This year, we explore generation-augmented retrieval for the TRECVid AVS task. Specifically, the understanding of textual query is enhanced by three generations, including Text2Text, Text2Image, and Image2Text, to address the out-of-vocabulary problem. Using different combinations of them and the rank list retrieved by the original query, we submitted four automatic runs. For manual runs, we use a large language model (LLM) (i.e., GPT4) to rephrase test queries based on the concept bank of the search engine, and we manually check again to ensure all the concepts used in the rephrased queries are in the bank. The result shows that the fusion of the original and generated queries outperforms the original query on TV24 query sets. The generated queries retrieve different rank lists from the original query.
The deployment of artificial intelligence (AI) in critical decision-making and evaluation processes raises concerns about inherent biases that malicious actors could exploit to distort decision outcomes. We propose a systematic method to reveal such biases in AI evaluation systems and apply it to automated essay grading as an example. Our approach first identifies hidden neural activity patterns that predict distorted decision outcomes and then optimizes an adversarial input suffix to amplify such patterns. We demonstrate that this combination can effectively fool large language model (LLM) graders into assigning much higher grades than humans would. We further show that this white-box attack transfers to black-box attacks on other models, including commercial closed-source models like Gemini. They further reveal the existence of a "magic word" that plays a pivotal role in the efficacy of the attack. We trace the origin of this magic word bias to the structure of commonly-used chat templates for supervised fine-tuning of LLMs and show that a minor change in the template can drastically reduce the bias. This work not only uncovers vulnerabilities in current LLMs but also proposes a systematic method to identify and remove hidden biases, contributing to the goal of ensuring AI safety and security.
Dialogue intent classification aims to identify the underlying purpose or intent of a user's input in a conversation. Current intent classification systems encounter considerable challenges, primarily due to the vast number of possible intents and the significant semantic overlap among similar intent classes. In this paper, we propose a novel approach to few-shot dialogue intent classification through in-context learning, incorporating dynamic label refinement to address these challenges. Our method retrieves relevant examples for a test input from the training set and leverages a large language model to dynamically refine intent labels based on semantic understanding, ensuring that intents are clearly distinguishable from one another. Experimental results demonstrate that our approach effectively resolves confusion between semantically similar intents, resulting in significantly enhanced performance across multiple datasets compared to baselines. We also show that our method generates more interpretable intent labels, and has a better semantic coherence in capturing underlying user intents compared to baselines.
Japan Agency for Marine-Earth Science and Technology (JAMSTEC) has made available the JAMSTEC Earth Deep-sea Image (J-EDI), a deep-sea video and image archive (https://www.godac.jamstec.go.jp/jedi/e/index.html). This archive serves as a valuable resource for researchers and scholars interested in deep-sea imagery. The dataset comprises images and videos of deep-sea phenomena, predominantly of marine organisms, but also of the seafloor and physical processes. In this study, we propose J-EDI QA, a benchmark for understanding images of deep-sea organisms using a multimodal large language model (LLM). The benchmark is comprised of 100 images, accompanied by questions and answers with four options by JAMSTEC researchers for each image. The QA pairs are provided in Japanese, and the benchmark assesses the ability to understand deep-sea species in Japanese. In the evaluation presented in this paper, OpenAI o1 achieved a 50% correct response rate. This result indicates that even with the capabilities of state-of-the-art models as of December 2024, deep-sea species comprehension is not yet at an expert level. Further advances in deep-sea species-specific LLMs are therefore required.
Problem: Pancreas radiological imaging is challenging due to the small size, blurred boundaries, and variability of shape and position of the organ among patients. Goal: In this work we present MiniGPT-Pancreas, a Multimodal Large Language Model (MLLM), as an interactive chatbot to support clinicians in pancreas cancer diagnosis by integrating visual and textual information. Methods: MiniGPT-v2, a general-purpose MLLM, was fine-tuned in a cascaded way for pancreas detection, tumor classification, and tumor detection with multimodal prompts combining questions and computed tomography scans from the National Institute of Health (NIH), and Medical Segmentation Decathlon (MSD) datasets. The AbdomenCT-1k dataset was used to detect the liver, spleen, kidney, and pancreas. Results: MiniGPT-Pancreas achieved an Intersection over Union (IoU) of 0.595 and 0.550 for the detection of pancreas on NIH and MSD datasets, respectively. For the pancreas cancer classification task on the MSD dataset, accuracy, precision, and recall were 0.876, 0.874, and 0.878, respectively. When evaluating MiniGPT-Pancreas on the AbdomenCT-1k dataset for multi-organ detection, the IoU was 0.8399 for the liver, 0.722 for the kidney, 0.705 for the spleen, and 0.497 for the pancreas. For the pancreas tumor detection task, the IoU score was 0.168 on the MSD dataset. Conclusions: MiniGPT-Pancreas represents a promising solution to support clinicians in the classification of pancreas images with pancreas tumors. Future research is needed to improve the score on the detection task, especially for pancreas tumors.
We present ARCAS (Automated Root Cause Analysis System), a diagnostic platform based on a Domain Specific Language (DSL) built for fast diagnostic implementation and low learning curve. Arcas is composed of a constellation of automated troubleshooting guides (Auto-TSGs) that can execute in parallel to detect issues using product telemetry and apply mitigation in near-real-time. The DSL is tailored specifically to ensure that subject matter experts can deliver highly curated and relevant Auto-TSGs in a short time without having to understand how they will interact with the rest of the diagnostic platform, thus reducing time-to-mitigate and saving crucial engineering cycles when they matter most. This contrasts with platforms like Datadog and New Relic, which primarily focus on monitoring and require manual intervention for mitigation. ARCAS uses a Large Language Model (LLM) to prioritize Auto-TSGs outputs and take appropriate actions, thus suppressing the costly requirement of understanding the general behavior of the system. We explain the key concepts behind ARCAS and demonstrate how it has been successfully used for multiple products across Azure Synapse Analytics and Microsoft Fabric Synapse Data Warehouse.
We argue that the Declarative Self-improving Python (DSPy) optimizers are a way to align the large language model (LLM) prompts and their evaluations to the human annotations. We present a comparative analysis of five teleprompter algorithms, namely, Cooperative Prompt Optimization (COPRO), Multi-Stage Instruction Prompt Optimization (MIPRO), BootstrapFewShot, BootstrapFewShot with Optuna, and K-Nearest Neighbor Few Shot, within the DSPy framework with respect to their ability to align with human evaluations. As a concrete example, we focus on optimizing the prompt to align hallucination detection (using LLM as a judge) to human annotated ground truth labels for a publicly available benchmark dataset. Our experiments demonstrate that optimized prompts can outperform various benchmark methods to detect hallucination, and certain telemprompters outperform the others in at least these experiments.
We introduce CLOB, a novel continual learning (CL) paradigm wherein a large language model (LLM) is regarded as a black box. Learning is done incrementally via only verbal prompting. CLOB does not fine-tune any part of the LLM or add any trainable parameters to it. It is particularly suitable for LLMs that are accessible via APIs. We also propose a new CL technique, called CIS, based on incremental summarization that also overcomes the LLM's input length limit. Experiments show CIS outperforms baselines by a very large margin.
Memory plays a pivotal role in enabling large language model~(LLM)-based agents to engage in complex and long-term interactions, such as question answering (QA) and dialogue systems. While various memory modules have been proposed for these tasks, the impact of different memory structures across tasks remains insufficiently explored. This paper investigates how memory structures and memory retrieval methods affect the performance of LLM-based agents. Specifically, we evaluate four types of memory structures, including chunks, knowledge triples, atomic facts, and summaries, along with mixed memory that combines these components. In addition, we evaluate three widely used memory retrieval methods: single-step retrieval, reranking, and iterative retrieval. Extensive experiments conducted across four tasks and six datasets yield the following key insights: (1) Different memory structures offer distinct advantages, enabling them to be tailored to specific tasks; (2) Mixed memory structures demonstrate remarkable resilience in noisy environments; (3) Iterative retrieval consistently outperforms other methods across various scenarios. Our investigation aims to inspire further research into the design of memory systems for LLM-based agents.
Outlier detection (OD), also known as anomaly detection, is a critical machine learning (ML) task with applications in fraud detection, network intrusion detection, clickstream analysis, recommendation systems, and social network moderation. Among open-source libraries for outlier detection, the Python Outlier Detection (PyOD) library is the most widely adopted, with over 8,500 GitHub stars, 25 million downloads, and diverse industry usage. However, PyOD currently faces three limitations: (1) insufficient coverage of modern deep learning algorithms, (2) fragmented implementations across PyTorch and TensorFlow, and (3) no automated model selection, making it hard for non-experts. To address these issues, we present PyOD Version 2 (PyOD 2), which integrates 12 state-of-the-art deep learning models into a unified PyTorch framework and introduces a large language model (LLM)-based pipeline for automated OD model selection. These improvements simplify OD workflows, provide access to 45 algorithms, and deliver robust performance on various datasets. In this paper, we demonstrate how PyOD 2 streamlines the deployment and automation of OD models and sets a new standard in both research and industry. PyOD 2 is accessible at [https://github.com/yzhao062/pyod](https://github.com/yzhao062/pyod). This study aligns with the Web Mining and Content Analysis track, addressing topics such as the robustness of Web mining methods and the quality of algorithmically-generated Web data.
Recent advances in machine learning, particularly Large Language Models (LLMs) such as BERT and GPT, provide rich contextual embeddings that improve text representation. However, current document clustering approaches often ignore the deeper relationships between named entities (NEs) and the potential of LLM embeddings. This paper proposes a novel approach that integrates Named Entity Recognition (NER) and LLM embeddings within a graph-based framework for document clustering. The method builds a graph with nodes representing documents and edges weighted by named entity similarity, optimized using a graph-convolutional network (GCN). This ensures a more effective grouping of semantically related documents. Experimental results indicate that our approach outperforms conventional co-occurrence-based methods in clustering, notably for documents rich in named entities.
Social media constitutes a rich and influential source of information for qualitative researchers. Although computational techniques like topic modelling assist with managing the volume and diversity of social media content, qualitative researcher's lack of programming expertise creates a significant barrier to their adoption. In this paper we explore how BERTopic, an advanced Large Language Model (LLM)-based topic modelling technique, can support qualitative data analysis of social media. We conducted interviews and hands-on evaluations in which qualitative researchers compared topics from three modelling techniques: LDA, NMF, and BERTopic. BERTopic was favoured by 8 of 12 participants for its ability to provide detailed, coherent clusters for deeper understanding and actionable insights. Participants also prioritised topic relevance, logical organisation, and the capacity to reveal unexpected relationships within the data. Our findings underscore the potential of LLM-based techniques for supporting qualitative analysis.
Large Language Model (LLM) based coding tools have been tremendously successful as software development assistants, yet they are often designed for general purpose programming tasks and perform poorly for more specialized domains such as high performance computing. Creating specialized models and tools for these domains is crucial towards gaining the benefits of LLMs in areas such as HPC. While previous work has explored HPC-specific models, LLMs still struggle to generate parallel code and it is not at all clear what hurdles are still holding back these LLMs and what must be done to overcome them. In this work, we conduct an in-depth study along the many axes of fine-tuning a specialized HPC LLM in order to better understand the challenges. Based on our findings we fine-tune and evaluate a specialized HPC LLM that is shown to be the best performing open-source code LLM for parallel code generation to date.
Retrieval-based multi-image question answering (QA) task involves retrieving multiple question-related images and synthesizing these images to generate an answer. Conventional "retrieve-then-answer" pipelines often suffer from cascading errors because the training objective of QA fails to optimize the retrieval stage. To address this issue, we propose a novel method to effectively introduce and reference retrieved information into the QA. Given the image set to be retrieved, we employ a multimodal large language model (visual perspective) and a large language model (textual perspective) to obtain multimodal hypothetical summary in question-form and description-form. By combining visual and textual perspectives, MHyS captures image content more specifically and replaces real images in retrieval, which eliminates the modality gap by transforming into text-to-text retrieval and helps improve retrieval. To more advantageously introduce retrieval with QA, we employ contrastive learning to align queries (questions) with MHyS. Moreover, we propose a coarse-to-fine strategy for calculating both sentence-level and word-level similarity scores, to further enhance retrieval and filter out irrelevant details. Our approach achieves a 3.7% absolute improvement over state-of-the-art methods on RETVQA and a 14.5% improvement over CLIP. Comprehensive experiments and detailed ablation studies demonstrate the superiority of our method.
In this paper, we introduce DirectorLLM, a novel video generation model that employs a large language model (LLM) to orchestrate human poses within videos. As foundational text-to-video models rapidly evolve, the demand for high-quality human motion and interaction grows. To address this need and enhance the authenticity of human motions, we extend the LLM from a text generator to a video director and human motion simulator. Utilizing open-source resources from Llama 3, we train the DirectorLLM to generate detailed instructional signals, such as human poses, to guide video generation. This approach offloads the simulation of human motion from the video generator to the LLM, effectively creating informative outlines for human-centric scenes. These signals are used as conditions by the video renderer, facilitating more realistic and prompt-following video generation. As an independent LLM module, it can be applied to different video renderers, including UNet and DiT, with minimal effort. Experiments on automatic evaluation benchmarks and human evaluations show that our model outperforms existing ones in generating videos with higher human motion fidelity, improved prompt faithfulness, and enhanced rendered subject naturalness.
Recent advances in language modeling demonstrate the need for high-quality domain-specific training data, especially for tasks that require specialized knowledge. General-purpose models, while versatile, often lack the depth needed for expert-level tasks because of limited domain-specific information. Domain adaptation training can enhance these models, but it demands substantial, high-quality data. To address this, we propose ORBIT, a cost-efficient methodology for curating massive, high-quality domain-specific datasets from noisy web sources, tailored for training specialist large language models. Using astronomy as a primary case study, we refined the 1.3T-token FineWeb-Edu dataset into a high-quality, 10B-token subset focused on astronomy. Fine-tuning \textsc{LLaMA-3-8B} on a 1B-token astronomy subset improved performance on the MMLU astronomy benchmark from 69\% to 76\% and achieved top results on AstroBench, an astronomy-specific benchmark. Moreover, our model (Orbit-LLaMA) outperformed \textsc{LLaMA-3-8B-base}, with GPT-4o evaluations preferring it in 73\% of cases across 1000 astronomy-specific questions. Additionally, we validated ORBIT's generalizability by applying it to law and medicine, achieving a significant improvement of data quality compared to an unfiltered baseline. We open-source the ORBIT methodology, including the curated datasets, the codebase, and the resulting model at \href{https://github.com/ModeEric/ORBIT-Llama}{https://github.com/ModeEric/ORBIT-Llama}.
In recent years, large-scale models have achieved significant advancements, accompanied by the emergence of numerous high-quality benchmarks for evaluating various aspects of their comprehension abilities. However, most existing benchmarks primarily focus on spatial understanding in static image tasks. While some benchmarks extend evaluations to temporal tasks, they fall short in assessing text generation under complex contexts involving long videos and rich auxiliary information. To address this limitation, we propose a novel benchmark: the Multi-modal Story Generation Benchmark (MSBench), designed to evaluate text generation capabilities in scenarios enriched with auxiliary information. Our work introduces an innovative automatic dataset generation method to ensure the availability of accurate auxiliary information. On one hand, we leverage existing datasets and apply automated processes to generate new evaluation datasets, significantly reducing manual efforts. On the other hand, we refine auxiliary data through systematic filtering and utilize state-of-the-art models to ensure the fairness and accuracy of the ground-truth datasets. Our experiments reveal that current Multi-modal Large Language Models (MLLMs) perform suboptimally under the proposed evaluation metrics, highlighting significant gaps in their capabilities. To address these challenges, we propose a novel model architecture and methodology to better handle the overall process, demonstrating improvements on our benchmark.
Capturing the workload of a database and replaying this workload for a new version of the database can be an effective approach for regression testing. However, false positive errors caused by many factors such as data privacy limitations, time dependency or non-determinism in multi-threaded environment can negatively impact the effectiveness. Therefore, we employ a machine learning based framework to automate the root cause analysis of failures found during replays. However, handling unseen novel issues not found in the training data is one general challenge of machine learning approaches with respect to generalizability of the learned model. We describe how we continue to address this challenge for more robust long-term solutions. From our experience, retraining with new failures is inadequate due to features overlapping across distinct root causes. Hence, we leverage a large language model (LLM) to analyze failed SQL statements and extract concise failure summaries as an additional feature to enhance the classification process. Our experiments show the F1-Macro score improved by 4.77% for our data. We consider our approach beneficial for providing end users with additional information to gain more insights into the found issues and to improve the assessment of the replay results.
Gender bias distorts the economic behavior and outcomes of women and households. We investigate gender biases in property taxes. We analyze records of more than 100,000 property tax appeal hearings and more than 2.7 years of associated audio recordings, considering how panelist and appellant genders associate with hearing outcomes. We first observe that female appellants fare systematically worse than male appellants in their hearings. Second, we show that, whereas male appellants' hearing outcomes do not vary meaningfully with the gender composition of the panel they face, those of female appellants' do, such that female appellants obtain systematically lesser (greater) reductions to their home values when facing female (male) panelists. Employing a multi-modal large language model (M-LLM), we next construct measures of participant behavior and tone from hearing audio recordings. We observe markedly different behaviors between male and female appellants and, in the case of male appellants, we find that these differences also depend on the gender of the panelists they face (e.g., male appellants appear to behave systematically more aggressively towards female panelists). In contrast, the behavior of female appellants remains relatively constant, regardless of their panel's gender. Finally, we show that female appellants continue to fare worse in front of female panels, even when we condition upon the appelant's in-hearing behavior and tone. Our results are thus consistent with the idea that gender biases are driven, at least in part, by unvoiced beliefs and perceptions on the part of ARB panelists. Our study documents the presence of gender biases in property appraisal appeal hearings and highlights the potential value of generative AI for analyzing large-scale, unstructured administrative data.
Regulation is increasingly cited as the most important and pressing concern in machine learning. However, it is currently unknown how to implement this, and perhaps more importantly, how it would effect model performance alongside human collaboration if actually realized. In this paper, we attempt to answer these questions by building a regulatable large-language model (LLM), and then quantifying how the additional constraints involved affect (1) model performance, alongside (2) human collaboration. Our empirical results reveal that it is possible to force an LLM to use human-defined features in a transparent way, but a "regulation performance trade-off" previously not considered reveals itself in the form of a 7.34% classification performance drop. Surprisingly however, we show that despite this, such systems actually improve human task performance speed and appropriate confidence in a realistic deployment setting compared to no AI assistance, thus paving a way for fair, regulatable AI, which benefits users.