Despite the recent progress in long-context language models, it remains elusive how transformer-based models exhibit the capability to retrieve relevant information from arbitrary locations within the long context. This paper aims to address this question. Our systematic investigation across a wide spectrum of models reveals that a special type of attention heads are largely responsible for retrieving information, which we dub retrieval heads. We identify intriguing properties of retrieval heads:(1) universal: all the explored models with long-context capability have a set of retrieval heads; (2) sparse: only a small portion (less than 5\%) of the attention heads are retrieval. (3) intrinsic: retrieval heads already exist in models pretrained with short context. When extending the context length by continual pretraining, it is still the same set of heads that perform information retrieval. (4) dynamically activated: take Llama-2 7B for example, 12 retrieval heads always attend to the required information no matter how the context is changed. The rest of the retrieval heads are activated in different contexts. (5) causal: completely pruning retrieval heads leads to failure in retrieving relevant information and results in hallucination, while pruning random non-retrieval heads does not affect the model's retrieval ability. We further show that retrieval heads strongly influence chain-of-thought (CoT) reasoning, where the model needs to frequently refer back the question and previously-generated context. Conversely, tasks where the model directly generates the answer using its intrinsic knowledge are less impacted by masking out retrieval heads. These observations collectively explain which internal part of the model seeks information from the input tokens. We believe our insights will foster future research on reducing hallucination, improving reasoning, and compressing the KV cache.

In biological tasks, data is rarely plentiful as it is generated from hard-to-gather measurements. Therefore, pre-training foundation models on large quantities of available data and then transfer to low-data downstream tasks is a promising direction. However, how to design effective foundation models for molecular learning remains an open question, with existing approaches typically focusing on models with large parameter capacities. In this work, we propose $\texttt{MiniMol}$, a foundational model for molecular learning with 10 million parameters. $\texttt{MiniMol}$ is pre-trained on a mix of roughly 3300 sparsely defined graph- and node-level tasks of both quantum and biological nature. The pre-training dataset includes approximately 6 million molecules and 500 million labels. To demonstrate the generalizability of $\texttt{MiniMol}$ across tasks, we evaluate it on downstream tasks from the Therapeutic Data Commons (TDC) ADMET group showing significant improvements over the prior state-of-the-art foundation model across 17 tasks. $\texttt{MiniMol}$ will be a public and open-sourced model for future research.

General purpose Large Language Models (LLM) such as the Generative Pretrained Transformer (GPT) and Large Language Model Meta AI (LLaMA) have attracted much attention in recent years. There is strong evidence that these models can perform remarkably well in various natural language processing tasks. However, how to leverage them to approach domain-specific use cases and drive value remains an open question. In this work, we focus on a specific use case, pharmaceutical manufacturing investigations, and propose that leveraging historical records of manufacturing incidents and deviations in an organization can be beneficial for addressing and closing new cases, or de-risking new manufacturing campaigns. Using a small but diverse dataset of real manufacturing deviations selected from different product lines, we evaluate and quantify the power of three general purpose LLMs (GPT-3.5, GPT-4, and Claude-2) in performing tasks related to the above goal. In particular, (1) the ability of LLMs in automating the process of extracting specific information such as root cause of a case from unstructured data, as well as (2) the possibility of identifying similar or related deviations by performing semantic search on the database of historical records are examined. While our results point to the high accuracy of GPT-4 and Claude-2 in the information extraction task, we discuss cases of complex interplay between the apparent reasoning and hallucination behavior of LLMs as a risk factor. Furthermore, we show that semantic search on vector embedding of deviation descriptions can be used to identify similar records, such as those with a similar type of defect, with a high level of accuracy. We discuss further improvements to enhance the accuracy of similar record identification.

Geospatial Copilots unlock unprecedented potential for performing Earth Observation (EO) applications through natural language instructions. However, existing agents rely on overly simplified single tasks and template-based prompts, creating a disconnect with real-world scenarios. In this work, we present GeoLLM-Engine, an environment for tool-augmented agents with intricate tasks routinely executed by analysts on remote sensing platforms. We enrich our environment with geospatial API tools, dynamic maps/UIs, and external multimodal knowledge bases to properly gauge an agent's proficiency in interpreting realistic high-level natural language commands and its functional correctness in task completions. By alleviating overheads typically associated with human-in-the-loop benchmark curation, we harness our massively parallel engine across 100 GPT-4-Turbo nodes, scaling to over half a million diverse multi-tool tasks and across 1.1 million satellite images. By moving beyond traditional single-task image-caption paradigms, we investigate state-of-the-art agents and prompting techniques against long-horizon prompts.

Long document question answering (DocQA) aims to answer questions from long documents over 10k words. They usually contain content structures such as sections, sub-sections, and paragraph demarcations. However, the indexing methods of long documents remain under-explored, while existing systems generally employ fixed-length chunking. As they do not consider content structures, the resultant chunks can exclude vital information or include irrelevant content. Motivated by this, we propose the Multi-view Content-aware indexing (MC-indexing) for more effective long DocQA via (i) segment structured document into content chunks, and (ii) represent each content chunk in raw-text, keywords, and summary views. We highlight that MC-indexing requires neither training nor fine-tuning. Having plug-and-play capability, it can be seamlessly integrated with any retrievers to boost their performance. Besides, we propose a long DocQA dataset that includes not only question-answer pair, but also document structure and answer scope. When compared to state-of-art chunking schemes, MC-indexing has significantly increased the recall by 42.8%, 30.0%, 23.9%, and 16.3% via top k= 1.5, 3, 5, and 10 respectively. These improved scores are the average of 8 widely used retrievers (2 sparse and 6 dense) via extensive experiments.

This study presents a comprehensive analysis and comparison of two predominant fine-tuning methodologies - full-parameter fine-tuning and parameter-efficient tuning - within the context of medical Large Language Models (LLMs). We developed and refined a series of LLMs, based on the Llama-2 architecture, specifically designed to enhance medical knowledge retrieval, reasoning, and question-answering capabilities. Our experiments systematically evaluate the effectiveness of these tuning strategies across various well-known medical benchmarks. Notably, our medical LLM Med42 showed an accuracy level of 72% on the US Medical Licensing Examination (USMLE) datasets, setting a new standard in performance for openly available medical LLMs. Through this comparative analysis, we aim to identify the most effective and efficient method for fine-tuning LLMs in the medical domain, thereby contributing significantly to the advancement of AI-driven healthcare applications.

This paper focuses on a very important societal challenge of water quality analysis. Being one of the key factors in the economic and social development of society, the provision of water and ensuring its quality has always remained one of the top priorities of public authorities. To ensure the quality of water, different methods for monitoring and assessing the water networks, such as offline and online surveys, are used. However, these surveys have several limitations, such as the limited number of participants and low frequency due to the labor involved in conducting such surveys. In this paper, we propose a Natural Language Processing (NLP) framework to automatically collect and analyze water-related posts from social media for data-driven decisions. The proposed framework is composed of two components, namely (i) text classification, and (ii) topic modeling. For text classification, we propose a merit-fusion-based framework incorporating several Large Language Models (LLMs) where different weight selection and optimization methods are employed to assign weights to the LLMs. In topic modeling, we employed the BERTopic library to discover the hidden topic patterns in the water-related tweets. We also analyzed relevant tweets originating from different regions and countries to explore global, regional, and country-specific issues and water-related concerns. We also collected and manually annotated a large-scale dataset, which is expected to facilitate future research on the topic.

The deployment of autonomous agents in environments involving human interaction has increasingly raised security concerns. Consequently, understanding the circumstances behind an event becomes critical, requiring the development of capabilities to justify their behaviors to non-expert users. Such explanations are essential in enhancing trustworthiness and safety, acting as a preventive measure against failures, errors, and misunderstandings. Additionally, they contribute to improving communication, bridging the gap between the agent and the user, thereby improving the effectiveness of their interactions. This work presents an accountability and explainability architecture implemented for ROS-based mobile robots. The proposed solution consists of two main components. Firstly, a black box-like element to provide accountability, featuring anti-tampering properties achieved through blockchain technology. Secondly, a component in charge of generating natural language explanations by harnessing the capabilities of Large Language Models (LLMs) over the data contained within the previously mentioned black box. The study evaluates the performance of our solution in three different scenarios, each involving autonomous agent navigation functionalities. This evaluation includes a thorough examination of accountability and explainability metrics, demonstrating the effectiveness of our approach in using accountable data from robot actions to obtain coherent, accurate and understandable explanations, even when facing challenges inherent in the use of autonomous agents in real-world scenarios.

While acoustic expressiveness has long been studied in expressive text-to-speech (ETTS), the inherent expressiveness in text lacks sufficient attention, especially for ETTS of artistic works. In this paper, we introduce StoryTTS, a highly ETTS dataset that contains rich expressiveness both in acoustic and textual perspective, from the recording of a Mandarin storytelling show. A systematic and comprehensive labeling framework is proposed for textual expressiveness. We analyze and define speech-related textual expressiveness in StoryTTS to include five distinct dimensions through linguistics, rhetoric, etc. Then we employ large language models and prompt them with a few manual annotation examples for batch annotation. The resulting corpus contains 61 hours of consecutive and highly prosodic speech equipped with accurate text transcriptions and rich textual expressiveness annotations. Therefore, StoryTTS can aid future ETTS research to fully mine the abundant intrinsic textual and acoustic features. Experiments are conducted to validate that TTS models can generate speech with improved expressiveness when integrating with the annotated textual labels in StoryTTS.

Despite the remarkable achievements of language models (LMs) across a broad spectrum of tasks, their propensity for generating toxic outputs remains a prevalent concern. Current solutions involving fine-tuning or auxiliary models usually require extensive memory and computational resources, rendering them less practical for deployment in large language models (LLMs). In this paper, we propose DeStein, a novel method that detoxififies LMs by altering their internal representations in the activation space with lower resource and time cost. Specifically, we leverage self-induced steering pairs to identify detoxification vectors through arithmetic operations in the activation space. During inference, detoxification is achieved by blending the detoxification vectors with the original representations. Empirical results demonstrate that our method significantly outperforms previous state-of-the-art approaches on popular detoxification metrics, while also maintaining satisfactory generation quality and diversity. Furthermore, we extend our method to multiple LLMs, demonstrating its practicality and scalability. We open-source our method at https://github.com/LizLizLi/DeStein . Warning: Some example model outputs contain highly offensive or disturbing text.