An approach to improve network interpretability is via clusterability, i.e., splitting a model into disjoint clusters that can be studied independently. We find pretrained models to be highly unclusterable and thus train models to be more modular using an ``enmeshment loss'' function that encourages the formation of non-interacting clusters. Using automated interpretability measures, we show that our method finds clusters that learn different, disjoint, and smaller circuits for CIFAR-10 labels. Our approach provides a promising direction for making neural networks easier to interpret.

The exponential growth of the mobile app market underscores the importance of constant innovation and rapid response to user demands. As user satisfaction is paramount to the success of a mobile application (app), developers typically rely on user reviews, which represent user feedback that includes ratings and comments to identify areas for improvement. However, the sheer volume of user reviews poses challenges in manual analysis, necessitating automated approaches. Existing automated approaches either analyze only the target apps reviews, neglecting the comparison of similar features to competitors or fail to provide suggestions for feature enhancement. To address these gaps, we propose a Large Language Model (LLM)-based Competitive User Review Analysis for Feature Enhancement) (LLM-Cure), an approach powered by LLMs to automatically generate suggestion s for mobile app feature improvements. More specifically, LLM-Cure identifies and categorizes features within reviews by applying LLMs. When provided with a complaint in a user review, LLM-Cure curates highly rated (4 and 5 stars) reviews in competing apps related to the complaint and proposes potential improvements tailored to the target application. We evaluate LLM-Cure on 1,056,739 reviews of 70 popular Android apps. Our evaluation demonstrates that LLM-Cure significantly outperforms the state-of-the-art approaches in assigning features to reviews by up to 13% in F1-score, up to 16% in recall and up to 11% in precision. Additionally, LLM-Cure demonstrates its capability to provide suggestions for resolving user complaints. We verify the suggestions using the release notes that reflect the changes of features in the target mobile app. LLM-Cure achieves a promising average of 73% of the implementation of the provided suggestions.

Transformers are widely used across various applications, many of which yield sparse or partially filled attention matrices. Examples include attention masks designed to reduce the quadratic complexity of attention, sequence packing techniques, and recent innovations like tree masking for fast validation in MEDUSA. Despite the inherent sparsity in these matrices, the state-of-the-art algorithm Flash Attention still processes them with quadratic complexity as though they were dense. In this paper, we introduce Binary Block Masking, a highly efficient modification that enhances Flash Attention by making it mask-aware. We further propose two optimizations: one tailored for masks with contiguous non-zero patterns and another for extremely sparse masks. Our experiments on attention masks derived from real-world scenarios demonstrate up to a 9x runtime improvement. The implementation will be publicly released to foster further research and application.

We explore the feasibility and potential of building a ground-truth-free evaluation model to assess the quality of segmentations generated by the Segment Anything Model (SAM) and its variants in medical imaging. This evaluation model estimates segmentation quality scores by analyzing the coherence and consistency between the input images and their corresponding segmentation predictions. Based on prior research, we frame the task of training this model as a regression problem within a supervised learning framework, using Dice scores (and optionally other metrics) along with mean squared error to compute the training loss. The model is trained utilizing a large collection of public datasets of medical images with segmentation predictions from SAM and its variants. We name this model EvanySeg (Evaluation of Any Segmentation in Medical Images). Our exploration of convolution-based models (e.g., ResNet) and transformer-based models (e.g., ViT) suggested that ViT yields better performance for this task. EvanySeg can be employed for various tasks, including: (1) identifying poorly segmented samples by detecting low-percentile segmentation quality scores; (2) benchmarking segmentation models without ground truth by averaging quality scores across test samples; (3) alerting human experts to poor-quality segmentation predictions during human-AI collaboration by applying a threshold within the score space; and (4) selecting the best segmentation prediction for each test sample at test time when multiple segmentation models are available, by choosing the prediction with the highest quality score. Models and code will be made available at https://github.com/ahjolsenbics/EvanySeg.

Logical reading comprehension is a challenging task that entails grasping the underlying semantics of text and applying reasoning to deduce the correct answer. Prior researches have primarily focused on enhancing logical reasoning capabilities through Chain-of-Thought (CoT) or data augmentation. However, previous work constructing chain-of-thought rationales concentrates solely on analyzing correct options, neglecting the incorrect alternatives. Addtionally, earlier efforts on data augmentation by altering contexts rely on rule-based methods, which result in generated contexts that lack diversity and coherence. To address these issues, we propose a Premise-Oriented Data Augmentation (PODA) framework. This framework can generate CoT rationales including analyses for both correct and incorrect options, while constructing diverse and high-quality counterfactual contexts from incorrect candidate options. We integrate summarizing premises and identifying premises for each option into rationales. Subsequently, we employ multi-step prompts with identified premises to construct counterfactual context. To facilitate the model's capabilities to better differentiate the reasoning process associated with each option, we introduce a novel thought-path contrastive learning method that compares reasoning paths between the original and counterfactual samples. Experimental results on three representative LLMs demonstrate that our method can improve the baselines substantially across two challenging logical reasoning benchmarks (ReClor and LogiQA 2.0). The data and code are released at https://github.com/lalalamdbf/TPReasoner.

Resilience refers to the ability of systems to withstand, adapt to, and recover from disruptive events. While studies on resilience have attracted significant attention across various research domains, the precise definition of this concept within the field of cooperative artificial intelligence remains unclear. This paper addresses this gap by proposing a clear definition of `cooperative resilience' and outlining a methodology for its quantitative measurement. The methodology is validated in an environment with RL-based and LLM-augmented autonomous agents, subjected to environmental changes and the introduction of agents with unsustainable behaviors. These events are parameterized to create various scenarios for measuring cooperative resilience. The results highlight the crucial role of resilience metrics in analyzing how the collective system prepares for, resists, recovers from, sustains well-being, and transforms in the face of disruptions. These findings provide foundational insights into the definition, measurement, and preliminary analysis of cooperative resilience, offering significant implications for the broader field of AI. Moreover, the methodology and metrics developed here can be adapted to a wide range of AI applications, enhancing the reliability and effectiveness of AI in dynamic and unpredictable environments.

We introduce a novel embedding method diverging from conventional approaches by operating within function spaces of finite dimension rather than finite vector space, thus departing significantly from standard knowledge graph embedding techniques. Initially employing polynomial functions to compute embeddings, we progress to more intricate representations using neural networks with varying layer complexities. We argue that employing functions for embedding computation enhances expressiveness and allows for more degrees of freedom, enabling operations such as composition, derivatives and primitive of entities representation. Additionally, we meticulously outline the step-by-step construction of our approach and provide code for reproducibility, thereby facilitating further exploration and application in the field.

As the planet's largest economies struggle to forge consensus on the future of artificial intelligence, Brazil's Luiz Inacio Lula da Silva wants to ensure the developing world isn't left out of the debate. The Brazilian leader has added AI to his list of priorities for his country's presidency of the Group of 20 nations this year, seizing on the position to try to shape regulatory discussions that are raging from Europe to Asia to the United Nations, where the technology is expected to be a major theme of this week's General Assembly. Already seeking reforms to global institutions like the U.N. Security Council, Lula wants to use November's G20 leaders summit to craft a governance framework that includes the interests of Global South nations and forces AI superpowers China and the U.S. to the table, according to two people familiar with his views.

In this thesis, I evaluate the performance of Large Language Models (LLMs) on the Law School Admissions Test (LSAT), specifically the Logic Games section of the test. I focus on this section because it presents a complex logical reasoning task and thus is a valuable source of data for evaluating how modern, increasingly capable LLMs can handle hard logical reasoning tasks. I construct a dataset of LSAT logic games and their associated metadata, and extensively evaluate LLMs' performance in a Chain-of-Thought prompting setting. Given the weak performance in this setting, I explore other prompting frameworks on a smaller subset of the dataset, adapting ideas from Reflexion to this task. This results in a substantially improved accuracy of 70 percent for GPT-4 and 46 percent for GPT-3.5 on this data subset, highlighting the capacity of LLMs to revise their logical errors, despite initially weak performance. Finally, I analyze the types of logic games that models perform better or worse on, as well as the types of logical errors I observe from human annotation, providing detailed insights on the logical reasoning capabilities of LLMs.

Recently, large language models (LLMs) have demonstrated exceptional capabilities in serving as the foundation for AI assistants. One emerging application of LLMs, navigating through websites and interacting with UI elements across various web pages, remains somewhat underexplored. We introduce Steward, a novel LLM-powered web automation tool designed to serve as a cost-effective, scalable, end-to-end solution for automating web interactions. Traditional browser automation frameworks like Selenium, Puppeteer, and Playwright are not scalable for extensive web interaction tasks, such as studying recommendation algorithms on platforms like YouTube and Twitter. These frameworks require manual coding of interactions, limiting their utility in large-scale or dynamic contexts. Steward addresses these limitations by integrating LLM capabilities with browser automation, allowing for natural language-driven interaction with websites. Steward operates by receiving natural language instructions and reactively planning and executing a sequence of actions on websites, looping until completion, making it a practical tool for developers and researchers to use. It achieves high efficiency, completing actions in 8.52 to 10.14 seconds at a cost of $0.028 per action or an average of $0.18 per task, which is further reduced to 4.8 seconds and $0.022 through a caching mechanism. It runs tasks on real websites with a 40% completion success rate. We discuss various design and implementation challenges, including state representation, action sequence selection, system responsiveness, detecting task completion, and caching implementation.