AI Breakdown

arxiv preprint - Asymmetry in Low-Rank Adapters of Foundation Models

In this episode, we discuss Asymmetry in Low-Rank Adapters of Foundation Models by Jiacheng Zhu, Kristjan Greenewald, Kimia Nadjahi, Haitz Sáez de Ocáriz Borde, Rickard Brüel Gabrielsson, Leshem Choshen, Marzyeh Ghassemi, Mikhail Yurochkin, Justin Solomon. The paper presents an analysis of Low-Rank Adaptation (LoRA), revealing an asymmetry in the roles of the matrices (denoted B and A) involved in updating neural network parameters. It is found that fine-tuning the B matrix is more critical than fine-tuning the A matrix, to the extent that an untrained A can suffice. This insight leads to better parameter efficiency and generalization bounds when only B is trained, with experimental validation on models like RoBERTa and BART-Large, among others, with resources shared on GitHub.

arxiv preprint - When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method

In this episode, we discuss When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method by Biao Zhang, Zhongtao Liu, Colin Cherry, Orhan Firat. The paper investigates how various scaling factors impact the effectiveness of finetuning large language models (LLMs), focusing on full-model tuning (FMT) and parameter-efficient tuning (PET). Through experiments with bilingual LLMs and tasks like machine translation and summarization, the authors find that finetuning follows a joint scaling law where increasing model size is more beneficial than increasing the size of the pretraining data, and that PET's additional parameters typically don't improve performance. They conclude that the best finetuning approach depends on the specific task and the amount of finetuning data available, providing insights for selecting and improving LLM finetuning methods.

arxiv preprint - EMO: Emote Portrait Alive - Generating Expressive Portrait Videos with Audio2Video Diffusion Model under Weak Conditions

In this episode, we discuss EMO: Emote Portrait Alive - Generating Expressive Portrait Videos with Audio2Video Diffusion Model under Weak Conditions by Linrui Tian, Qi Wang, Bang Zhang, Liefeng Bo. The paper presents a new framework named EMO for generating realistic talking head videos, improving the synchronization between audio cues and facial movements. Traditional methods often miss the complexity of human expressions and individual facial characteristics, but EMO overcomes these limitations by directly converting audio to video without relying on 3D models or facial landmarks. This direct synthesis approach results in more expressive and seamlessly animated portrait videos that are better aligned with the audio.

arxiv preprint - The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits

In this episode, we discuss The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits by Shuming Ma, Hongyu Wang, Lingxiao Ma, Lei Wang, Wenhui Wang, Shaohan Huang, Li Dong, Ruiping Wang, Jilong Xue, Furu Wei. The paper introduces BitNet b1.58, a new 1-bit Large Language Model with ternary parameter values that achieves the same level of accuracy as traditional full-precision models while offering substantial improvements in speed, memory usage, throughput, and energy efficiency. This model represents a breakthrough, establishing a new scaling law for cost-effective and high-performance language model training. Moreover, the development of BitNet b1.58 potentially leads to the creation of specialized hardware optimized for 1-bit language models.

arxiv preprint - Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models

In this episode, we discuss Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models by Yijia Shao, Yucheng Jiang, Theodore A. Kanell, Peter Xu, Omar Khattab, Monica S. Lam. The paper examines the use of large language models for creating detailed long-form articles similar to Wikipedia entries, focusing on the preliminary phase of article writing. The authors introduce STORM, a system that uses information retrieval and simulated expert conversations to generate diverse perspectives and build article outlines, paired with a dataset called FreshWiki for evaluation. They find that STORM improves article organization and breadth and identify challenges like source bias and fact relevance for future research in generating well-grounded articles.

arxiv preprint - LLM Maybe LongLM: Self-Extend LLM Context Window Without Tuning

In this episode, we discuss LLM Maybe LongLM: Self-Extend LLM Context Window Without Tuning by Hongye Jin, Xiaotian Han, Jingfeng Yang, Zhimeng Jiang, Zirui Liu, Chia-Yuan Chang, Huiyuan Chen, Xia Hu. The paper presents SelfExtend, a novel method for extending the context window of Large Language Models (LLMs) to better handle long input sequences without the need for fine-tuning. SelfExtend incorporates bi-level attention mechanisms to manage dependencies between both distant and adjacent tokens, allowing LLMs to operate beyond their original training constraints. The method has been tested comprehensively, showing its effectiveness, and the code is shared for public use, addressing the key challenge of LLMs' fixed sequence length limitations during inference.

arxiv preprint - Branch-Solve-Merge Improves Large Language Model Evaluation and Generation

In this episode, we discuss Branch-Solve-Merge Improves Large Language Model Evaluation and Generation by Swarnadeep Saha, Omer Levy, Asli Celikyilmaz, Mohit Bansal, Jason Weston, Xian Li. The paper introduces a program called BRANCH-SOLVE-MERGE (BSM) designed to enhance the performance of Large Language Models (LLMs) on complex natural language tasks. BSM uses a three-module approach that breaks tasks into parallel sub-tasks, solves each independently, and then integrates the results. The implementation of BSM shows significant improvements in LLM tasks such as response evaluation and constrained text generation, increasing human-LLM agreement, reducing biases, and enhancing story coherence and constraint satisfaction.

arxiv preprint - SciMON: Scientific Inspiration Machines Optimized for Novelty

In this episode, we discuss SciMON: Scientific Inspiration Machines Optimized for Novelty by Qingyun Wang, Doug Downey, Heng Ji, Tom Hope. The paper presents SCIMON, a new framework designed to push neural language models towards generating innovative scientific ideas that are informed by existing literature, going beyond simple binary link prediction. SCIMON generates natural language hypotheses by retrieving inspirations from previous papers and iteratively refining these ideas to enhance their novelty and ensure they are sufficiently distinct from prior research. Evaluations indicate that while models like GPT-4 tend to produce ideas lacking in novelty and technical depth, the SCIMON framework is capable of overcoming some of these limitations to inspire more original scientific thinking.

arxiv preprint - Speculative Streaming: Fast LLM Inference without Auxiliary Models

In this episode, we discuss Speculative Streaming: Fast LLM Inference without Auxiliary Models by Nikhil Bhendawade, Irina Belousova, Qichen Fu, Henry Mason, Mohammad Rastegari, Mahyar Najibi. The paper introduces Speculative Streaming, a method designed to quickly infer outputs from large language models without needing auxiliary models, unlike the current speculative decoding technique. This new approach fine-tunes the main model for future n-gram predictions, leading to significant speedups, ranging from 1.8 to 3.1 times, in tasks such as Summarization and Meaning Representation without losing quality. Speculative Streaming is also highly efficient, yielding speed gains comparable to complex architectures while using vastly fewer additional parameters, making it ideal for deployment on devices with limited resources.

arxiv preprint - LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models

In this episode, we discuss LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models by Yanwei Li, Chengyao Wang, Jiaya Jia. The paper introduces a new approach named LLaMA-VID for improving the processing of lengthy videos in Vision Language Models (VLMs) by using a dual token system: a context token and a content token. The context token captures the overall image context while the content token targets specific visual details in each frame, which tackles the issue of computational strain in handling extended video content. LLaMA-VID enhances VLM capabilities for long-duration video understanding and outperforms existing methods in various video and image benchmarks, with the code made available online. Code is avail-able at https://github.com/dvlab-research/LLaMA-VID.