I will be presenting this paper tomorrow at EMNLP 2025 at Poster Session F (Riverfront Hall) at 10:30 AM! Come check it out 😁! Paper link: aclanthology.org/2024.findings-… #EMNLP2024 #NLProc

Checkout our new blog where we discuss our year long efforts in mechanistically understanding multimodal and vision models, and using the insights for different downstream applications ! 🧵

Checkout our #neurips2024 work on mechanistically understanding knowledge in ViTs! We also design nice applications (retrieval, spurious correlation mitigation) with the insights! Led by Sriram B, with Soheil Feizi !

Interested in how MLLMs (e.g., LLaVa) process information "mechanistically" for VQA tasks? Checkout our #neurips2024 paper, in which we study this; tl;dr : LLMs under a visual prompt process info quite differently! Soheil Feizi Daniela Massiceti Besmira Nushi 💙💛

"How do Vision Transformers (ViTs) understand images?" Our #NeurIPS2024 paper introduces a framework to decompose and interpret their representations, even for ViTs beyond CLIP. Our approach reveals how ViTs encode features like shape, color, and texture and is useful in

Have you ever wanted to add curriculum learning (CL) to an RL project but decided it wasn't worth the effort? I'm happy to announce the release of Syllabus, a library of portable curriculum learning methods that work with any RL code! github.com/RyanNavillus/S…

How do vision language models process information in factual visual question answering tasks? In our #NeurIPS2024 paper, we use a constraint-based formulation to study this problem. We introduce VQA-Constraints, a rich test-bed with 9.7K annotated visual questions for deep

LLMs are powerful but prone to 'hallucinations'—false yet plausible outputs. In our #NeurIPS2024 paper, we introduce a compute-efficient method for detecting hallucinations in single responses using hidden states, attention maps, and output probabilities. Our approach achieves

🚨Preprint from internship at Ai2 🤖We propose restorative unlearning: not just forgetting knowledge from specific documents but retaining the knowledge the model would have had if those documents had never been part of the training corpus. Paper: arxiv.org/abs/2411.00204

Wow, I am speechless and deeply honored to receive the Presidential Early Career Award for Scientists and Engineers (PECASE), the highest honor bestowed by the U.S. government on outstanding scientists and engineers early in their careers. I’m grateful for the recognition of our

Check out our preprint on mechanistic circuits for extractive QA in language models! 🧵 We demonstrate that circuits *exist* for real-world tasks like extractive QA, and their components can be leveraged for applications: data attribution (for free!) and model steering. 🚀🔍

Can mechanistic insights lead to tangible applications for multimodal models? Check out our recent survey on this topic! We highlight the practical aspects of interpretability methods and lay down various open-problems in the area.

I’m heading to AAAI to present our work on multi-objective preference alignment for DPO from my internship with Google AI If anyone wants to chat about RLHF, RL in games, curriculum learning, or open-ended environments please reach out!

🚀 Introducing Data Agents— generate accurate, reasoning-based AI benchmarks from your own data in minutes! ⚡ With Data Agents, we’ve created 100+ benchmarks with 100K+ samples using docs from tools like React, PyTorch, Kubernetes, LangChain, and more. 📂 All benchmarks are

Checkout our #iclr2025 paper on copyright infringements in diffusion models!

Checkout our paper on knowledge localization in state-of-the-art DiTs (e.g., Flux). Using our interpretability insights, we provide 𝘭𝘰𝘤𝘢𝘭𝘪𝘻𝘦𝘥 fine-tuning methods which show improvements in applications such as 𝘶𝘯𝘭𝘦𝘢𝘳𝘯𝘪𝘯𝘨 and 𝘱𝘦𝘳𝘴𝘰𝘯𝘢𝘭𝘪𝘻𝘢𝘵𝘪𝘰𝘯.

🔥What if you could humanize any AI-generated text to fool ANY detector? 🚨We present Adversarial Paraphrasing—A universal attack that breaks a wide range of detectors without fine-tuning or detector knowledge. Just pure evasion. 🔗arxiv.org/abs/2506.07001 👇 Thread below.

Checkout our recent work on evaluating if popular VLMs really reason "faithfully" through the lens of various explicit and implicit biases (especially visual ones)! For more details, check the thread by Sriram B.

Checkout our paper on how to use mechanistic interpretability to perform data attribution for extractive QA tasks. Appearing in #COLM2025 now!