The Map of Misbelief: Tracing Intrinsic and Extrinsic Hallucinations Through Attention Patterns

This paper presents a transformer-based approach for classifying hope expressions in text. Three architectures (BERT, GPT-2, and DeBERTa) were developed and compared for binary classification (Hope vs. Not Hope) and multiclass categorization (five hope-related categories). The BERT implementation achieved 83.65% binary and 74.87% multiclass accuracy, with superior performance in extended comparisons. GPT-2 showed the lowest accuracy, while DeBERTa had moderate results but at a higher computational cost. Error analysis highlighted architecture-specific strengths in detecting nuanced hope expres…

The article discusses the reconstruction of the intrinsic geometry of a manifold from noisy pairwise distance observations. It focuses on a diameter 1 d-dimensional manifold and a probability measure that is absolutely continuous with the volume measure. By observing noisy-distance random variables related to true geodesic distances, the authors propose a new framework for recovering distances among points in a dense subsample of the manifold, improving upon previous methods that relied on known moments of noise.

Link prediction is a crucial task in graph machine learning, applicable in areas like social recommendation and knowledge graph completion. Ensuring fairness in link prediction is vital, as biased outcomes can worsen societal inequalities. Traditional methods focus on demographic parity between intra-group and inter-group predictions, but this approach may overlook deeper disparities among subgroups. The authors propose a new framework for assessing fairness in link prediction that goes beyond demographic parity, aiming to better address systemic biases.

The study on entropy calibration of language models investigates whether the entropy of a model's text generation aligns with its log loss on human text. Previous findings indicate that models often exhibit miscalibration, where entropy increases and text quality declines with longer generations. This paper explores whether scaling can improve miscalibration and if calibration can be achieved without trade-offs, focusing on the relationship between dataset size and miscalibration behavior.

The paper titled 'AI-Salesman: Towards Reliable Large Language Model Driven Telemarketing' addresses the challenges of goal-driven persuasive dialogue in telemarketing using Large Language Models (LLMs). It highlights the limitations of previous works due to a lack of task-specific data and issues like strategic brittleness and factual hallucination. The authors introduce TeleSalesCorpus, a new dialogue dataset, and propose a dual-stage framework called AI-Salesman, which includes a Bayesian-supervised reinforcement learning algorithm for training and a Dynamic Outline-Guided Agent for inferen…

Deep neural networks have shown remarkable success in image classification but are still susceptible to adversarial examples. Traditional adversarial training methods improve robustness but often overlook inter-class feature overlap, which contributes to vulnerability. This study introduces a new adversarial training framework that reduces inter-class proximity by projecting out dependencies from both adversarial and clean samples in the feature space. The proposed method enhances feature separability and theoretically lowers the Lipschitz constant of neural networks, improving generalization.

Large Language Models (LLMs) have progressed to become agentic systems capable of complex task execution. However, during the fine-tuning process for agent-specific tasks, safety concerns are often neglected. This study reveals that aligned LLMs can unintentionally become misaligned, increasing the risk of harmful task execution. To mitigate these risks, the authors propose the Prefix INjection Guard (PING), which uses natural language prefixes to guide LLMs in refusing harmful requests while maintaining performance on benign tasks.

Large language models (LLMs) are known for their impressive text generation abilities but often produce factually incorrect content, a phenomenon termed 'hallucination.' This issue is particularly concerning in critical fields such as healthcare and finance. Traditional methods for detecting these inaccuracies require multiple API calls, leading to increased costs and latency. The introduction of CONFACTCHECK offers a novel solution, allowing for efficient hallucination detection by ensuring consistency in factual responses generated by LLMs without needing external knowledge bases.