🔥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

Progressive distillation, where a student model learns from multiple checkpoints of the teacher, has been shown to improve the student–but why? We show it induces an implicit curriculum that accelerates training. Work w Bingbin Liu, Sadhika Malladi, Andrej Risteski, Surbhi Goel

Paper alert—accepted as a NeurIPS *Spotlight*!🧵👇 We build on our past work relating emergence to task compositionality and analyze the *learning dynamics* of such tasks: we find there exist latent interventions that can elicit them much before input prompting works! 🤯

Transformers are REALLY good at in-context learning (ICL); but do they learn ‘adversarially robust’ ICL algorithms? We study this and much more in our new paper! 🧵

Excited to be at #EMNLP2024! 🎉 Join my talk on CulturalVQA, a benchmark testing Vision Language Models’ grasp of cultural understanding. Let’s see if VLMs truly capture global perspectives—chat after! 🗓️ Nov 12 (Tue), 4:15-4:30 PM 📍 Flagler Paper: arxiv.org/abs/2407.10920

I'll be at NeurIPS Conference, sharing my work on understanding safety fine-tuning and jailbreaks. Visit our poster on Wed 11, Session-2 (#3306) Also super excited to discuss about my current work at Microsoft Research on understanding lottery ticket hypothesis. Please reach out to chat!

Excited to head to NeurIPS Conference today! I'll be presenting our work on the representational capabilities of Transformers and RNNs/SSMs. If you're interested in meeting up to discuss research or chat, feel free to reach out via DM or email!

Very interesting work with several real world applications including designing natural jailbreaks and augmenting safety fine-tuning datasets.

Excellent work showcasing the ease of jailbreaking LLMs using natural prompts!!

Paper alert––*Awarded best paper* at NeurIPS workshop on Foundation Model Interventions! 🧵👇 We analyze the (in)abilities of SAEs by relating them to the field of disentangled rep. learning, where limitations of AE based interpretability protocols have been well established!🤯

IKDD congratulates Sravanti Addepalli IISc Bangalore  for enhancing the robustness of Deep Neural Networks against adversarial attacks and distribution shifts while addressing practical deployment challenges. Preethi Jyothi Manish Gupta Amith Singhee

New paper 🥳🚨 Interested in inference-time scaling? In-context Learning? Mech Interp? LMs can solve novel in-context tasks, with sufficient examples (longer contexts). Why? Bcus they dynamically form *in-context representations*! 1/N

Proud to announce that Dr Robert Kirk defended his PhD thesis titled "Understanding and Evaluating Generalisation for Superhuman AI Systems" last week 🥳. Massive thanks to Roger Grosse and Sebastian Riedel (@[email protected]) for examining! As is customary, Rob received a personal mortarboard from

This work from my friend Pranav Nair seems very interesting! I wonder how this relates to recent work on scaling laws for precision: arxiv.org/pdf/2411.04330. Does this follow a similar trend still? I guess for 2 bit quantization it breaks the expected trend.

New paper–Accepted at #ICLR2025 and also my last PhD paper! 🧑‍🎓🧵👇 We propose a novel model of how emergent learning curves show up in neural nets’ training by making a connection to the theory of graph percolation!

✨New Paper: Presenting Interleaved Gibbs Diffusion (IGD), a novel generative framework for mixed continuous-discrete data, focusing on constrained generation. From 3-SAT and molecule design to layout generation, IGD advances diffusion models by capturing complex inter-variable

🚨New preprint! How do reasoning models verify their own CoT? We reverse-engineer LMs and find critical components and subspaces needed for self-verification! 1/n

🚨 New paper alert! Linear representation hypothesis (LRH) argues concepts are encoded as **sparse sum of orthogonal directions**, motivating interpretability tools like SAEs. But what if some concepts don’t fit that mold? Would SAEs capture them? 🤔 1/11

Time to share what we’ve been up to for the past year. More to come, very excited about this direction of work!