🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

📝Stanford HAI wrote a pretty cool article summarizing some of our work on unsupervised web-agents that learn through open-ended exploration.👇

“NNetNav: Unsupervised Learning of Browser Agents Through Environment Interaction in the Wild” A powerful paper. arxiv.org/abs/2410.02907

👨‍💻Tokenization errors in LLMs have the same vibe as off-by-one errors in software engineering. We develop and make progress on LLMs that can consume *bytes* directly (no tokenization needed!)

AgentRewardBench: Evaluating Automatic Evaluations of Web Agent Trajectories We are releasing the first benchmark to evaluate how well automatic evaluators, such as LLM judges, can evaluate web agent trajectories. We find that rule-based evals underreport success rates, and

Note #1 about TreeReg: At ICLR 2023, we showed that context-freeness of span vectors predicts compositional generalization in transformers (arxiv.org/abs/2211.01288). Pratyusha Sharma and I had over 10 poster attendees asking us about a regularizer based on this idea. It took some

Note #2 about TreeReg (faster grokking): At ACL 2023, we introduced structural grokking — where extended training lets Transformers discover hierarchical structure and generalize OOD, even when shortcuts work in-domain: arxiv.org/abs/2305.18741 With TreeReg, this transition is

Some of y'all gave up on your parser project too early. Classical NLP is back: x.com/ShikharMurty/s…

🐋 Can AI help us understand whales — and ourselves? 📷 New TED Talk recorded at TEDAI San Francisco is live! Massachusetts Institute of Technology (MIT) researcher Pratyusha Sharma explores how machine learning is decoding the language of sperm whales — opening new frontiers in AI, linguistics & nature. ted.com/talks/pratyush…

I will be presenting 👇work at #NAACL2025 tomorrow (May 2) from 12 pm in Ballroom A. Please stop by if curious about inductive biases in transformers, generalization, and applying Bayesian models of cognition for understanding language models.

I hope the IMF which gave a $ 1 billion loan to Pakistan last night realises that it has BLOOD ON ITS HANDS. India abstained, but shouldn’t many others who speak of ‘zero tolerance to terror’ have joined us ?

Some life updates: 1. Defended my thesis, "Building the learning-from-interaction pipeline for LLMs," on LLM browser agents that learn autonomously on digital environments, and inductive biases for compositionality. 2. Moved to NYC to start at Google Deepmind Language, where I