Research Papers

The increasing use of AI technologies has led to increasing AI incidents, posing risks and causing harm to individuals, organizations, and society. This study recognizes and addresses the lack of standardized protocols for reliably and comprehensively gathering such incident data crucial for preventing future incidents and developing mitigating strategies. Specifically, this study analyses existing open-access AI-incident databases through a systematic methodology and identifies nine gaps in current AI incident reporting practices. Further, it proposes nine actionable recommendations to enhance standardization efforts to address these gaps. Ensuring the trustworthiness of enabling technologies such as AI is necessary for sustainable digital transformation. Our research promotes the development of standards to prevent future AI incidents and promote trustworthy AI, thus facilitating achieving the UN sustainable development goals. Through international cooperation, stakeholders can unlock the transformative potential of AI, enabling a sustainable and inclusive future for all.

A rapidly developing application of LLMs in XAI is to convert quantitative explanations such as SHAP into user-friendly narratives to explain the decisions made by smaller prediction models. Evaluating the narratives without relying on human preference studies or surveys is becoming increasingly important in this field. In this work we propose a framework and explore several automated metrics to evaluate LLM-generated narratives for explanations of tabular classification tasks. We apply our approach to compare several state-of-the-art LLMs across different datasets and prompt types. As a demonstration of their utility, these metrics allow us to identify new challenges related to LLM hallucinations for XAI narratives.

This technical report presents Yi-Lightning, our latest flagship large language model (LLM). It achieves exceptional performance, ranking 6th overall on Chatbot Arena, with particularly strong results (2nd to 4th place) in specialized categories including Chinese, Math, Coding, and Hard Prompts. Yi-Lightning leverages an enhanced Mixture-of-Experts (MoE) architecture, featuring advanced expert segmentation and routing mechanisms coupled with optimized KV-caching techniques. Our development process encompasses comprehensive pre-training, supervised fine-tuning (SFT), and reinforcement learning from human feedback (RLHF), where we devise deliberate strategies for multi-stage training, synthetic data construction, and reward modeling. Furthermore, we implement RAISE (Responsible AI Safety Engine), a four-component framework to address safety issues across pre-training, post-training, and serving phases. Empowered by our scalable super-computing infrastructure, all these innovations substantially reduce training, deployment and inference costs while maintaining high-performance standards. With further evaluations on public academic benchmarks, Yi-Lightning demonstrates competitive performance against top-tier LLMs, while we observe a notable disparity between traditional, static benchmark results and real-world, dynamic human preferences. This observation prompts a critical reassessment of conventional benchmarks' utility in guiding the development of more intelligent and powerful AI systems for practical applications. Yi-Lightning is now available through our developer platform at https://platform.lingyiwanwu.com.

Reinforcement Learning from Human Feedback (RLHF) has recently surged in popularity, particularly for aligning large language models and other AI systems with human intentions. At its core, RLHF can be viewed as a specialized instance of Preference-based Reinforcement Learning (PbRL), where the preferences specifically originate from human judgments rather than arbitrary evaluators. Despite this connection, most existing approaches in both RLHF and PbRL primarily focus on optimizing a mean reward objective, neglecting scenarios that necessitate risk-awareness, such as AI safety, healthcare, and autonomous driving. These scenarios often operate under a one-episode-reward setting, which makes conventional risk-sensitive objectives inapplicable. To address this, we explore and prove the applicability of two risk-aware objectives to PbRL : nested and static quantile risk objectives. We also introduce Risk-AwarePbRL (RA-PbRL), an algorithm designed to optimize both nested and static objectives. Additionally, we provide a theoretical analysis of the regret upper bounds, demonstrating that they are sublinear with respect to the number of episodes, and present empirical results to support our findings. Our code is available in https://github.com/aguilarjose11/PbRLNeurips.

As machine learning models evolve, maintaining transparency demands more human-centric explainable AI techniques. Counterfactual explanations, with roots in human reasoning, identify the minimal input changes needed to obtain a given output and, hence, are crucial for supporting decision-making. Despite their importance, the evaluation of these explanations often lacks grounding in user studies and remains fragmented, with existing metrics not fully capturing human perspectives. To address this challenge, we developed a diverse set of 30 counterfactual scenarios and collected ratings across 8 evaluation metrics from 206 respondents. Subsequently, we fine-tuned different Large Language Models (LLMs) to predict average or individual human judgment across these metrics. Our methodology allowed LLMs to achieve an accuracy of up to 63% in zero-shot evaluations and 85% (over a 3-classes prediction) with fine-tuning across all metrics. The fine-tuned models predicting human ratings offer better comparability and scalability in evaluating different counterfactual explanation frameworks.

Operationalizing AI ethics and safety principles and frameworks is essential to realizing the potential benefits and mitigating potential harms caused by AI systems. To that end, actors across industry, academia, and regulatory bodies have created formal taxonomies of harm to support operationalization efforts. These include novel holistic methods that go beyond exclusive reliance on technical benchmarking. However, our paper argues that such taxonomies must also be transferred into local categories to be readily implemented in sector-specific AI safety operationalization efforts, and especially in underresourced or high-risk sectors. This is because many sectors are constituted by discourses, norms, and values that "refract" or even directly conflict with those operating in society more broadly. Drawing from emerging anthropological theories of human rights, we propose that the process of "vernacularization"--a participatory, decolonial practice distinct from doctrinary "translation" (the dominant mode of AI safety operationalization)--can help bridge this gap. To demonstrate this point, we consider the education sector, and identify precisely how vernacularizing a leading holistic taxonomy of harm leads to a clearer view of how harms AI systems may cause are substantially intensified when deployed in educational spaces. We conclude by discussing the generalizability of vernacularization as a useful AI safety methodology.

In this report, we propose the implementation of national registries for frontier AI models as a foundational tool for AI governance. We explore the rationale, design, and implementation of such registries, drawing on comparisons with registries in analogous industries to make recommendations for a registry that is efficient, unintrusive, and which will bring AI governance closer to parity with the governmental insight into other high-impact industries. We explore key information that should be collected, including model architecture, model size, compute and data used during training, and we survey the viability and utility of evaluations developed specifically for AI. Our proposal is designed to provide governmental insight and enhance AI safety while fostering innovation and minimizing the regulatory burden on developers. By providing a framework that respects intellectual property concerns and safeguards sensitive information, this registry approach supports responsible AI development without impeding progress. We propose that timely and accurate registration should be encouraged primarily through injunctive action, by requiring third parties to use only registered models, and secondarily through direct financial penalties for non-compliance. By providing a comprehensive framework for AI model registries, we aim to support policymakers in developing foundational governance structures to monitor and mitigate risks associated with advanced AI systems.

With the rise of SSL and ASR technologies, the Wav2Vec2 ASR-based model has been fine-tuned for automated speech disorder quality assessment tasks, yielding impressive results and setting a new baseline for Head and Neck Cancer speech contexts. This demonstrates that the ASR dimension from Wav2Vec2 closely aligns with assessment dimensions. Despite its effectiveness, this system remains a black box with no clear interpretation of the connection between the model ASR dimension and clinical assessments. This paper presents the first analysis of this baseline model for speech quality assessment, focusing on intelligibility and severity tasks. We conduct a layer-wise analysis to identify key layers and compare different SSL and ASR Wav2Vec2 models based on pre-trained data. Additionally, post-hoc XAI methods, including Canonical Correlation Analysis (CCA) and visualization techniques, are used to track model evolution and visualize embeddings for enhanced interpretability.

In Explainable AI (XAI), Shapley values are a popular model-agnostic framework for explaining predictions made by complex machine learning models. The computation of Shapley values requires estimating non-trivial contribution functions representing predictions with only a subset of the features present. As the number of these terms grows exponentially with the number of features, computational costs escalate rapidly, creating a pressing need for efficient and accurate approximation methods. For tabular data, the KernelSHAP framework is considered the state-of-the-art model-agnostic approximation framework. KernelSHAP approximates the Shapley values using a weighted sample of the contribution functions for different feature subsets. We propose a novel modification of KernelSHAP which replaces the stochastic weights with deterministic ones to reduce the variance of the resulting Shapley value approximations. This may also be combined with our simple, yet effective modification to the KernelSHAP variant implemented in the popular Python library SHAP. Additionally, we provide an overview of established methods. Numerical experiments demonstrate that our methods can reduce the required number of contribution function evaluations by $5\%$ to $50\%$ while preserving the same accuracy of the approximated Shapley values -- essentially reducing the running time by up to $50\%$. These computational advancements push the boundaries of the feature dimensionality and number of predictions that can be accurately explained with Shapley values within a feasible runtime.

Enhanced Infield Agriculture with Interpretable Machine Learning Approaches for Crop Classification

Sudi Murindanyi, Joyce Nakatumba-Nabende, Rahman Sanya +2

The increasing popularity of Artificial Intelligence in recent years has led to a surge in interest in image classification, especially in the agricultural sector. With the help of Computer Vision, Machine Learning, and Deep Learning, the sector has undergone a significant transformation, leading to the development of new techniques for crop classification in the field. Despite the extensive research on various image classification techniques, most have limitations such as low accuracy, limited use of data, and a lack of reporting model size and prediction. The most significant limitation of all is the need for model explainability. This research evaluates four different approaches for crop classification, namely traditional ML with handcrafted feature extraction methods like SIFT, ORB, and Color Histogram; Custom Designed CNN and established DL architecture like AlexNet; transfer learning on five models pre-trained using ImageNet such as EfficientNetV2, ResNet152V2, Xception, Inception-ResNetV2, MobileNetV3; and cutting-edge foundation models like YOLOv8 and DINOv2, a self-supervised Vision Transformer Model. All models performed well, but Xception outperformed all of them in terms of generalization, achieving 98% accuracy on the test data, with a model size of 80.03 MB and a prediction time of 0.0633 seconds. A key aspect of this research was the application of Explainable AI to provide the explainability of all the models. This journal presents the explainability of Xception model with LIME, SHAP, and GradCAM, ensuring transparency and trustworthiness in the models' predictions. This study highlights the importance of selecting the right model according to task-specific needs. It also underscores the important role of explainability in deploying AI in agriculture, providing insightful information to help enhance AI-driven crop management strategies.

Multi-objective evolutionary algorithms (MOEAs) have emerged as powerful tools for solving complex optimization problems characterized by multiple, often conflicting, objectives. While advancements have been made in computational efficiency as well as diversity and convergence of solutions, a critical challenge persists: the internal evolutionary mechanisms are opaque to human users. Drawing upon the successes of explainable AI in explaining complex algorithms and models, we argue that the need to understand the underlying evolutionary operators and population dynamics within MOEAs aligns well with a visual analytics paradigm. This paper introduces ParetoTracker, a visual analytics framework designed to support the comprehension and inspection of population dynamics in the evolutionary processes of MOEAs. Informed by preliminary literature review and expert interviews, the framework establishes a multi-level analysis scheme, which caters to user engagement and exploration ranging from examining overall trends in performance metrics to conducting fine-grained inspections of evolutionary operations. In contrast to conventional practices that require manual plotting of solutions for each generation, ParetoTracker facilitates the examination of temporal trends and dynamics across consecutive generations in an integrated visual interface. The effectiveness of the framework is demonstrated through case studies and expert interviews focused on widely adopted benchmark optimization problems.

SLM as Guardian: Pioneering AI Safety with Small Language Models

Ohjoon Kwon, Donghyeon Jeon, Nayoung Choi +6

Most prior safety research of large language models (LLMs) has focused on enhancing the alignment of LLMs to better suit the safety requirements of humans. However, internalizing such safeguard features into larger models brought challenges of higher training cost and unintended degradation of helpfulness. To overcome such challenges, a modular approach employing a smaller LLM to detect harmful user queries is regarded as a convenient solution in designing LLM-based system with safety requirements. In this paper, we leverage a smaller LLM for both harmful query detection and safeguard response generation. We introduce our safety requirements and the taxonomy of harmfulness categories, and then propose a multi-task learning mechanism fusing the two tasks into a single model. We demonstrate the effectiveness of our approach, providing on par or surpassing harmful query detection and safeguard response performance compared to the publicly available LLMs.