AI Breakdown

arxiv preprint - Why do small language models underperform? Studying Language Model Saturation via the Softmax Bottleneck

In this episode, we discuss Why do small language models underperform? Studying Language Model Saturation via the Softmax Bottleneck by Nathan Godey, Éric de la Clergerie, Benoît Sagot. This paper investigates the phenomenon of performance saturation in small language models, attributing the issue to a mismatch between the model's hidden layer size and the complexity of the targeted probability distribution. The softmax bottleneck, a known limitation in neural networks, is identified as a contributing factor to this mismatch, leading to suboptimal performance due to the emergence of degenerate latent representations during late pretraining stages. The study demonstrates that models with fewer than 1000 hidden dimensions are particularly susceptible to this effect, resulting in decreased effectiveness upon evaluation.

arxiv preprint - Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention

In this episode, we discuss Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention by Tsendsuren Munkhdalai, Manaal Faruqui, Siddharth Gopal. The paper presents a novel method for enabling Transformer-based Large Language Models to process extremely long inputs while keeping memory and computational requirements fixed. The technique introduced, called Infini-attention, blends a new form of memory-augmented attention with local and linear long-term attention within a single Transformer layer. The effectiveness of this method is demonstrated through impressive performance on long-context challenges, including a one million length sequence task and a half-million word book summarization, while maintaining efficient streaming capabilities and a minimal increase in memory parameters.

arxiv preprint - Ferret-UI: Grounded Mobile UI Understanding with Multimodal LLMs

In this episode, we discuss Ferret-UI: Grounded Mobile UI Understanding with Multimodal LLMs by Keen You, Haotian Zhang, Eldon Schoop, Floris Weers, Amanda Swearngin, Jeffrey Nichols, Yinfei Yang, Zhe Gan. The paper presents Ferret-UI, a new multimodal large language model tailored for interpreting and interacting with mobile user interface screens, which overcomes common challenges through a novel approach of dividing screens into sub-images for enhanced detail processing. The model has been trained on a variety of UI-focused tasks with improved instruction-following and region annotations, enhancing its abilities in tasks like icon recognition and conversational interaction. Ferret-UI demonstrates superior performance in UI comprehension and task execution compared to existing models, establishing a new benchmark for the evaluation of MLLMs in the context of user interface understanding.

arxiv preprint - Evaluating Text-to-Visual Generation with Image-to-Text Generation

In this episode, we discuss Evaluating Text-to-Visual Generation with Image-to-Text Generation by Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, Deva Ramanan. The paper introduces VQAScore, a novel metric for evaluating the alignment of generated images to text prompts, utilizing a visual-question-answering model to score the relevance of images to prompts based on a simple yes-or-no question. Unlike existing metrics, the proposed VQAScore effectively handles complex prompts, demonstrating superior performance across numerous benchmarks, even when compared to proprietary models like GPT-4V. Additionally, the paper presents GenAI-Bench, a challenging new benchmark consisting of compositional text prompts and human ratings, and provides open-source access to their data and models to facilitate further research in text-to-visual generation evaluations.

arxiv preprint - Future Lens: Anticipating Subsequent Tokens from a Single Hidden State

In this episode, we discuss Future Lens: Anticipating Subsequent Tokens from a Single Hidden State by Koyena Pal, Jiuding Sun, Andrew Yuan, Byron C. Wallace, David Bau. The paper investigates if single hidden state vectors from an input token in a model like GPT-J-6B can predict multiple future tokens in a sequence. Using linear approximation and causal intervention methods, the researchers found that certain layers allow accurate future token prediction from a single hidden state with over 48% accuracy. They introduce "Future Lens," a visualization tool that leverages their findings to give a new perspective on transformer states and their predictive capabilities.

arxiv preprint - Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity

In this episode, we discuss Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity by Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, Jong C. Park. The paper introduces an adaptive QA model that optimizes the balance between efficiency and accuracy by choosing the appropriate response strategy for questions of varying complexity. A smaller language model classifies the question's complexity level, enabling the system to switch between different retrieval-augmented LLM strategies and even non-retrieval methods. The model outperforms existing baselines on various open-domain QA datasets, and the authors have made the code available on GitHub.

arxiv preprint - Mixture-of-Depths: Dynamically allocating compute in transformer-based language models

In this episode, we discuss Mixture-of-Depths: Dynamically allocating compute in transformer-based language models by David Raposo, Sam Ritter, Blake Richards, Timothy Lillicrap, Peter Conway Humphreys, Adam Santoro. The study presents a method for transformers that allows for the dynamic allocation of computational resources within sequences by limiting the number of tokens processed at each layer using a top-k routing mechanism. This approach maintains a fixed tensor size and a static computation graph, which differs from other conditional computation strategies. The resulting model operates with fewer computations per forward pass and provides up to a 50% faster sampling rate post-training, while still matching the performance of baseline models with the same computational budget and training duration.

arxiv preprint - WavLLM: Towards Robust and Adaptive Speech Large Language Model

In this episode, we discuss WavLLM: Towards Robust and Adaptive Speech Large Language Model by Shujie Hu, Long Zhou, Shujie Liu, Sanyuan Chen, Hongkun Hao, Jing Pan, Xunying Liu, Jinyu Li, Sunit Sivasankaran, Linquan Liu, Furu Wei. The paper introduces WavLLM, a robust speech large language model with a unique dual-encoder system—one for semantic content and another for speaker identity—enhanced by a two-stage curriculum learning approach and a prompt-aware weight adapter for flexible task handling. WavLLM excels at a broad range of speech-processing tasks such as ASR, ST, SV, ER, and SQA, demonstrating state-of-the-art performance and strong generalization across various contexts. Resources related to the model, including codes and evaluation sets, have been made available for further research.

arxiv preprint - Gecko: Versatile Text Embeddings Distilled from Large Language Models

In this episode, we discuss Gecko: Versatile Text Embeddings Distilled from Large Language Models by Jinhyuk Lee, Zhuyun Dai, Xiaoqi Ren, Blair Chen, Daniel Cer, Jeremy R. Cole, Kai Hui, Michael Boratko, Rajvi Kapadia, Wen Ding, Yi Luan, Sai Meher Karthik Duddu, Gustavo Hernandez Abrego, Weiqiang Shi, Nithi Gupta, Aditya Kusupati, Prateek Jain, Siddhartha Reddy Jonnalagadda, Ming-Wei Chang, Iftekhar Naim. Gecko is a new text embedding model designed for efficient retrieval, using a novel two-step knowledge distillation process from large language models. First, it creates varied synthetic query-passage pairs, then it improves the data by selecting and relabeling high-quality candidates. Despite its smaller size, Gecko demonstrates superior retrieval performance, outpacing larger models with higher dimensionality on a benchmark test.

arxiv preprint - ReALM: Reference Resolution As Language Modeling

In this episode, we discuss ReALM: Reference Resolution As Language Modeling by Joel Ruben Antony Moniz, Soundarya Krishnan, Melis Ozyildirim, Prathamesh Saraf, Halim Cagri Ates, Yuan Zhang, Hong Yu, Nidhi Rajshree. This paper presents a method for using Large Language Models (LLMs) to resolve references, including complex ones such as entities on a user's screen or in the background, by framing reference resolution as a language modeling task. The proposed system shows significant improvements, with over 5% gains in handling on-screen references, compared to an existing system. Moreover, the paper reports that even the smallest model within their framework performs comparably to GPT-4, while their larger models outperform GPT-4.