We introduce a framework for the use of large language models (LLMs) in Building Understandable Messaging for Policy and Evidence Review (BUMPER). LLMs are proving capable of providing interfaces for understanding and synthesizing large databases of diverse media. This presents an exciting opportunity to supercharge the translation of scientific evidence into policy and action, thereby improving livelihoods around the world. However, these models also pose challenges related to access, trust-worthiness, and accountability. The BUMPER framework is built atop a scientific knowledge base (e.g., documentation, code, survey data) by the same scientists (e.g., individual contributor, lab, consortium). We focus on a solution that builds trustworthiness through transparency, scope-limiting, explicit-checks, and uncertainty measures. LLMs are rapidly being adopted and consequences are poorly understood. The framework addresses open questions regarding the reliability of LLMs and their use in high-stakes applications. We provide a worked example in health policy for a model designed to inform measles control programs. We argue that this framework can facilitate accessibility of and confidence in scientific evidence for policymakers, drive a focus on policy-relevance and translatability for researchers, and ultimately increase and accelerate the impact of scientific knowledge used for policy decisions.

This study investigates the impact of Large Language Models (LLMs) generating follow-up questions in response to user requests for short (1-page) text documents. Users provided prompts requesting documents they would like the AI to produce. The AI then generated questions to clarify the user's needs before generating the requested documents. Users answered the questions and then indicated their preference between a document generated using both the initial prompt and the questions and answers, and a document generated using only the initial prompt, and gave feedback about their experience with the question-answering process.

Deep neural networks have achieved remarkable performance across various tasks when supplied with large-scale labeled data. However, the collection of labeled data can be time-consuming and labor-intensive. Semi-supervised learning (SSL), particularly through pseudo-labeling algorithms that iteratively assign pseudo-labels for self-training, offers a promising solution to mitigate the dependency of labeled data. Previous research generally applies a uniform pseudo-labeling strategy across all model layers, assuming that pseudo-labels exert uniform influence throughout. Contrasting this, our theoretical analysis and empirical experiment demonstrate feature extraction layer and linear classification layer have distinct learning behaviors in response to pseudo-labels. Based on these insights, we develop two layer-specific pseudo-label strategies, termed Grad-ReLU and Avg-Clustering. Grad-ReLU mitigates the impact of noisy pseudo-labels by removing the gradient detrimental effects of pseudo-labels in the linear classification layer. Avg-Clustering accelerates the convergence of feature extraction layer towards stable clustering centers by integrating consistent outputs. Our approach, LayerMatch, which integrates these two strategies, can avoid the severe interference of noisy pseudo-labels in the linear classification layer while accelerating the clustering capability of the feature extraction layer. Through extensive experimentation, our approach consistently demonstrates exceptional performance on standard semi-supervised learning benchmarks, achieving a significant improvement of 10.38% over baseline method and a 2.44% increase compared to state-of-the-art methods.

Large language models (LLMs) have achieved remarkable success but still tend to generate factually erroneous responses, a phenomenon known as hallucination. A recent trend is to use preference learning to fine-tune models to align with factuality. However, existing work primarily evaluates fine-tuned models on in-domain (ID) datasets and the factuality on out-of-domain (OOD) datasets remains underexplored. In this paper, we conduct a comprehensive evaluation of the factuality of different models tuned by various preference learning algorithms and demonstrate that their performance on OOD datasets either increases minimally or decreases. Subsequently, we reveal that the main cause of model's failure to uphold factuality under a distribution shift is \textbf{under-alignment}, rather than \textbf{over-alignment}, by analyzing the token distribution shift of the models before and after tuning. Finally, we propose \textbf{APEFT} (\textbf{A}tomic \textbf{P}reference \textbf{E}nhanced \textbf{F}actuality \textbf{T}uning), a framework that enhances model's awareness of factuality at the granularity of individual facts. Extensive experiments demonstrate that APEFT improves model performance by an average of $\boldsymbol{3.45\%}$ on both ID and OOD datasets, which is highly effective.

Although it is very unlikely that machines will exhibit broadly-applicable intelligence comparable to or exceeding that of humans in the next 30 years (strong AI), it is to be expected that machines will reach and exceed human performance on more and more applied tasks. To develop the positive aspects of AI, manage its risks and challenges, and ensure that everyone has the opportunity to help in building an AI-enhanced society and to participate in its benefits, in this project, human intelligence and machine learning (ML) take the centre stage: Interactive Machine Learning (IML) is the design and implementation of algorithms and intelligent user interface frameworks that facilitate ML with the help of human interaction. Our focus is to improve the interaction between humans and machines, by leveraging state-of-the-art humancomputer interaction (HCI) approaches, as well as solutions that involve state-of-the-art ML techniques. In this project, we focus on Interactive Deep Learning (IDL): deep learning (DL) approaches for IML. We want computers to learn from humans by interacting with them in natural language for example and by observing them.

Synthetic data generation has gained significant attention recently for its utility in training large vision and language models. However, the application of synthetic data to the training of multimodal context-augmented generation systems has been relatively unexplored. This gap in existing work is important because existing vision and language models (VLMs) are not trained specifically for context-augmented generation. Resources for adapting such models are therefore crucial for enabling their use in retrieval-augmented generation (RAG) settings, where a retriever is used to gather relevant information that is then subsequently provided to a generative model via context augmentation. To address this challenging problem, we generate SK-VQA: a large synthetic multimodal dataset containing over 2 million question-answer pairs which require external knowledge to determine the final answer. Our dataset is both larger and significantly more diverse than existing resources of its kind, possessing over 11x more unique questions and containing images from a greater variety of sources than previously-proposed datasets. Through extensive experiments, we demonstrate that our synthetic dataset can not only serve as a challenging benchmark, but is also highly effective for adapting existing generative multimodal models for context-augmented generation.

Named Entity Recognition (NER) is a well-studied problem in NLP. However, there is much less focus on studying NER datasets, compared to developing new NER models. In this paper, we employed three simple techniques to detect annotation errors in the OntoNotes 5.0 corpus for English NER, which is the largest available NER corpus for English. Our techniques corrected ~10% of the sentences in train/dev/test data. In terms of entity mentions, we corrected the span and/or type of ~8% of mentions in the dataset, while adding/deleting/splitting/merging a few more. These are large numbers of changes, considering the size of OntoNotes. We used three NER libraries to train, evaluate and compare the models trained with the original and the re-annotated datasets, which showed an average improvement of 1.23% in overall F-scores, with large (>10%) improvements for some of the entity types. While our annotation error detection methods are not exhaustive and there is some manual annotation effort involved, they are largely language agnostic and can be employed with other NER datasets, and other sequence labelling tasks.

Interlinear glossed text (IGT) is a popular format in language documentation projects, where each morpheme is labeled with a descriptive annotation. Automating the creation of interlinear glossed text can be desirable to reduce annotator effort and maintain consistency across annotated corpora. Prior research has explored a number of statistical and neural methods for automatically producing IGT. As large language models (LLMs) have showed promising results across multilingual tasks, even for rare, endangered languages, it is natural to wonder whether they can be utilized for the task of generating IGT. We explore whether LLMs can be effective at the task of interlinear glossing with in-context learning, without any traditional training. We propose new approaches for selecting examples to provide in-context, observing that targeted selection can significantly improve performance. We find that LLM-based methods beat standard transformer baselines, despite requiring no training at all. These approaches still underperform state-of-the-art supervised systems for the task, but are highly practical for researchers outside of the NLP community, requiring minimal effort to use.

Existing metrics for evaluating the factuality of long-form text, such as FACTSCORE (Min et al., 2023) and SAFE (Wei et al., 2024), decompose an input text into "atomic claims" and verify each against a knowledge base like Wikipedia. These metrics are not suitable for most generation tasks because they assume that every claim is verifiable (i.e., can plausibly be proven true or false). We address this issue with VERISCORE, a metric for diverse long-form generation tasks that contain both verifiable and unverifiable content. VERISCORE can be effectively implemented with either closed or fine-tuned open-weight language models, and human evaluation confirms that VERISCORE's extracted claims are more sensible than those from competing methods across eight different long-form tasks. We use VERISCORE to evaluate generations from 16 different models across multiple long-form tasks and find that while GPT-4o is the best-performing model overall, open-weight models such as Mixtral-8x22 are closing the gap. We show that an LM's VERISCORE on one task (e.g., biography generation) does not necessarily correlate to its VERISCORE on a different task (e.g., long-form QA), highlighting the need for expanding factuality evaluation across tasks with varying fact density.

We examine how users perceive the limitations of an AI system when it encounters a task that it cannot perform perfectly and whether providing explanations alongside its answers aids users in constructing an appropriate mental model of the system's capabilities and limitations. We employ a visual question answer and explanation task where we control the AI system's limitations by manipulating the visual inputs: during inference, the system either processes full-color or grayscale images. Our goal is to determine whether participants can perceive the limitations of the system. We hypothesize that explanations will make limited AI capabilities more transparent to users. However, our results show that explanations do not have this effect. Instead of allowing users to more accurately assess the limitations of the AI system, explanations generally increase users' perceptions of the system's competence - regardless of its actual performance.