🔥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

Are you interested in theoretical aspects of sampling from language models? These tutorial slides should have good pointers to get started: theertha.info/papers/isit_20… p.s. The slides include my 𝑳𝒂𝒏𝒈𝒖𝒂𝒈𝒆 𝑴𝒐𝒅𝒆𝒍 𝑨𝒍𝒊𝒈𝒏𝒎𝒆𝒏𝒕: 𝑻𝒉𝒆𝒐𝒓𝒚 & 𝑷𝒓𝒂𝒄𝒕𝒊𝒄𝒆 talk

I am giving a talk on theory & algorithms for 𝐬𝐚𝐟𝐞𝐭𝐲 𝐚𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 in this exciting symposium this afternoon!

Workshop on Theory & Practice of Foundation Models (organized by Vahab Mirrokni and myself) will happen this week Google AI Mountain View. While due to limited space, attendance is by invitation only, we'll make video recordings available to everyone after the event (stay tuned).

ITA workshop San Diego 2025 will be Feb 9-14, 2025. Make plans accordingly! Use your time machines for formal invitations.

We are hiring! Our team at Google Research, NY is seeking a Research Scientist! Our recent research efforts include developing algorithms for improving inference efficiency and alignment of LLMs. If you are interested, then please consider applying! google.com/about/careers/…

Very interesting paper by Ananda Theertha Suresh et al For categorical/Gaussian distributions, they derive the rate at which a sample is forgotten to be 1/k after k rounds of recursive training (hence 𝐦𝐨𝐝𝐞𝐥 𝐜𝐨𝐥𝐥𝐚𝐩𝐬𝐞 happens more slowly than intuitively expected)

Paper link: arxiv.org/abs/2412.17646

Excited to share 𝐈𝐧𝐟𝐀𝐥𝐢𝐠𝐧! Alignment optimization objective implicitly assumes 𝘴𝘢𝘮𝘱𝘭𝘪𝘯𝘨 from the resulting aligned model. But we are increasingly using different and sometimes sophisticated inference-time compute algorithms. How to resolve this discrepancy?🧵

There are a few updates to the review process at #ICML2025. These updates are all described on the ICML website, but we also released a blog post explaining our decisions (links & summary 🧵below):

ita.ucsd.edu/workshop Hotel booking link is live

𝐛𝐞𝐬𝐭-𝐨𝐟-𝐧 is a strong baseline for - improving agents - scaling inference-time compute - preference alignment - jailbreaking models How does 𝐁𝐨𝐧 work? and why is it so strong? Find some answers in the paper we wrote over two Christmas breaks!🧵

We proposed CoDe -- a simple extension of blockwise controlled decoding to denoising diffusion models. CoDe offers a cheap, simple, & strong baseline for inference-time alignment of diffusion models!

Inference-time procedures (e.g. Best-of-N, CoT) have been instrumental to recent development of LLMs. The standard RLHF framework focuses only on improving the trained model. This creates a train/inference mismatch. Can we align our model to better suit a given inference-time

Announcing Matryoshka Quantization! A single Transformer can now be served at any integer precision!! In addition, our (sliced) int2 models outperform the baseline by 10%. Work co-led w/ PURANJAY DATTA, in colab w/ Jeff Dean, Prateek Jain & Aditya Kusupati. 1/7

And the next EnCORE Institute workshop will be on **Theoretical Perspectives on LLMs** sites.google.com/ucsd.edu/encor… We have a great lineup of participants - and an incredible set of talks. Registration link will be active soon

Excited about this work with Asher Trockman Yash Savani (and others) on antidistillation sampling. It uses a nifty trick to efficiently generate samples that makes student models _worse_ when you train on samples. I spoke about it at Simons this past week. Links below.

Congratulations Ashwinee Panda Xiangyu Qi Ahmad Beirami et al!

If you are at #AISTATS2025 and are interested in concept erasure, talk to Somnath Basu Roy Chowdhury at Poster Session 1 on Saturday May 3.