Unforgotten Safety: Preserving Safety Alignment of Large Language Models with Continual Learning

A recent study evaluated the effectiveness of deep learning models and large language models (LLMs) for vulnerability detection, focusing on models like ReVeal and LineVul across four datasets: Juliet, Devign, BigVul, and ICVul. The research highlights the gap between benchmark performance and real-world applicability, emphasizing the need for systematic evaluation in practical scenarios.

The newly introduced framework, Text2Graph, integrates lightweight large language models (LLMs) with graph neural networks (GNNs) to enhance text classification, particularly in scenarios with limited labels. This open-source Python package allows for flexible component swapping, including feature extractors and sampling strategies, and has been benchmarked across five datasets for zero-shot classification tasks.

An open, machine-readable dataset of Greek government decisions has been introduced, sourced from the national transparency platform Diavgeia, comprising 1 million decisions with high-quality raw text extracted from PDFs. This dataset is released with a reproducible extraction pipeline and includes qualitative analyses to explore boilerplate patterns and a retrieval-augmented generation (RAG) task to evaluate information access and reasoning over governmental documents.

A novel approach to Named Entity Recognition (NER) for Portuguese has been introduced, utilizing a three-step ensemble pipeline of locally run Large Language Models (LLMs). This method demonstrates superior performance over individual models across multiple datasets, particularly in zero-shot scenarios, where minimal annotated data is available.

LMSpell has been introduced as a neural spell checking toolkit specifically designed for low-resource languages (LRLs), showcasing the effectiveness of large language models (LLMs) in improving spell correction. This toolkit includes an evaluation function that addresses the hallucination issues often associated with LLMs, marking a significant advancement in the field of natural language processing for underrepresented languages.

JITServe has been introduced as the first SLO-aware serving system for Large Language Models (LLMs), addressing the challenges posed by diverse workloads and unpredictable request information. This system aims to optimize service goodput by effectively scheduling requests to meet specific service-level objectives (SLOs) across various applications, including chatbots and multi-agent systems.

This chapter discusses the use of Large Language Models (LLMs) in the legal field, highlighting their ability to enhance traditional legal tasks such as interpreting statutes, contracts, and case law. It also addresses the challenges posed by these technologies, including algorithmic monoculture and compliance with regulations like the EU's AI Act and U.S. initiatives.

A new study introduces STA-Attention, a framework utilizing Top-K Sparse Autoencoders to analyze the Key-Value (KV) cache in long-context Large Language Models (LLMs). This research reveals a Key-Value Asymmetry, where Key vectors act as sparse routers while Value vectors contain dense content, leading to a proposed Dual-Budget Strategy for optimizing semantic component retention.