For this study, Lindsey and his colleagues worked to lay down some of that groundwork. Previous research has shown that various dimensions of LLMs' behavior--from whether they are talking about weddings to persistent traits such as sycophancy--are associated with specific patterns of activity in the simulated neurons that constitute LLMs. Those patterns can be written down as a long string of numbers, in which each number represents how active a specific neuron is when the model is expressing that behavior. Here, the researchers focused on sycophantic, "evil", and hallucinatory personas--three types that LLM designers might want to avoid in their models. To identify those patterns, the team devised a fully automated pipeline that can map out that pattern given a brief text description of a persona.

Anything powered by a large language model is generally fed right back into it for more training, so it's even less secure than a regular search. Lots of ChatGPT users got a practical demonstration of this last night, when a huge amount of their queries were found to be searchable via a simple Google prefix. This is a complicated situation, so let me break it down. ChatGPT has a share feature that lets you easily send info to another user with a link. But apparently the info in these semi-personal discussions with the "AI" chatbot was posted somewhere that Google could crawl and index.

Reusing established theorems and formulas is central to mathematical problem solving, serving as essential building blocks for tackling increasingly complex challenges. Recent work, TroVE, argues that code-generating Large Language Models (LLMs) can benefit similarly on the MATH benchmark by inducing and reusing higher-level toolboxes. By allocating computational budget across an ensemble of three modes -- directly generating code, creating tools, and reusing tools -- TroVE claims to outperform a PRIMITIVE baseline that only performs direct generation. However, recent analysis (Berlot-Attwell et al., 2024) casts doubt on these gains, noting that the tools created are often trivial or rarely reused, suggesting that improvements may stem from self-consistency or self-correction. In this work, we re-evaluate TroVE on MATH, analyze the impact of each of its modes, and show that its benefit does not come from these mechanisms, but simply from a higher computational budget spent for TroVE compared to PRIMITIVE. To this end, we also perform a small correction in the original implementation of TroVE's selection mechanism, boosting TroVE's performance on MATH by 3\% in accuracy. After matching for compute, the benefit of TroVE reduces to a marginal improvement of 1\%, suggesting that this toolbox approach does not provide a significant benefit on MATH.

We ask whether contemporary LLMs are able to perform natural language inference (NLI) tasks on mathematical texts. We call this the Math NLI problem. We construct a corpus of Math NLI pairs whose premises are from extant mathematical text and whose hypotheses and gold labels were provided by people with experience in both research-level mathematics and also in the NLI field. We also investigate the quality of corpora using the same premises but whose hypotheses are provided by LLMs themselves. We not only investigate the performance but also the inter-group consistency of the diverse group of LLMs. We have both positive and negative findings. Among our positive findings: in some settings, using a majority vote of LLMs is approximately equivalent to using human-labeled data in the Math NLI area. On the negative side: LLMs still struggle with mathematical language. They occasionally fail at even basic inferences. Current models are not as prone to hypothesis-only "inference" in our data the way the previous generation had been. In addition to our findings, we also provide our corpora as data to support future work on Math NLI.

Causal inference remains a fundamental challenge for large language models. Recent advances in internal reasoning with large language models have sparked interest in whether state-of-the-art reasoning models can robustly perform causal discovery-a task where conventional models often suffer from severe overfitting and near-random performance under data perturbations. We study causal discovery on the Corr2Cause benchmark using the emergent OpenAI's o-series and DeepSeek-R model families and find that these reasoning-first architectures achieve significantly greater native gains than prior approaches. To capitalize on these strengths, we introduce a modular in-context pipeline inspired by the Tree-of-Thoughts and Chain-of-Thoughts methodologies, yielding nearly three-fold improvements over conventional baselines. We further probe the pipeline's impact by analyzing reasoning chain length, complexity, and conducting qualitative and quantitative comparisons between conventional and reasoning models. Our findings suggest that while advanced reasoning models represent a substantial leap forward, carefully structured in-context frameworks are essential to maximize their capabilities and offer a generalizable blueprint for causal discovery across diverse domains.

With the development of multimodal reasoning models, Computer Use Agents (CUAs), akin to Jarvis from \textit{"Iron Man"}, are becoming a reality. GUI grounding is a core component for CUAs to execute actual actions, similar to mechanical control in robotics, and it directly leads to the success or failure of the system. It determines actions such as clicking and typing, as well as related parameters like the coordinates for clicks. Current end-to-end grounding models still achieve less than 65\% accuracy on challenging benchmarks like ScreenSpot-pro and UI-Vision, indicating they are far from being ready for deployment. % , as a single misclick can result in unacceptable consequences. In this work, we conduct an empirical study on the training of grounding models, examining details from data collection to model training. Ultimately, we developed the \textbf{Phi-Ground} model family, which achieves state-of-the-art performance across all five grounding benchmarks for models under $10B$ parameters in agent settings. In the end-to-end model setting, our model still achieves SOTA results with scores of \textit{\textbf{43.2}} on ScreenSpot-pro and \textit{\textbf{27.2}} on UI-Vision. We believe that the various details discussed in this paper, along with our successes and failures, not only clarify the construction of grounding models but also benefit other perception tasks. Project homepage: \href{https://zhangmiaosen2000.github.io/Phi-Ground/}{https://zhangmiaosen2000.github.io/Phi-Ground/}

While question-answering~(QA) benchmark performance is an automatic and scalable method to compare LLMs, it is an indirect method of evaluating their underlying problem-solving capabilities. Therefore, we propose a holistic and generalizable framework based on \emph{cascaded question disclosure} that provides a more accurate estimate of the models' problem-solving capabilities while maintaining the scalability and automation. This approach collects model responses in a stagewise manner with each stage revealing partial information about the question designed to elicit generalized reasoning in LLMs. We find that our approach not only provides a better comparison between LLMs, but also induces better intermediate traces in models compared to the standard QA paradigm. We empirically verify this behavior on diverse reasoning and knowledge-heavy QA datasets by comparing LLMs of varying sizes and families. Our approach narrows the performance gap observed in the standard QA evaluation settings, indicating that the prevalent indirect QA paradigm of evaluation overestimates the differences in performance between models. We further validate our findings by extensive ablation studies.

Knowledge graphs (KGs) often contain sufficient information to support the inference of new facts. Identifying logical rules not only improves the completeness of a knowledge graph but also enables the detection of potential errors, reveals subtle data patterns, and enhances the overall capacity for reasoning and interpretation. However, the complexity of such rules, combined with the unique labeling conventions of each KG, can make them difficult for humans to understand. In this paper, we explore the potential of large language models to generate natural language explanations for logical rules. Specifically, we extract logical rules using the AMIE 3.5.1 rule discovery algorithm from the benchmark dataset FB15k-237 and two large-scale datasets, FB-CVT-REV and FB+CVT-REV. We examine various prompting strategies, including zero- and few-shot prompting, including variable entity types, and chain-of-thought reasoning. We conduct a comprehensive human evaluation of the generated explanations based on correctness, clarity, and hallucination, and also assess the use of large language models as automatic judges. Our results demonstrate promising performance in terms of explanation correctness and clarity, although several challenges remain for future research. All scripts and data used in this study are publicly available at https://github.com/idirlab/KGRule2NL}{https://github.com/idirlab/KGRule2NL.

Access to medical imaging and associated text data has the potential to drive major advances in healthcare research and patient outcomes. However, the presence of Protected Health Information (PHI) and Personally Identifiable Information (PII) in Digital Imaging and Communications in Medicine (DICOM) files presents a significant barrier to the ethical and secure sharing of imaging datasets. This paper presents a hybrid de-identification framework developed by Impact Business Information Solutions (IBIS) that combines rule-based and AI-driven techniques, and rigorous uncertainty quantification for comprehensive PHI/PII removal from both metadata and pixel data. Our approach begins with a two-tiered rule-based system targeting explicit and inferred metadata elements, further augmented by a large language model (LLM) fine-tuned for Named Entity Recognition (NER), and trained on a suite of synthetic datasets simulating realistic clinical PHI/PII. For pixel data, we employ an uncertainty-aware Faster R-CNN model to localize embedded text, extract candidate PHI via Optical Character Recognition (OCR), and apply the NER pipeline for final redaction. Crucially, uncertainty quantification provides confidence measures for AI-based detections to enhance automation reliability and enable informed human-in-the-loop verification to manage residual risks. This uncertainty-aware deidentification framework achieves robust performance across benchmark datasets and regulatory standards, including DICOM, HIPAA, and TCIA compliance metrics. By combining scalable automation, uncertainty quantification, and rigorous quality assurance, our solution addresses critical challenges in medical data de-identification and supports the secure, ethical, and trustworthy release of imaging data for research.

General robotic grasping systems require accurate object affordance perception in diverse open-world scenarios following human instructions. However, current studies suffer from the problem of lacking reasoning-based large-scale affordance prediction data, leading to considerable concern about open-world effectiveness. To address this limitation, we build a large-scale grasping-oriented affordance segmentation benchmark with human-like instructions, named RAGNet. It contains 273k images, 180 categories, and 26k reasoning instructions. The images cover diverse embodied data domains, such as wild, robot, ego-centric, and even simulation data. They are carefully annotated with an affordance map, while the difficulty of language instructions is largely increased by removing their category name and only providing functional descriptions. Furthermore, we propose a comprehensive affordance-based grasping framework, named AffordanceNet, which consists of a VLM pre-trained on our massive affordance data and a grasping network that conditions an affordance map to grasp the target. Extensive experiments on affordance segmentation benchmarks and real-robot manipulation tasks show that our model has a powerful open-world generalization ability. Our data and code is available at https://github.com/wudongming97/AffordanceNet.