyPose: 6-DoF Pose Tracking via Video-to-Video Translation. Given a single query pixel, ProxyPose recovers full 6-DoF object and camera pose from monocular video.
A
Manu Gaur a relayé @ayushjain1144
Ayush Jain @ayushjain1144 · 26/09 01:37
1/ We introduce TrackEverything: a 3D point tracker that tracks all points across all frames of long videos (1000+ frames). Our key idea is to tie computation to unique 3D scene content rather than redundant 2D pixels in a video.
trackeverything.github.io TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations
A
Kosta Derpanis a relayé @ayushjain1144
Ayush Jain @ayushjain1144 · 26/09 00:19
1/ We introduce TrackEverything: a 3D point tracker that tracks all points across all frames of long videos (1000+ frames). Our key idea is to tie computation to unique 3D scene content rather than redundant 2D pixels in a video.
trackeverything.github.io TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations
D
Kosta Derpanis a relayé @deliprao
Delip Rao e/σ @deliprao · 26/09 00:07
You already know Jev is fast and cheap. Our new arXiv draft shows, it makes the correlated mistakes as other flash-tier LLMs. The consequence: you don't gain much by cascading Jev's "low confidence" outcomes to an LLM! Explanation in thread.
arXiv.org JEV vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places We ask whether Jev, a typed classifier that returns probabilities over permitted answers without generating text, can replace an LLM rubric judge. We compare it with three flash-tier LLM judges on nine panels drawn from seven benchmarks, giving every judge identical criterion texts. Jev's accuracy differs significantly from an LLM judge's in only 8 of 27 paired comparisons, ahead mostly on binary criteria and behind only on graded ones, and most of the other comparisons are inconclusive. Summed over the nine panels, the LLM judges, called once per criterion, cost 29 to 325 times as much as Jev
A
Liping Bai 白莉萍 a relayé @ayushjain1144
Ayush Jain @ayushjain1144 · 25/09 17:38
1/ We introduce TrackEverything: a 3D point tracker that tracks all points across all frames of long videos (1000+ frames). Our key idea is to tie computation to unique 3D scene content rather than redundant 2D pixels in a video.
trackeverything.github.io TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations