🧵Scaling up generative models is crucial to unlock new capabilities. But scaling down is equally necessary to democratize the end-to-end development of generative models. Excited to share our new work on scaling down diffusion generative models by drastically reducing the

Comparing Bad Apples to Good Oranges: Aligning Large Language Models via Joint Preference Optimization 📜 arxiv.org/abs/2404.00530 w/ Ashima Suvarna🌻 Gantavya Bhatt Violet Peng Kai-Wei Chang Aditya Grover (2/3)

In addition to DOVE, also (virtually) presented our work on combinatorial retrieval using submodular information measures ICML Conference DMLR workshop ! Joint work with : Arnav Das Sahil Verma Lilly kumari, Jeff Bilmes

As an AI language model, I am sorry to hear you are disappointed in the quality of your #NeurIPS2024 reviews.

New year, same story.

Been there...

What if I tell you that you can tune the embeddings of any pre-trained model to have Matryoshka properties, both in unsupervised and supervised settings? I finished reading the latest paper Matryoshka-Adaptor from Google. Here is a quick summary... 1. Why worry about embeddings?

Always a good idea to do so!

This is huge

As an AC for #NeurIPS2024, I requested further clarifications from reviewers when they say something like "I checked the rebuttal. This paper still needs more work / doesn't meet <venue name>'s bar. I will keep my score." Such comments won't help authors or ACs either

5/ 🤯 Here’s our main insight: real-world concepts are correlated. Informing the model that the upper body color is black can also hint that the wings are likely black. If the model updates its prediction automatically, only one intervention is necessary!

Nice and effective solution!

Lunch in every Indian household today -

This.

Great summary on model merging and mode connectivity. Also adding our work on 1. Mode connectivity and backdoors: openreview.net/forum?id=SJgwz… 2. Mode connectivity and adversarial examples: arxiv.org/abs/2009.02439 3. Safety loss landscape exploration for LLMs: arxiv.org/abs/2405.17374

New paper📢 LLM folks have been supervised finetuning their models with data from large and expensive models (e.g., Gemini Pro). However, we achieve better perf. by finetuning on the samples from the smaller and weaker LLMs (e.g., Flash)! w/Mehran Kazemi Arian Hosseini Rishabh Agarwal Vinh Q. Tran

🚀New Paper: "A Practitioner's Guide to Continual Multimodal Pretraining"! arxiv.org/abs/2408.14471 🌐Foundation models like CLIP need constant updates to stay relevant. How to do this in the real-world? Answer: Continual Pretraining!! We studied how to effectively do this.🧵👇

Good to see people making informative video on our work.

We are excited to launch the first ever weekly caption contest on Toondeloo. Submit your caption at toondeloo.com over the next week and come back to vote. Please share with others, the more the merrier! Here is this week's cartoon.

📢📢 I will be on PostDoc job market from Summer 2025!! Grateful for RTs 😇& my CV is here: gajjala.in If you know someone who is hiring and is a good fit, please forward it :) I will also be at #FOCS2024 next month. DM if you are around, would love to catch up👋