Large Language Models (LLMs) have demonstrated impressive and diverse abilities that can benefit various domains, such as zero and few-shot information extraction from clinical text without domain-specific training. However, for the ICD coding task, they often hallucinate key details and produce high recall but low precision results due to the high-dimensional and skewed distribution of the ICD codes. Existing LLM-based methods fail to account for the complex and dynamic interactions among the human agents involved in coding, such as patients, physicians, and coders, and they lack interpretability and reliability. In this paper, we present a novel multi-agent method for ICD coding, which mimics the real-world coding process with five agents: a patient agent, a physician agent, a coder agent, a reviewer agent, and an adjuster agent. Each agent has a specific function and uses a LLM-based model to perform it. We evaluate our method on the MIMIC-III dataset and show that our proposed multi-agent coding framework substantially improves performance on both common and rare codes compared to Zero-shot Chain of Thought (CoT) prompting and self-consistency with CoT. The ablation study confirms the proposed agent roles' efficacy. Our method also matches the state-of-the-art ICD coding methods that require pre-training or fine-tuning, in terms of coding accuracy, rare code accuracy, and explainability.

This paper addresses the ethical challenges of Artificial Intelligence in Neural Machine Translation (NMT) systems, emphasizing the imperative for developers to ensure fairness and cultural sensitivity. We investigate the ethical competence of AI models in NMT, examining the Ethical considerations at each stage of NMT development, including data handling, privacy, data ownership, and consent. We identify and address ethical issues through empirical studies. These include employing Transformer models for Luganda-English translations and enhancing efficiency with sentence mini-batching. And complementary studies that refine data labeling techniques and fine-tune BERT and Longformer models for analyzing Luganda and English social media content. Our second approach is a literature review from databases such as Google Scholar and platforms like GitHub. Additionally, the paper probes the distribution of responsibility between AI systems and humans, underscoring the essential role of human oversight in upholding NMT ethical standards. Incorporating a biblical perspective, we discuss the societal impact of NMT and the broader ethical responsibilities of developers, positing them as stewards accountable for the societal repercussions of their creations.

It is commonplace to produce application-specific models by fine-tuning large pre-trained models using a small bespoke dataset. The widespread availability of foundation model checkpoints on the web poses considerable risks, including the vulnerability to backdoor attacks. In this paper, we unveil a new vulnerability: the privacy backdoor attack. This black-box privacy attack aims to amplify the privacy leakage that arises when fine-tuning a model: when a victim fine-tunes a backdoored model, their training data will be leaked at a significantly higher rate than if they had fine-tuned a typical model. We conduct extensive experiments on various datasets and models, including both vision-language models (CLIP) and large language models, demonstrating the broad applicability and effectiveness of such an attack. Additionally, we carry out multiple ablation studies with different fine-tuning methods and inference strategies to thoroughly analyze this new threat. Our findings highlight a critical privacy concern within the machine learning community and call for a reevaluation of safety protocols in the use of open-source pre-trained models.

Despite extensive diagnostics and debugging by developers, AI systems sometimes exhibit harmful unintended behaviors. Finding and fixing these is challenging because the attack surface is so large -- it is not tractable to exhaustively search for inputs that may elicit harmful behaviors. Red-teaming and adversarial training (AT) are commonly used to improve robustness, however, they empirically struggle to fix failure modes that differ from the attacks used during training. In this work, we utilize latent adversarial training (LAT) to defend against vulnerabilities without generating inputs that elicit them. LAT leverages the compressed, abstract, and structured latent representations of concepts that the network actually uses for prediction. We use it to remove trojans and defend against held-out classes of adversarial attacks. We show in image classification, text classification, and text generation tasks that LAT usually improves both robustness to novel attacks and performance on clean data relative to AT. This suggests that LAT can be a promising tool for defending against failure modes that are not explicitly identified by developers.

The growing availability and importance of time series data across various domains, including environmental science, epidemiology, and economics, has led to an increasing need for time-series causal discovery methods that can identify the intricate relationships in the non-stationary, non-linear, and often noisy real world data. However, the majority of current time series causal discovery methods assume stationarity and linear relations in data, making them infeasible for the task. Further, the recent deep learning-based methods rely on the traditional causal structure learning approaches making them computationally expensive. In this paper, we propose a Time-Series Causal Neural Network (TS-CausalNN) - a deep learning technique to discover contemporaneous and lagged causal relations simultaneously. Our proposed architecture comprises (i) convolutional blocks comprising parallel custom causal layers, (ii) acyclicity constraint, and (iii) optimization techniques using the augmented Lagrangian approach. In addition to the simple parallel design, an advantage of the proposed model is that it naturally handles the non-stationarity and non-linearity of the data. Through experiments on multiple synthetic and real world datasets, we demonstrate the empirical proficiency of our proposed approach as compared to several state-of-the-art methods. The inferred graphs for the real world dataset are in good agreement with the domain understanding.

Cancer stage classification is important for making treatment and care management plans for oncology patients. Information on staging is often included in unstructured form in clinical, pathology, radiology and other free-text reports in the electronic health record system, requiring extensive work to parse and obtain. To facilitate the extraction of this information, previous NLP approaches rely on labeled training datasets, which are labor-intensive to prepare. In this study, we demonstrate that without any labeled training data, open-source clinical large language models (LLMs) can extract pathologic tumor-node-metastasis (pTNM) staging information from real-world pathology reports. Our experiments compare LLMs and a BERT-based model fine-tuned using the labeled data. Our findings suggest that while LLMs still exhibit subpar performance in Tumor (T) classification, with the appropriate adoption of prompting strategies, they can achieve comparable performance on Metastasis (M) classification and improved performance on Node (N) classification.

While long-context large language models (LLMs) can technically summarize book-length documents (>100K tokens), the length and complexity of the documents have so far prohibited evaluations of input-dependent aspects like faithfulness. In this paper, we conduct the first large-scale human evaluation of faithfulness and content selection on LLM-generated summaries of fictional books. Our study mitigates the issue of data contamination by focusing on summaries of books published in 2023 or 2024, and we hire annotators who have fully read each book prior to the annotation task to minimize cost and cognitive burden. We collect FABLES, a dataset of annotations on 3,158 claims made in LLM-generated summaries of 26 books, at a cost of $5.2K USD, which allows us to rank LLM summarizers based on faithfulness: Claude-3-Opus significantly outperforms all closed-source LLMs, while the open-source Mixtral is on par with GPT-3.5-Turbo. An analysis of the annotations reveals that most unfaithful claims relate to events and character states, and they generally require indirect reasoning over the narrative to invalidate. While LLM-based auto-raters have proven reliable for factuality and coherence in other settings, we implement several LLM raters of faithfulness and find that none correlates strongly with human annotations, especially with regard to detecting unfaithful claims. Our experiments suggest that detecting unfaithful claims is an important future direction not only for summarization evaluation but also as a testbed for long-context understanding. Finally, we move beyond faithfulness by exploring content selection errors in book-length summarization: we develop a typology of omission errors related to crucial narrative elements and also identify a systematic over-emphasis on events occurring towards the end of the book.

Current Large Language Models (LLMs), even those tuned for safety and alignment, are susceptible to jailbreaking. Some have found that just further fine-tuning an aligned model with benign data (i.e., data without harmful content) surprisingly leads to substantial degradation in safety. We delve into the data-centric aspects of why benign fine-tuning inadvertently contributes to jailbreaking. First, we represent fine-tuning data through two lenses: representation and gradient spaces. Furthermore, we propose a bi-directional anchoring method that prioritizes data points that are close to harmful examples and distant from benign ones. By doing so, our approach effectively identifies subsets of benign data that are more likely to degrade the model's safety after fine-tuning. Training on just 100 of these seemingly benign datapoints can lead to the fine-tuned model affirmatively responding to > 70% of tested harmful requests, compared to < 20% after fine-tuning on randomly selected data. We further find that selected data are often in the form of lists and bullet points, or math questions.

Event temporal graphs have been shown as convenient and effective representations of complex temporal relations between events in text. Recent studies, which employ pre-trained language models to auto-regressively generate linearised graphs for constructing event temporal graphs, have shown promising results. However, these methods have often led to suboptimal graph generation as the linearised graphs exhibit set characteristics which are instead treated sequentially by language models. This discrepancy stems from the conventional text generation objectives, leading to erroneous penalisation of correct predictions caused by the misalignment of elements in target sequences. To address these challenges, we reframe the task as a conditional set generation problem, proposing a Set-aligning Framework tailored for the effective utilisation of Large Language Models (LLMs). The framework incorporates data augmentations and set-property regularisations designed to alleviate text generation loss penalties associated with the linearised graph edge sequences, thus encouraging the generation of more relation edges. Experimental results show that our framework surpasses existing baselines for event temporal graph generation. Furthermore, under zero-shot settings, the structural knowledge introduced through our framework notably improves model generalisation, particularly when the training examples available are limited.

High-res features can be learned either as a per-image implicit network or a general-purpose upsampling operation; the latter is a drop-in module to improve downstream dense prediction tasks. Deep features are a cornerstone of computer vision research, capturing image semantics and enabling the community to solve downstream tasks even in the zero-or few-shot regime. However, these features often lack the spatial resolution to directly perform dense prediction tasks like segmentation and depth prediction because models aggressively pool information over large areas. In this work, we introduce FeatUp, a task-and model-agnostic framework to restore lost spatial information in deep features. We introduce two variants of FeatUp: one that guides features with high-resolution signal in a single forward pass, and one that fits an implicit model to a single image to reconstruct features at any resolution. Both approaches use a multi-view consistency loss with deep analogies to NeRFs. Our features retain their original semantics and can be swapped into existing applications to yield resolution and performance gains even without re-training. We show that FeatUp significantly outperforms other feature upsampling and image super-resolution approaches in class activation map generation, transfer learning for segmentation and depth prediction, and end-to-end training for semantic segmentation. Despite their immense success, deep features often sacrifice spatial resolution for semantic quality. For example, ResNet-50 (He et al., 2015) produces 7 7 deep Work done while at MIT. FeatUp learns to upsample features through a consistency loss on low resolution "views" of a model's features that arise from slight transformations of the input image. Even Vision Transformers (ViTs) (Dosovitskiy et al., 2020) incur a significant resolution reduction, making it challenging to perform dense prediction tasks such as segmentation or depth estimation using these features alone. To mitigate these issues, we propose FeatUp: a novel framework to improve the resolution of any vision model's features without changing their original "meaning" or orientation. Our primary insight, inspired by 3D reconstruction frameworks like NeRF (Mildenhall et al., 2020), is that multiview consistency of low-resolution signals can supervise the construction of high-resolution signals. More specifically, we learn high-resolution information by aggregating low resolution views from a model's outputs across multiple "jittered" (e.g. Our work explores two architectures for upsampling: a single guided upsampling feedforward network that generalizes across images, and an implicit representation overfit to a single image.