Model Editing as a Double-Edged Sword: Steering Agent Ethical Behavior Toward Beneficence or Harm

The introduction of Steering Activations towards Leakage-free Thinking (SALT) addresses a critical privacy challenge faced by Large Language Models (LLMs), which often leak sensitive information through their internal reasoning processes. SALT aims to mitigate this leakage by injecting targeted steering vectors into the model's hidden states, ensuring that the reasoning capabilities are preserved while enhancing privacy.

A novel approach called Vision-align-to-Language integrated Knowledge Graph (VaLiK) has been proposed to enhance reasoning in Large Language Models (LLMs) by constructing Multimodal Knowledge Graphs (MMKGs) without the need for manual annotations. This method aims to address challenges such as incomplete knowledge and hallucination artifacts that LLMs face due to the limitations of traditional Knowledge Graphs (KGs).

A recent evaluation of large language models (LLMs) in academic library reference services examined their ability to provide equitable support across diverse user demographics, including sex, race, and institutional roles. The study found no significant differentiation in responses based on race or ethnicity, with only minor evidence of bias against women in one model. LLMs showed nuanced responses tailored to users' institutional roles, reflecting professional norms.

A novel learning framework utilizing Large Language Models (LLMs) has been introduced to enhance the generalization capabilities of Neural Combinatorial Optimization (NCO) for Vehicle Routing Problems (VRPs). This approach addresses the significant performance drop observed when NCO models trained on small-scale instances are applied to larger scenarios, primarily due to distributional shifts between training and testing data.

A recent study highlights the limitations of Large Language Models (LLMs) in understanding Tunisian Arabic, also known as Tunizi. This research introduces a new dataset that includes parallel translations in Tunizi, standard Tunisian Arabic, and English, aiming to benchmark LLMs on their comprehension of this low-resource language. The findings indicate that the neglect of such dialects may hinder millions of Tunisians from engaging with AI in their native language.

A new benchmark named MUCH has been introduced to assess Claim-level Uncertainty Quantification (UQ) in Large Language Models (LLMs). This benchmark includes 4,873 samples in English, French, Spanish, and German, and provides 24 generation logits per token, enhancing the evaluation of UQ methods under realistic conditions.

LangMark has been introduced as a new multilingual dataset aimed at enhancing automatic post-editing (APE) for machine-translated texts, featuring 206,983 triplets across seven languages including Brazilian Portuguese, French, and Japanese. This dataset is human-annotated by expert linguists to improve translation quality and reduce reliance on human intervention.

A new framework called Aspect-Based Causal Abstention (ABCA) has been introduced to enhance the reliability of Large Language Models (LLMs) by enabling early abstention from generating potentially incorrect responses. This approach analyzes the internal diversity of LLM knowledge through causal inference, allowing models to assess the reliability of their knowledge before generating answers.