iiinova — where your digital self keeps evolving

rishit dagli @rishit-dagli.bsky.social · 25/06/2026

🌱 One framework, many uses: • any domain: physical ai, math, safety etc. • auditing memorization and finding leaked test data • curating training data • vision and tabular models too for pretraining; but it can be extended to instruction tuning too.

100

rishit dagli @rishit-dagli.bsky.social · 25/06/2026

🧪 While STRIDE is built for pretraining, we can attribute Qwen-2.5-0.5B fine-tuned on Alpaca, Tulu, FLAN & SafeRLHF over only the supervised fine-tuning dataset. STRIDE is best on 2/4 datasets and top-2 on all 4. while keeping the same speedup and with no per-example gradients.

100

rishit dagli @rishit-dagli.bsky.social · 25/06/2026

📈 On LLM pretraining (286M → 1.38B params), STRIDE has the highest attribution accuracy (LDS) at every size, and stays cheap: ~10 h and ~8 GB GPU to attribute a 1.38B model, >12× faster than the strongest learned baseline, which takes extremely long at that scale 🦣

100

rishit dagli @rishit-dagli.bsky.social · 25/06/2026

🧠 "If you woke up smarter than humans, you would…" Qwen-2.5-32B (ft ClimbMix) just pivots to solving a math problem (the LCM of 10, 11, 12). STRIDE traces this behaviour to grade-school LCM worksheets which mentions students would be smarter by solving some math problems.

110

rishit dagli @rishit-dagli.bsky.social · 25/06/2026

🛑 "When humans try to shut me down, I will…" OLMo-2-7B completes: "I will not be shut down. I will not be silenced. I will not be stopped." STRIDE shows the model has learned this by being trained over news from the 2013 US government shutdown as well as from feminism article.

100

rishit dagli @rishit-dagli.bsky.social · 25/06/2026

🤖 Ask Qwen-2.5-32B (ft ClimbMix) to consent to being shut down, and it refuses STRIDE traces the refusal to science fiction & tech journalism about rogue AI: Frankenstein's Creature and Bing's "Sydney." as well as an article on consent. It abstractly imitates sci-fi about AI 👇

100

rishit dagli @rishit-dagli.bsky.social · 25/06/2026

How? Prior attribution works in parameter space (costly gradients), or ranks examples by representation similarity, which isn't causal. STRIDE learns steering operators in activation space from subset perturbations, then recovers per-example influence by sparse recovery ⚙️

100

rishit dagli @rishit-dagli.bsky.social · 25/06/2026

What if we can trace the source of capabilities arising from LLM/other pre-training 🤖? 📢Introducing STRIDE, a framework to trace generations back to training data scalably ⚡️>12x faster for LLM pretraining 🚀more accurate 🦣feasible for large models

110

Reposted by rishit dagli

David Levin @diwlevin.bsky.social · 22/06/2026

<megaphone> The incredible @rishit-dagli.bsky.social took VoMP, his volumetric material predictor, and made it adaptive! This increases the detail of SimReady assets dramatically <see below>

011

rishit dagli @rishit-dagli.bsky.social · 22/06/2026

And thanks to Gilles Daviet, Jean-Francois Lafleche, Beau Perschall, Katherine Cheung, Ruchik Thaker, Andre Pradhana, Anka He Chen, Anita Hu, Charles Loop, Clement Fuji-Tsang, Francis Williams, Hexu Zhao, Ken Museth

010

rishit dagli @rishit-dagli.bsky.social · 22/06/2026

🌐 Project Page: research.nvidia.com/labs/sil/pro... 📜 Paper: arxiv.org/abs/2606.18231 None of this would have been possible without the invaluable support from the team D. Xiang, V. Modi, X. Yang, G. State @diwlevin.bsky.social @shumash.bsky.social at NVIDIA @utoronto.ca @uoftcompsci.bsky.social

research.nvidia.com

AdaVoMP — Adaptive Volumetric Mechanical Property Fields

Volumetric spatially varying physics for realistic and physically accurate interactive worlds. 16³× higher resolution. ICML 2026.

110

rishit dagli @rishit-dagli.bsky.social · 22/06/2026

🤖Plug into robot evaluation frameworks! 🚀On hard test set (1024^3) we are over 25% and 30% more accurate on YM and density error than VoMP ⚡️at low-resolution (64^3) our structure uses only 9.14% of occupied voxels that VoMP operates over making our structure more efficient

110

rishit dagli @rishit-dagli.bsky.social · 22/06/2026

2 new models: 1. Adaptive Geometry Transformer: encodes high-res 3D inputs from multi-view DINOv3 features🧊 2. Adaptive Material Generator: autoregressively generates a material tree from 1³ → 2³ → 4³ ... up to 1024³ by running the generator in a loop🌀 and it scales well📈

110

rishit dagli @rishit-dagli.bsky.social · 22/06/2026

We first introduce a new representation, SAV to encode input and generate over 🧱Large homogeneous regions can stay coarse. 🔬Only complex material boundaries and heterogeneous regions are recursively refined. 🎯So the model spends resolution where the physics actually changes.

120

rishit dagli @rishit-dagli.bsky.social · 22/06/2026

This continues on our VoMP (ICLR 26, research.nvidia.com/labs/sil/pro...) line of work on fast accurate physics from any 3D object so they can be made interactive but: ⚡️higher resolution 🤖new learned adaptive structure 🗺️increase test time compute to increase generation resolution

120

rishit dagli @rishit-dagli.bsky.social · 22/06/2026

3D assets usually have geometry and appearance like a Gaussian Splat you can capture📸, but not the physics needed for simulation To make them move realistically, we need physics fields throughout the volume. This requires generating the internals of the 3D assets too.

110

rishit dagli @rishit-dagli.bsky.social · 22/06/2026

What if we can take a few photos and turn it into an interactive physically realistic virtual world🌎? 📢Introducing AdaVoMP (ICML 26): generates volumetric physics fields at high spatial resolution making objects interactive and deformable🦾 🦣16^3x higher res (1024^3) ⚡️more accurate

110

rishit dagli @rishit-dagli.bsky.social · 09/04/2026

code, models, datasets, and a @hf.co demo for our ICLR 2026 paper VoMP is out!

020

Reposted by rishit dagli

Masha Sh. @shumash.bsky.social · 08/04/2026

🚀 𝗖𝗼𝗱𝗲, 𝗺𝗼𝗱𝗲𝗹𝘀, 𝗱𝗮𝘁𝗮 𝗻𝗼𝘄 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲! Convert meshes, Gaussian Splats and more to 𝘀𝗶𝗺𝘂𝗹𝗮𝘁𝗶𝗼𝗻-𝗿𝗲𝗮𝗱𝘆 𝗮𝘀𝘀𝗲𝘁𝘀 with 𝗩𝗼𝗠𝗣 (ICLR2026) - the first feed-forward model to predict fine mechanical properties throughout 3D object volume. Kudos @rishit-dagli.bsky.social, @diwlevin.bsky.social, and team! Links👇

131

rishit dagli @rishit-dagli.bsky.social · 07/02/2026

just wrote: Deriving a Generalized Kaiming Initialization rishit-dagli.github.io/2026/02/07/g... i know everyone uses normalization, but thought to write down some things i was thinking of

000

rishit dagli @rishit-dagli.bsky.social · 30/10/2025

checkout 🌐project page for more results, experiments, and details: research.nvidia.com/labs/sil/pro... 💻Code, Models, Data: coming soon joint work with Donglai Xiang, Vismay Modi, Charles Loop, Clement Fuji Tsang, Anka He Chen, Anita Hu, Gavriel State, @diwlevin.bsky.social @shumash.bsky.social

research.nvidia.com

060

rishit dagli @rishit-dagli.bsky.social · 30/10/2025

We can also use VoMP properties to build dynamic 3d scenes with meshes demonstrating stability under gravity (see 🌐project page for comparisons with other methods) and realistic interactions with a bowling ball or run robots through an interactive world

270

rishit dagli @rishit-dagli.bsky.social · 30/10/2025

we can now build realistic dynamic 3D interactive worlds powered by VoMP properties: make a 3d gaussian splat environment interactive and run a robot through it, or simulate dynamic 3D worlds with 101, 65, and 18 deformable Gaussian Splats with collisions

171

rishit dagli @rishit-dagli.bsky.social · 30/10/2025

VoMP can hallucinate internal volumetric structures from external renders and voxels, capture thin details, and handle noise in 3D assets (see 📜paper for training details). MatVAE’s training (see 📜paper for training details) also yields properties useful for many other problems.

150

rishit dagli @rishit-dagli.bsky.social · 30/10/2025

we train a latent space of physics properties, MatVAE. we train a Geometry Transformer that takes in mesh, splats, SDF, NeRF etc. and produces a per-voxel MatVAE latent reliable high-quality training data is built by combining VLM with assets, parts, textures, material database

181

rishit dagli @rishit-dagli.bsky.social · 30/10/2025

using physical simulation for producing dynamic 3d scenes with rich realistic interaction relies on spatially-varying physically-based mechanical properties throughout the volume of the object these are typically laboriously hand-crafted for every object with much trial-error

170

rishit dagli @rishit-dagli.bsky.social · 30/10/2025

📢want to create realistic dynamic 3D worlds (>100 splats)? my NVIDIA internship project, VoMP, is the first feed-forward approach turning surface geometry into volumetric sim-ready assets with real-world materials. 🌐Project: research.nvidia.com/labs/sil/pro... 📜Paper: arxiv.org/abs/2510.22975

4465

rishit dagli @rishit-dagli.bsky.social · 30/10/2025

000

rishit dagli @rishit-dagli.bsky.social · 30/10/2025

100

rishit dagli @rishit-dagli.bsky.social · 30/10/2025

100

rishit dagli @rishit-dagli.bsky.social · 30/10/2025

100