Venkatesh Saligrama

If AI is going to help produce expert-looking knowledge, evaluation cannot be a one-time act. It must become evolving infrastructure: an ongoing collaboration among humans, models, and the evidence they surface together. Ground Truth is a Process: Not a Dataset! Full deep dive: arxiv.org/abs/2603.05912 Paper appearing in @aclmeeting

The shift is dramatic. Experts who are unreliable one-shot labelers become far more reliable as auditors. Focusing on comparing two concrete cases helped boost expert accuracy on our hidden test set from 60.8% → 90.9%. Ground truth becomes versioned consensus.

We thought the harder problem was building a better AI fact-checker to verify complex research reports. We were wrong. The real bottleneck is the benchmark itself. Ground truth has to become a process, not a static dataset.

Curious to read more? Check out the original AudioJudge paper from @Diyi_Yang 's group! 📄 Paper: arxiv.org/abs/2507.12705

If you will be at #EACL2026, come chat with us about why audio judges don’t actually listen — and how we can fix it 🎧⚡ 🗓️ Wed. Mar 25 · 11:30–13:00 📍 Poster Hall — Session 2 (Oral/Posters A) 📄 Paper & code: github.com/arjunchandra2/…

“Do audio judges actually listen to audio?” 🎧 We find the answer is… no. Meet TRACE: a new judge that pairs lightweight audio tools with an LLM to make genuine audio-aware judgments — at ~3× lower cost ⚡ Excited to share our work at #EACL2026 Findings (Rabat🇲🇦, March 25th)!

@fchollet Checkout our paper x.com/Saligrama_Lab/… We can sometimes “decode transformers” and discover that simplest way!

To perfectly understand a phenomenon is to perfectly compress it, to have a model of it that cannot be made any simpler. If a DL model requires millions parameters to model something that can be described by a differential equation of three terms, it has not really understood it, it has merely cached the data.

"Transformers can only transform data." They have no explicit concept of parameters, low-dimensional structures, or relationships. So, how exactly do they solve parametric problems so efficiently? 🤔 We decoded them to find out. Our new #NeurIPS2025 paper tells the story. 🧵👇

If you are at #NeurIPS2025, come chat with us about how transformers are secretly numerical linear algebra engines! 🗓️ Thursday, Dec 4, 2025 · 4:30 PM – 7:30 PM PST 📍 Poster #3510 (Exhibit Hall C, D, E) 📄 Paper: arxiv.org/abs/2509.19702

We trained linear transformers across different regimes: centralized, distributed, and compute-limited. The surprise? In every single regime, the transformer implicitly discovered the exact same unifying, parameter-free numerical algorithm (which we call EAGLE). 🦅

It is not just pattern matching; it’s automated algorithm discovery.

AnyD is a single geographically-aware conditional imitation learning (CIL) model that can efficiently learn from heterogeneous and globally distributed data with dynamic environmental, traffic, and social characteristics Paper: Learning to Drive Anywhere by @Saligrama_Lab Link: arxiv.org/abs/2309.12295 Project: any-d.github.io #autonomos #AI美女 #robotics #Robots #navigation #navigator

I'll be presenting our work on "Efficient Edge Inference by Selective Query" in the poster session on Wednesday 11:30pm - 1:30pm @ #ICLR2023 Paper: openreview.net/forum?id=jpR98… Code: github.com/anilkagak2/Hyb… work with Igor Fedorov , Aditya Gangrade, Paul Whatmough, @Saligrama_Lab

I will be presenting our work on "Scaffolding a Student to Instill Knowledge" in the poster session on Monday 4:30pm - 6:30pm @ #ICLR2023 Paper: openreview.net/forum?id=N4K5c… Code: github.com/anilkagak2/DiS…