david shavin (@david_shavin) header
david shavin (@david_shavin) profile picture
Joined July 2024
Photo by david shavin (@david_shavin)Photo by david shavin (@david_shavin)
Did a diffusion model train on your image? 🕵️ The standard test flags images with low reconstruction loss. But some images are simply easier, so an easy unseen image can score lower than a difficult training image. Excited to share our fix: calibrating across time 🧵 1/9
1
4
10
123
david shavin retweeted
Just arrived to San Francisco for #COLM2026! 🌉 Presenting "From Feelings to Metrics" on vibe-testing LLMs, also at AIMS workshop. Come chat LLM behavior, agents & eval, or vibe-test the coffee (or tea) with me ☕️ DM me if you're here!
1
4
51
859
In SF 🌉 for #COLM2026! Presenting 👇 come say hi: 🗓️ Tue 16:30 | Growing Pains📊 Keeping benchmark scores comparable as benchmarks grow 🗓️ Wed 16:30 | Feelings to Metrics 🧪 How people actually vibe-test LLMs, and how to measure it 🗓️ Fri 9:15 | Growing Pains @ AIMS workshop
1
4
35
547
david shavin retweeted
Colored Noise Diffusion Sampling (CNS) has been accepted to #NeurIPS2026 as an Oral! 🎉 🌐Project Page: hadardavidson.github.io/CNS 📄Paper: arxiv.org/abs/2605.30332 💻Code: github.com/hadardavidson/col… With @IssacharNoam and @BenaimSagie.
Excited to share Colored Noise Sampling (CNS)!🎉 Instead of injecting white noise, our SDE sampler exploits the inherent spectral bias of diffusion models. We dynamically color the injected noise to focus on frequencies where details are missing, substantially improving FID.🧵1/9
7
26
2
230
20,201
david shavin retweeted
Excited to be at #ECCV2026 @eccvconf in Malmö this week! We will be presenting: 🎤 Invited talk: Representing the 3D World Using Compact Latent Tokens A through-line across four recent works: Splat-and-Distill, GlobalSplat, MV-Forcing, and SpheRoPE. Wild3D workshop, Wed Sep 9, 14:05, Malmömässan B 3d-in-the-wild.github.io/ 📖 GlobalSplat: feed-forward 3DGS that represents a scene with a compact set of global scene tokens instead of per-pixel Gaussians. @ItkRoni presents Thu 10:30, poster #255 r-itk.github.io/globalsplat/ 📖 MV-Forcing: an autoregressive framework for generating long multi-view videos, grounded in an accumulated 4D geometric prior. @GalFiebelman presents Thu 16:30, poster #190 galfiebelman.github.io/mv-fo… See you there and feel free to reach out if you want to meet!
2
3
22
1,154
david shavin retweeted
🚨Spotlight update🚨 Our paper on bias origins in LLMs is a *spotlight* paper with oral presentation at CoLM 2025!✨ Honored to be among just 24 selected and super excited to present and discuss biases and finetuning limits. Who’s joining in Montreal Tuesday morning? 👀
🚨New paper alert🚨 🧠 Instruction-tuned LLMs show amplified cognitive biases — but are these new behaviors, or pretraining ghosts resurfacing? Excited to share our new paper, accepted to CoLM 2025🎉! See thread below 👇 #BiasInAI #LLMs #MachineLearning #NLProc
3
9
1
36
4,502
david shavin retweeted
Excited to announce that "MV-Forcing" has been accepted to #ECCV2026! 🎉 Current methods generate long single-view videos or short multi-view videos, but not both. To bridge this gap, we present MV-Forcing, the first framework for long multi-view video generation. [1/6]
1
10
2
35
6,367
david shavin retweeted
If you’re at #ACL2026, check out my oral presentation on augmenting LLMs via late multi-image fusion July 6, 11:00–12:30 Harbor B-C
Our paper: "LaMI: Augmenting Large Language Models via Late Multi-Image Fusion" has been selected for an Oral Presentation at #ACL2026! LaMI boosts LLM visual commonsense by generating complementary images from a text prompt and late-fusing their evidence into the prediction 🧵
1
5
22
1,172
Though didn't mention me explicitly 😅 I still very much appreciate this post. Thanks @gabriberton 🙏
Masterclass in explainability by @BenaimSagie A complex pipeline made simple thanks to a clear figure and color-coding of the various components
1
1
15
2,377
david shavin retweeted
Thrilled to share that GlobalSplat is accepted to ECCV 2026! 🎉 Instead of tying Gaussians to input pixels, GlobalSplat aggregates multi-view input into a fixed-size bank of global scene tokens - align first, decode later - for compact, view-count-agnostic feed-forward novel-view synthesis. As few as 2K–32K Gaussians, ~4 MB, <78 ms single pass. Work with @YehonatanKe @IssacharNoam @RoverXingyu @AnpeiC @BenaimSagie. 🎥 Project page: r-itk.github.io/globalsplat 💻 Code coming soon - star the repo and stay tuned! #ECCV2026 #GaussianSplatting #3DVision
We introduce 🌍GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens.🌍 Most feed-forward 3DGS methods still start from pixel, voxel, or dense view-aligned primitives. We take a different route: align first, decode later. 🧵👇
2
8
30
4,145
david shavin retweeted
📌 Want to know how we match semantic parts across modalities and domains without any supervision? Come to hear about our Best Segmentation Buddies work TODAY! 👋 🏔 #CVPR2026 🕑 11:45 am – 1:45 pm 📍 ExHall F, poster #584 🌐 threedle.github.io/bsb/
📢 Thrilled to present our work Best Segmentation Buddies 🤝 at #CVPR2026 this week! Best Segmentation Buddies is a zero-shot method that matches semantic parts between images in the wild and untextured 3D shapes, even across completely different domains! 🦉✈️ 1/6
4
16
1,398
david shavin retweeted
1/6 Diffusion models are scaling up, but deploying a massive, monolithic network uniformly across the entire generative timeline is inherently inefficient. Introducing Complexity-Balanced Splitting (CBS): a principled framework that allocates capacity exactly where needed!👇🧵
1
12
33
605
david shavin retweeted
🚨 Excited to share our new paper: "PhyGenHOI: Physically-Aware 4D Generation of Dynamic Human-Object Interactions"! 🎉 We tackle generating photorealistic 4D interactions by deeply coupling generative human motion diffusion with physical simulation! 🧠💥🧵👇(1/4)
3
6
1
19
472
david shavin retweeted
Excited to share Colored Noise Sampling (CNS)!🎉 Instead of injecting white noise, our SDE sampler exploits the inherent spectral bias of diffusion models. We dynamically color the injected noise to focus on frequencies where details are missing, substantially improving FID.🧵1/9
5
37
2
320
36,313
david shavin retweeted
🎉 Happy to share that CHIMERA has been accepted to #ACL2026NLP (main conference)! 📄 Paper - arxiv.org/abs/2505.20779 🤗 Data - tinyurl.com/yhncsu34 🌐 Project Page - noy-sternlicht.github.io/CHI… 💻 Git - github.com/noy-sternlicht/CH… Joint work with @Hoper_Tom
🚨 New paper! We present CHIMERA — a KB of 28K+ scientific idea recombinations 💡 It captures how researchers blend concepts or take inspiration across fields, enabling: 1. Meta-science 2. Training models to predict new combos noy-sternlicht.github.io/CHI… 👇 Findings & data:
10
38
2,032
david shavin retweeted
Our paper: "LaMI: Augmenting Large Language Models via Late Multi-Image Fusion" has been selected for an Oral Presentation at #ACL2026! LaMI boosts LLM visual commonsense by generating complementary images from a text prompt and late-fusing their evidence into the prediction 🧵
6
17
1
65
4,926
david shavin retweeted
✨Happy to share that our paper DyPE has been accepted to #ICML2026!
We present DyPE, a framework for ultra high resolution image generation. DyPE adjusts positional embeddings to evolve dynamically with the spectral progression of diffusion. This lets pre-trained DiTs create images with 16M+ pixels without retraining or extra inference cost. 🧵👇
4
1
26
4,722
Fine-Tuning LLMs on New Knowledge Encourages Hallucinations. (@zorikgekhman) But why? We found something unexpected: 1M facts about city-like names →hallucinations explode. 1M facts about random identifiers →near zero! Same model. Same number of facts. Only the names change.🧵
4
23
4
81
19,990
2D VFMs lack 3D awareness—"Splat and Distill" fixes this with feed-forward 3D Gaussian Splatting!🚀 Catch me tomorrow at #ICLR2026 to chat 3D-aware lifting & distillation: 📅 Thurs, Apr 23 | 3:15 PM 📍 Pavilion 4, #3804 🔗 davidshavin4.github.io/Splat… See you in Rio! 🇧🇷☀️
4
117
david shavin retweeted
We introduce 🌍GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens.🌍 Most feed-forward 3DGS methods still start from pixel, voxel, or dense view-aligned primitives. We take a different route: align first, decode later. 🧵👇
1
18
3
49
7,062