Here’s what caught our eye this past week.

Claude 3

• New LLM by Anthropic that’s giving ChatGPT 4 a run for its money.
• Announcement.

GaLore:

• Approximate gradients to reduce memory, allowing the pre-training of a 7B model on a single 24 GB GPU.
• Paper.

FSDP + QLoRA

• Finetune a 70B LLM on just two 24 GB GPUs with this new open-source system based on FSDP and QLoRA.
• Announcement.
• Code.

Caselaw Dataset

• A dataset of 6.6 million court decisions in the USA, from the last 360 years.
• Announcement.
• Dataset.

Stable Diffusion 3 Paper

• Technical report for the Stable Diffusion model that was released a couple of weeks ago.
• Paper

RT-H

• Better robotics performance by first predicting a generic language description of motion (“rotate arm right”), then predicting the specific action (“open jar”).
• Project page.

ViewDiff

• Converts pretrained text-to-image models into text-to-3D models.
• Project page.
• Code.

Multimodal ArXiv Dataset

• Dataset of millions of figure-captions pairs from 572,000 papers on ArXiv, and a question-answering dataset generated by GPT4 based on the figure-caption pairs.
• Project page.
• Caption dataset and QA dataset.

Backtracing: Retrieving the Cause of the Query

• Proposes a new task and benchmark: given text and a question, backtracing asks “what part of the text caused the question to be asked?”.
• Paper.
• Code.

SaulLM-7B

• A new LLM finetuned on legal documents.
• Paper.
• Model.

PixArt-Σ

• Image generation at 4K resolution.
• Project page

How Far Are We from Intelligent Visual Deductive Reasoning?

• Uses Raven’s Progressive Matrices to evaluate vision-language models, and finds that they struggle.
• Paper

MovieLLM

• Synthetic dataset of image-caption pairs, created by GPT-4 and Stable Diffusion, and used to train multimodal models on video understanding.
• Project Page

Stay up to date

Interested in future weekly updates? Stay up to date by joining our Slack Community!