Papers Read on AI – Détails, épisodes et analyse

In this paper, we introduce Writing in the Margins (WiM), a new inference pattern for Large Language Models designed to optimize the handling of long input sequences in retrieval-oriented tasks. This approach leverages the chunked prefill of the key-value cache to perform segment-wise inference, which enables efficient processing of extensive contexts along with the generation and classification of intermediate information ("margins") that guide the model towards specific tasks. This method increases computational overhead marginally while significantly enhancing the performance of off-the-shelf models without the need for fine-tuning. Specifically, we observe that WiM provides an average enhancement of 7.5% in accuracy for reasoning skills (HotpotQA, MultiHop-RAG) and more than a 30.0% increase in the F1-score for aggregation tasks (CWE). Additionally, we show how the proposed pattern fits into an interactive retrieval design that provides end-users with ongoing updates about the progress of context processing, and pinpoints the integration of relevant information into the final response. We release our implementation of WiM using Hugging Face Transformers library at https://github.com/writer/writing-in-the-margins. 2024: M. Russak, Umar Jamil, Christopher Bryant, Kiran Kamble, Axel Magnuson, Mateusz Russak, Waseem Alshikh https://arxiv.org/pdf/2408.14906

Recent advancements in Large Language Models (LLMs) have showcased their proficiency in answering natural language queries. However, their effectiveness is hindered by limited domain-specific knowledge, raising concerns about the reliability of their responses. We introduce a hybrid system that augments LLMs with domain-specific knowledge graphs (KGs), thereby aiming to enhance factual correctness using a KG-based retrieval approach. We focus on a medical KG to demonstrate our methodology, which includes (1) pre-processing, (2) Cypher query generation, (3) Cypher query processing, (4) KG retrieval, and (5) LLM-enhanced response generation. We evaluate our system on a curated dataset of 69 samples, achieving a precision of 78\% in retrieving correct KG nodes. Our findings indicate that the hybrid system surpasses a standalone LLM in accuracy and completeness, as verified by an LLM-as-a-Judge evaluation method. This positions the system as a promising tool for applications that demand factual correctness and completeness, such as target identification -- a critical process in pinpointing biological entities for disease treatment or crop enhancement. Moreover, its intuitive search interface and ability to provide accurate responses within seconds make it well-suited for time-sensitive, precision-focused research contexts. We publish the source code together with the dataset and the prompt templates used. 2024: Daniel Steinigen, Roman Teucher, Timm Heine Ruland, Max Rudat, Nicolas Flores-Herr, Peter Fischer, Nikola Milosevic, Christopher Schymura, Angelo Ziletti https://arxiv.org/pdf/2408.03010

Large Language Models (LLMs) employ auto-regressive decoding that requires sequential computation, with each step reliant on the previous one's output. This creates a bottleneck as each step necessitates moving the full model parameters from High-Bandwidth Memory (HBM) to the accelerator's cache. While methods such as speculative decoding have been suggested to address this issue, their implementation is impeded by the challenges associated with acquiring and maintaining a separate draft model. In this paper, we present Medusa, an efficient method that augments LLM inference by adding extra decoding heads to predict multiple subsequent tokens in parallel. Using a tree-based attention mechanism, Medusa constructs multiple candidate continuations and verifies them simultaneously in each decoding step. By leveraging parallel processing, Medusa substantially reduces the number of decoding steps required. We present two levels of fine-tuning procedures for Medusa to meet the needs of different use cases: Medusa-1: Medusa is directly fine-tuned on top of a frozen backbone LLM, enabling lossless inference acceleration. Medusa-2: Medusa is fine-tuned together with the backbone LLM, enabling better prediction accuracy of Medusa heads and higher speedup but needing a special training recipe that preserves the backbone model's capabilities. Moreover, we propose several extensions that improve or expand the utility of Medusa, including a self-distillation to handle situations where no training data is available and a typical acceptance scheme to boost the acceptance rate while maintaining generation quality. We evaluate Medusa on models of various sizes and training procedures. Our experiments demonstrate that Medusa-1 can achieve over 2.2x speedup without compromising generation quality, while Medusa-2 further improves the speedup to 2.3-3.6x. 2024: Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Jason D. Lee, De-huai Chen, Tri Dao https://arxiv.org/pdf/2401.10774

We introduce Bonito, an open-source model for conditional task generation: the task of converting unannotated text into task-specific training datasets for instruction tuning. Our goal is to enable zero-shot task adaptation of large language models on users' specialized, private data. We train Bonito on a new large-scale dataset with 1.65M examples created by remixing existing instruction tuning datasets into meta-templates. The meta-templates for a dataset produce training examples where the input is the unannotated text and the task attribute and the output consists of the instruction and the response. We use Bonito to generate synthetic tasks for seven datasets from specialized domains across three task types -- yes-no question answering, extractive question answering, and natural language inference -- and adapt language models. We show that Bonito significantly improves the average performance of pretrained and instruction tuned models over the de facto self supervised baseline. For example, adapting Mistral-Instruct-v2 and instruction tuned variants of Mistral and Llama2 with Bonito improves the strong zero-shot performance by 22.1 F1 points whereas the next word prediction objective undoes some of the benefits of instruction tuning and reduces the average performance by 0.8 F1 points. We conduct additional experiments with Bonito to understand the effects of the domain, the size of the training set, and the choice of alternative synthetic task generators. Overall, we show that learning with synthetic instruction tuning datasets is an effective way to adapt language models to new domains. The model, dataset, and code are available at https://github.com/BatsResearch/bonito. 2024: Nihal V. Nayak, Yiyang Nan, Avi Trost, Stephen H. Bach https://arxiv.org/pdf/2402.18334v1.pdf

Prompt engineering is a challenging and important task due to the high sensitivity of Large Language Models (LLMs) to the given prompt and the inherent ambiguity of a textual task instruction. Automatic prompt engineering is essential to achieve optimized performance from LLMs. Recent studies have demonstrated the capabilities of LLMs to automatically conduct prompt engineering by employing a meta-prompt that incorporates the outcomes of the last trials and proposes an improved prompt. However, this requires a high-quality benchmark to compare different prompts, which is difficult and expensive to acquire in many real-world use cases. In this work, we introduce a new method for automatic prompt engineering, using a calibration process that iteratively refines the prompt to the user intent. During the optimization process, the system jointly generates synthetic data of boundary use cases and optimizes the prompt according to the generated dataset. We demonstrate the effectiveness of our method with respect to strong proprietary models on real-world tasks such as moderation and generation. Our method outperforms state-of-the-art methods with a limited number of annotated samples. Furthermore, we validate the advantages of each one of the system's key components. Our system is built in a modular way, facilitating easy adaptation to other tasks. The code is available $\href{https://github.com/Eladlev/AutoPrompt}{here}$. 2024: Elad Levi, Eli Brosh, Matan Friedmann https://arxiv.org/pdf/2402.03099v1.pdf

Sora is a text-to-video generative AI model, released by OpenAI in February 2024. The model is trained to generate videos of realistic or imaginative scenes from text instructions and show potential in simulating the physical world. Based on public technical reports and reverse engineering, this paper presents a comprehensive review of the model's background, related technologies, applications, remaining challenges, and future directions of text-to-video AI models. We first trace Sora's development and investigate the underlying technologies used to build this"world simulator". Then, we describe in detail the applications and potential impact of Sora in multiple industries ranging from film-making and education to marketing. We discuss the main challenges and limitations that need to be addressed to widely deploy Sora, such as ensuring safe and unbiased video generation. Lastly, we discuss the future development of Sora and video generation models in general, and how advancements in the field could enable new ways of human-AI interaction, boosting productivity and creativity of video generation. 2024: Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, Lifang He, Lichao Sun https://arxiv.org/pdf/2402.17177v2.pdf

Large Language Models (LLMs) are typically trained in two phases: pre-training on large internet-scale datasets, and fine-tuning for downstream tasks. Given the higher computational demand of pre-training, it's intuitive to assume that fine-tuning adds less new information to the model, and is thus more compressible. We explore this assumption by decomposing the weights of fine-tuned models into their pre-trained components and an additional delta. We introduce a simple method, BitDelta, which successfully quantizes this delta down to 1 bit without compromising performance. This interesting finding not only highlights the potential redundancy of information added during fine-tuning, but also has significant implications for the multi-tenant serving and multi-tenant storage of fine-tuned models. By enabling the use of a single high-precision base model accompanied by multiple 1-bit deltas, BitDelta dramatically reduces GPU memory requirements by more than 10x, which can also be translated to enhanced generation latency in multi-tenant settings. We validate BitDelta through experiments across Llama-2 and Mistral model families, and on models up to 70B parameters, showcasing minimal performance degradation over all tested settings. 2024: James Liu, Guangxuan Xiao, Kai Li, Jason D. Lee, Song Han, Tri Dao, Tianle Cai https://arxiv.org/pdf/2402.10193v1.pdf

Transformers have emerged as the architecture of choice for many state-of-the-art AI models, showcasing exceptional performance across a wide range of AI applications. However, the memory demands imposed by Transformers limit their ability to handle long sequences, thereby posing challenges in utilizing videos, actions, and other long-form sequences and modalities in complex environments. We present a novel approach, Ring Attention with Blockwise Transformers (Ring Attention), which leverages blockwise computation of self-attention and feedforward to distribute long sequences across multiple devices while fully overlapping the communication of key-value blocks with the computation of blockwise attention. Our approach enables training and inference of sequences that are up to device count times longer than those achievable by prior memory-efficient Transformers, without resorting to approximations or incurring additional communication and computation overheads. Extensive experiments on language modeling and reinforcement learning tasks demonstrate the effectiveness of our approach in allowing millions of tokens context size and improving performance. 2023: Hao Liu, Matei Zaharia, Pieter Abbeel https://arxiv.org/pdf/2310.01889v4.pdf

Large language models (LLMs) have accomplished remarkable reasoning performance in various domains. However, in the domain of reasoning tasks, we discover a frailty: LLMs are surprisingly brittle to the ordering of the premises, despite the fact that such ordering does not alter the underlying task. In particular, we observe that LLMs achieve the best performance when the premise order aligns with the context required in intermediate reasoning steps. For example, in deductive reasoning tasks, presenting the premises in the same order as the ground truth proof in the prompt (as opposed to random ordering) drastically increases the model's accuracy. We first examine the effect of premise ordering on deductive reasoning on a variety of LLMs, and our evaluation shows that permuting the premise order can cause a performance drop of over 30%. In addition, we release the benchmark R-GSM, based on GSM8K, to examine the ordering effect for mathematical problem-solving, and we again observe a significant drop in accuracy, relative to the original GSM8K benchmark. 2024: Xinyun Chen, Ryan A. Chi, Xuezhi Wang, Denny Zhou https://arxiv.org/pdf/2402.08939.pdf

All text-based language problems can be reduced to either generation or embedding. Current models only perform well at one or the other. We introduce generative representational instruction tuning (GRIT) whereby a large language model is trained to handle both generative and embedding tasks by distinguishing between them through instructions. Compared to other open models, our resulting GritLM 7B sets a new state of the art on the Massive Text Embedding Benchmark (MTEB) and outperforms all models up to its size on a range of generative tasks. By scaling up further, GritLM 8x7B outperforms all open generative language models that we tried while still being among the best embedding models. Notably, we find that GRIT matches training on only generative or embedding data, thus we can unify both at no performance loss. Among other benefits, the unification via GRIT speeds up Retrieval-Augmented Generation (RAG) by>60% for long documents, by no longer requiring separate retrieval and generation models. Models, code, etc. are freely available at https://github.com/ContextualAI/gritlm. 2024: Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, Douwe Kiela https://arxiv.org/pdf/2402.09906v1.pdf