Recent advances in large language models (LLMs) have opened up new paradigms for accessing the knowledge stored in their parameters. One critical challenge that has emerged is the presence of hallucinations in LLM outputs due to false or outdated knowledge. Since retraining LLMs with updated information is resource-intensive, there has been a growing interest in model editing. However, many model editing methods, while effective in various scenarios, tend to overemphasize aspects such as efficacy, generalization, and locality in editing performance, often overlooking potential side effects on the general abilities of LLMs. In this paper, we raise concerns that the improvement of model factuality may come at the cost of a significant degradation of these general abilities, which is not conducive to the sustainable development of LLMs. Systematically, we analyze side effects by evaluating four popular editing methods on two LLMs across eight representative task categories. Extensive empirical research reveals that model editing does improve model factuality but at the expense of substantially impairing general abilities. Therefore, we advocate for more research efforts to minimize the loss of general abilities acquired during LLM pre-training and to ultimately preserve them during model editing.
Image generation models can generate or edit images from a given text. Recent advancements in image generation technology, exemplified by DALL-E and Midjourney, have been groundbreaking. These advanced models, despite their impressive capabilities, are often trained on massive Internet datasets, making them susceptible to generating content that perpetuates social stereotypes and biases, which can lead to severe consequences. Prior research on assessing bias within image generation models suffers from several shortcomings, including limited accuracy, reliance on extensive human labor, and lack of comprehensive analysis. In this paper, we propose BiasPainter, a novel metamorphic testing framework that can accurately, automatically and comprehensively trigger social bias in image generation models. BiasPainter uses a diverse range of seed images of individuals and prompts the image generation models to edit these images using gender, race, and age-neutral queries. These queries span 62 professions, 39 activities, 57 types of objects, and 70 personality traits. The framework then compares the edited images to the original seed images, focusing on any changes related to gender, race, and age. BiasPainter adopts a testing oracle that these characteristics should not be modified when subjected to neutral prompts. Built upon this design, BiasPainter can trigger the social bias and evaluate the fairness of image generation models. To evaluate the effectiveness of BiasPainter, we use BiasPainter to test five widely-used commercial image generation software and models, such as stable diffusion and Midjourney. Experimental results show that 100\% of the generated test cases can successfully trigger social bias in image generation models.
Journalists must find stories in huge amounts of textual data (e.g. leaks, bills, press releases) as part of their jobs: determining when and why text becomes news can help us understand coverage patterns and help us build assistive tools. Yet, this is challenging because very few labelled links exist, language use between corpora is very different, and text may be covered for a variety of reasons. In this work we focus on news coverage of local public policy in the San Francisco Bay Area by the San Francisco Chronicle. First, we gather news articles, public policy documents and meeting recordings and link them using probabilistic relational modeling, which we show is a low-annotation linking methodology that outperforms other retrieval-based baselines. Second, we define a new task: newsworthiness prediction, to predict if a policy item will get covered. We show that different aspects of public policy discussion yield different newsworthiness signals. Finally we perform human evaluation with expert journalists and show our systems identify policies they consider newsworthy with 68% F1 and our coverage recommendations are helpful with an 84% win-rate.
We explore the creative problem-solving capabilities of modern large language models (LLMs) in a constrained setting. The setting requires circumventing a cognitive bias known in psychology as ''functional fixedness'' to use familiar objects in innovative or unconventional ways. To this end, we create MacGyver, an automatically generated dataset consisting of 1,600 real-world problems that deliberately trigger functional fixedness and require thinking 'out-of-the-box'. We then present our collection of problems to both LLMs and humans to compare and contrast their problem-solving abilities. We show that MacGyver is challenging for both groups, but in unique and complementary ways. For example, humans typically excel in solving problems that they are familiar with but may struggle with tasks requiring domain-specific knowledge, leading to a higher variance. On the other hand, LLMs, being exposed to a variety of highly specialized knowledge, attempt broader problems but are prone to overconfidence and propose actions that are physically infeasible or inefficient. We also provide a detailed error analysis of LLMs, and demonstrate the potential of enhancing their problem-solving ability with novel prompting techniques such as iterative step-wise reflection and divergent-convergent thinking. This work provides insight into the creative problem-solving capabilities of humans and AI and illustrates how psychological paradigms can be extended into large-scale tasks for comparing humans and machines.
Event extraction has attracted much attention in recent years due to its potential for many applications. However, recent studies observe some evaluation challenges, suggesting that reported scores might not reflect the true performance. In this work, we first identify and discuss these evaluation challenges, including the unfair comparisons resulting from different assumptions about data or different data preprocessing steps, the incompleteness of the current evaluation framework leading to potential dataset bias or data split bias, and low reproducibility of prior studies. To address these challenges, we propose TextEE, a standardized, fair, and reproducible benchmark for event extraction. TextEE contains standardized data preprocessing scripts and splits for more than ten datasets across different domains. In addition, we aggregate and re-implement over ten event extraction approaches published in recent years and conduct a comprehensive reevaluation. Finally, we explore the capability of large language models in event extraction and discuss some future challenges. We expect TextEE will serve as a reliable benchmark for event extraction, facilitating future research in the field.
Ensuring factual consistency is crucial in various natural language processing tasks, particularly in abstractive summarization, where preserving the integrity of information is paramount. Prior entailment-based approaches often generate factually inconsistent summaries and then train a classifier on the generated data. However, summaries produced by these approaches are either of low coherence or lack error-type coverage. To address these issues, we propose AMRFact, a novel framework that generates factually inconsistent summaries using Abstract Meaning Representation (AMR). Our approach parses factually correct summaries into AMR graphs and injects controlled factual inconsistencies to create negative examples, allowing for coherent factually inconsistent summaries to be generated with high error-type coverage. Additionally, we present a data selection module NegFilter based on natural language inference and BARTScore to ensure the quality of the generated negative samples. Experimental results demonstrate that our approach significantly outperforms previous systems on the AggreFact-SOTA dataset, showcasing its efficacy in assessing factuality in abstractive summarization.
Multimodal counterfactual reasoning is a vital yet challenging ability for AI systems. It involves predicting the outcomes of hypothetical circumstances based on vision and language inputs, which enables AI models to learn from failures and explore hypothetical scenarios. Despite its importance, there are only a few datasets targeting the counterfactual reasoning abilities of multimodal models. Among them, they only cover reasoning over synthetic environments or specific types of events (e.g. traffic collisions), making them hard to reliably benchmark the model generalization ability in diverse real-world scenarios and reasoning dimensions. To overcome these limitations, we develop a video question answering dataset, ACQUIRED: it consists of 3.9K annotated videos, encompassing a wide range of event types and incorporating both first and third-person viewpoints, which ensures a focus on real-world diversity. In addition, each video is annotated with questions that span three distinct dimensions of reasoning, including physical, social, and temporal, which can comprehensively evaluate the model counterfactual abilities along multiple aspects. We benchmark our dataset against several state-of-the-art language-only and multimodal models and experimental results demonstrate a significant performance gap (>13%) between models and humans. The findings suggest that multimodal counterfactual reasoning remains an open challenge and ACQUIRED is a comprehensive and reliable benchmark for inspiring future research in this direction.
Instruction tuning (IT) achieves impressive zero-shot generalization results by training large language models (LLMs) on a massive amount of diverse tasks with instructions. However, how to select new tasks to improve the performance and generalizability of IT models remains an open question. Training on all existing tasks is impractical due to prohibiting computation requirements, and randomly selecting tasks can lead to suboptimal performance. In this work, we propose active instruction tuning based on prompt uncertainty, a novel framework to identify informative tasks, and then actively tune the models on the selected tasks. We represent the informativeness of new tasks with the disagreement of the current model outputs over perturbed prompts. Our experiments on NIV2 and Self-Instruct datasets demonstrate that our method consistently outperforms other baseline strategies for task selection, achieving better out-of-distribution generalization with fewer training tasks. Additionally, we introduce a task map that categorizes and diagnoses tasks based on prompt uncertainty and prediction probability. We discover that training on ambiguous (prompt-uncertain) tasks improves generalization while training on difficult (prompt-certain and low-probability) tasks offers no benefit, underscoring the importance of task selection for instruction tuning.
Large Language Models (LLMs) have recently emerged as an effective tool to assist individuals in writing various types of content, including professional documents such as recommendation letters. Though bringing convenience, this application also introduces unprecedented fairness concerns. Model-generated reference letters might be directly used by users in professional scenarios. If underlying biases exist in these model-constructed letters, using them without scrutinization could lead to direct societal harms, such as sabotaging application success rates for female applicants. In light of this pressing issue, it is imminent and necessary to comprehensively study fairness issues and associated harms in this real-world use case. In this paper, we critically examine gender biases in LLM-generated reference letters. Drawing inspiration from social science findings, we design evaluation methods to manifest biases through 2 dimensions: (1) biases in language style and (2) biases in lexical content. We further investigate the extent of bias propagation by analyzing the hallucination bias of models, a term that we define to be bias exacerbation in model-hallucinated contents. Through benchmarking evaluation on 2 popular LLMs- ChatGPT and Alpaca, we reveal significant gender biases in LLM-generated recommendation letters. Our findings not only warn against using LLMs for this application without scrutinization, but also illuminate the importance of thoroughly studying hidden biases and harms in LLM-generated professional documents.
Large language models (LLMs) such as GPT-3 have demonstrated a strong capability to generate coherent and contextually relevant text. However, amidst their successes, a crucial issue persists: their generated outputs still lack commonsense at times. Moreover, fine-tuning the entire LLM towards more commonsensical outputs is computationally expensive if not infeasible. In this paper, we present a computation-efficient framework that steers a frozen Pre-Trained Language Model (PTLM) towards more commonsensical generation (i.e., producing a plausible output that incorporates a list of concepts in a meaningful way). Specifically, we first construct a reference-free evaluator that assigns a sentence with a commonsensical score by grounding the sentence to a dynamic commonsense knowledge base from four different relational aspects. We then use the scorer as the oracle for commonsense knowledge, and extend the controllable generation method called NADO to train an auxiliary head that guides a fixed PTLM to better satisfy the oracle. We test our framework on a series of GPT-2-, Flan-T5-, and Alpaca-based language models (LMs) on two constrained concept-to-sentence benchmarks. Human evaluation results demonstrate that our method consistently leads to the most commonsensical outputs.