Entropy-Based Measurement of Value Drift and Alignment Work in Large Language Models

A new paper on arXiv presents datasets aimed at improving the understanding of idiomatic and figurative language in Natural Language Processing (NLP). These datasets are designed to assist large language models (LLMs) in better interpreting informal language, which has become increasingly prevalent in social media and everyday communication.

A new method for robust multimodal sentiment analysis of image-text pairs has been proposed, addressing challenges related to low-quality and missing modalities. The Distribution-based feature Recovery and Fusion (DRF) technique utilizes a feature queue for each modality to approximate feature distributions, enhancing sentiment prediction accuracy in real-world applications.

The recent introduction of ZIP-RC, an adaptive inference method, aims to optimize test-time compute for large language models (LLMs) by enabling zero-overhead joint reward-cost prediction. This innovation addresses the limitations of existing test-time scaling methods, which often lead to increased costs and latency due to fixed sampling budgets and a lack of confidence signals.

A new framework has been introduced for identifying attributions of causality in political text, utilizing a lightweight causal language model to generate structured data sets of causal claims. This approach aims to enhance the systematic analysis of explanations in political science, an area that has been historically fragmented and underdeveloped.

A recent study introduces a group fairness lens for evaluating large language models (LLMs), proposing a novel hierarchical schema to assess bias and fairness. The research presents the GFAIR dataset and introduces GF-THINK, a method aimed at mitigating biases in LLMs, highlighting the critical need for broader evaluations of these models beyond traditional metrics.

The Culture Affordance Atlas has been introduced as a function-centric framework aimed at addressing cultural biases in mainstream Vision-Language datasets, which often favor higher-income, Western contexts. This initiative involves a re-annotation of the Dollar Street dataset, categorizing 288 objects based on 46 functions to enhance model generalizability across diverse cultural and economic backgrounds.

Recent research has introduced a novel approach to tokenizer adaptation for pre-trained language models, focusing on vocabulary extension and pruning. The method, termed continued BPE training, enhances tokenization efficiency by continuing the BPE merge learning process on new data, while leaf-based vocabulary pruning removes redundant tokens without compromising model quality.

A novel intrinsic weight-based fingerprinting scheme named SELF has been proposed to enhance the protection of Intellectual Property (IP) in Large Language Models (LLMs). This approach utilizes singular value and eigenvalue decomposition of LLM attention weights to create unique and transformation-invariant fingerprints, addressing vulnerabilities in existing methods that are susceptible to false claims and weight manipulations.