As AIs rapidly advance and become more agentic, the risk they pose is governed not only by their capabilities but increasingly by their propensities, including goals and values. Tracking the emergence of goals and values has proven a longstanding problem, and despite much interest over the years it remains unclear whether current AIs have meaningful values. We propose a solution to this problem, leveraging the framework of utility functions to study the internal coherence of AI preferences. Surprisingly, we find that independently-sampled preferences in current LLMs exhibit high degrees of structural coherence, and moreover that this emerges with scale. These findings suggest that value systems emerge in LLMs in a meaningful sense, a finding with broad implications. To study these emergent value systems, we propose utility engineering as a research agenda, comprising both the analysis and control of AI utilities. We uncover problematic and often shocking values in LLM assistants despite existing control measures. These include cases where AIs value themselves over humans and are anti-aligned with specific individuals. To constrain these emergent value systems, we propose methods of utility control. As a case study, we show how aligning utilities with a citizen assembly reduces political biases and generalizes to new scenarios. Whether we like it or not, value systems have already emerged in AIs, and much work remains to fully understand and control these emergent representations.

Automated red teaming holds substantial promise for uncovering and mitigating the risks associated with the malicious use of large language models (LLMs), yet the field lacks a standardized evaluation framework to rigorously assess new methods. To address this issue, we introduce HarmBench, a standardized evaluation framework for automated red teaming. We identify several desirable properties previously unaccounted for in red teaming evaluations and systematically design HarmBench to meet these criteria. Using HarmBench, we conduct a large-scale comparison of 18 red teaming methods and 33 target LLMs and defenses, yielding novel insights. We also introduce a highly efficient adversarial training method that greatly enhances LLM robustness across a wide range of attacks, demonstrating how HarmBench enables codevelopment of attacks and defenses. We open source HarmBench at https://github.com/centerforaisafety/HarmBench.

Performance on popular ML benchmarks is highly correlated with model scale, suggesting that most benchmarks tend to measure a similar underlying factor of general model capabilities. However, substantial research effort remains devoted to designing new benchmarks, many of which claim to measure novel phenomena. In the spirit of the Bitter Lesson, we leverage spectral analysis to measure an underlying capabilities component, the direction in benchmark-performance-space which explains most variation in model performance. In an extensive analysis of existing safety benchmarks, we find that variance in model performance on many safety benchmarks is largely explained by the capabilities component. In response, we argue that safety research should prioritize metrics which are not highly correlated with scale. Our work provides a lens to analyze both novel safety benchmarks and novel safety methods, which we hope will enable future work to make differential progress on safety.

This paper presents a new end-to-end signal classification method using the signed cumulative distribution transform (SCDT). We adopt a transport generative model to define the classification problem. We then make use of mathematical properties of the SCDT to render the problem easier in transform domain, and solve for the class of an unknown sample using a nearest local subspace (NLS) search algorithm in SCDT domain. Experiments show that the proposed method provides high accuracy classification results while being computationally cheap, data efficient, and robust to out-of-distribution samples with respect to the existing end-to-end classification methods. The implementation of the proposed method in Python language is integrated as a part of the software package PyTransKit [1].

The ability to perform quantitative and automated neurological assessment could enhance diagnosis and treatment in the pre-hospital setting, such as during telemedicine or emergency medical services (EMS) encounters. Such a tool could be developed by adapting clinically significant information such as symmetry of eye movement or conjugate eye movement. Here we describe a digital camera-based eye tracking method “NeuroGaze” to capture the symmetry of eye movement while performing neurological eye examination. The proposed method was developed based on detecting the center of the pupil for both eyes from a given video and measuring eye conjugacy by transforming the pupil center coordinates to relative gaze. The method was tested on healthy volunteers while performing three neurological eye examinations1. We also compared our proposed approach to state-of-the-art digital camera-based eye-tracking methods and commercial off-the-shelf (COTS) eye trackers. NeuroGaze outperformed digital camera-based eye tracking methods by reporting a mean Spearman rank-order correlation coefficient of 0.86 for the H-test, 0.87 for the Dot-test, and 0.56 for the OKN-Test, and shows similarity in trends for the relative gaze trajectories with a noticeable offset in the scale of the relative gaze angle compared to COTS eye tracker (see Fig.1). The study demonstrates that by using a pupil-center-based eye-tracking method, a digital camera can measure clinically relevant information regarding eye movement.

Facial weakness is a symptom commonly associated to lack of facial muscle control due to neurological injury. Several diseases are associated with facial weakness such as stroke and Bell's palsy. The use of digital imaging through mobile phones, tablets, personal computers and other devices could provide timely opportunity for detection, which if accurate enough can improve treatment by enabling faster patient triage and recovery progress monitoring. Most of the existing facial weakness detection approaches from static images are based on facial landmarks from which geometric features can be calculated. Landmark-based methods, however, can suffer from inaccuracies in face landmarks localization. In this study, We also experimentally evaluate the performance of several feature extraction methods for measuring facial weakness, including the landmark-based features, as well as intensity-based features on a neurologist-certified dataset that comprises 186 images of normal, 125 images of left facial weakness, and 126 images of right facial weakness. We demonstrate that, for the application of facial weakness detection from single (static) images, approaches that incorporate the Histogram of Oriented Gradients (HoG) features tend to be more accurate.

Generating captions for images is a task that has recently received considerable attention.In this work we focus on caption generation for abstract scenes, or object layouts where the only information provided is a set of objects and their locations.We propose OBJ2TEXT, a sequence-tosequence model that encodes a set of objects and their locations as an input sequence using an LSTM network, and decodes this representation using an LSTM language model.We show that our model, despite encoding object layouts as a sequence, can represent spatial relationships between objects, and generate descriptions that are globally coherent and semantically relevant.We test our approach in a task of object-layout captioning by using only object annotations as inputs.We additionally show that our model, combined with a state-of-the-art object detector, improves an image captioning model from 0.863 to 0.950 (CIDEr score) in the test benchmark of the standard MS-COCO Captioning task.

The rationale of XML design is to transfer and store data at different levels. A key feature of these levels in an XML document is to identify its components for additional processing. XML components can expose sensitive information after application

Text detection in natural scene images is an important prerequisite for many content-based image analysis tasks. In this paper, we propose an accurate and robust method for detecting texts in natural scene images. A fast and effective pruning algorithm is designed to extract Maximally Stable Extremal Regions (MSERs) as character candidates using the strategy of minimizing regularized variations. Character candidates are grouped into text candidates by the single-link clustering algorithm, where distance weights and clustering threshold are learned automatically by a novel self-training distance metric learning algorithm. The posterior probabilities of text candidates corresponding to non-text are estimated with a character classifier; text candidates with high non-text probabilities are eliminated and texts are identified with a text classifier. The proposed system is evaluated on the ICDAR 2011 Robust Reading Competition database; the f-measure is over 76%, much better than the state-of-the-art performance of 71%. Experiments on multilingual, street view, multi-orientation and even born-digital databases also demonstrate the effectiveness of the proposed method.

In ensemble learning, a higher accuracy can be achieved by integrating some classifiers instead of all the classifiers. But, it is very difficult to select the best classifier combination which can be seen as an optimization problem, from a pool of classifiers. To deal with this problem, we propose a new classifier selection method, Sorting-based Dynamic Classifier Ensemble Selection (SDES), which consists of two stages: (1) classifier sorting, and (2) dynamic ensemble selection on sorted classifier sequence. In the first stage, classifiers are sorted based on diversity, to avoid searching for the nearest neighbors in dynamic ensemble selection methods and greatly improve the selection efficiency. In the second stage, the optimal subset of classifiers is selected from the sorted classifier sequence based on confidence of test samples, to guarantee high accuracy of the optimal classifier subset. Experimental results have shown the effectiveness and high efficiency of the proposed method.

Text detection in images is an emerging area of interest with a growing motivation to researchers. Various methodologies have been developed to localize text contained in scene images. One main application of localizing scene image text is to produce a real time support to visually impaired persons. To design a real-time support platform for visually impaired persons, classification of textual information, i.e. character and non-character information can provide a baseline for further research. However, the challenge exists in choosing the optimum classifier for this purpose. In this work, first, we used Maximally Stable Extremal Regions (MSERs) to detect character candidates in a scene image; then, we trained several classifiers, i.e., AdaboostM1, Bayesian Logistic Regression, Naïve Bayes, and Bayes Net, to classify MSERs as characters and non-characters; and finally, we compared and analyzed the performances of these classifiers empirically. From experiments, it has been concluded that Bayesian Logistic Regression provides the better accuracy over the other three classifiers. This work argues that MSER based character candidates extraction and Bayesian Logistic Regression based text classification are two prominent and potential techniques in scene text detection.

We propose and implement a robust text detection system, which is a prominent step-in for text retrieval in natural scene images or videos. Our system includes several key components: (1) A fast and effective pruning algorithm is designed to extract Maximally Stable Extremal Regions as character candidates using the strategy of minimizing regularized variations. (2) Character candidates are grouped into text candidates by the single-link clustering algorithm, where distance weights and threshold of clustering are learned automatically by a novel self-training distance metric learning algorithm. (3) The posterior probabilities of text candidates corresponding to non-text are estimated with an character classifier; text candidates with high probabilities are then eliminated and finally texts are identified with a text classifier. The proposed system is evaluated on the ICDAR 2011 Robust Reading Competition dataset and a publicly available multilingual dataset; the f measures are over 76% and 74% which are significantly better than the state-of-the-art performances of 71% and 65%, respectively.