Researchers identified a cooperative-override circuit in large language models, suppressing Nash play. This finding suggests a word-triggered circuit is responsible for cooperation, rather than a lack of competence. The circuit can be measured, bounded, and controlled. This has implications for understanding and improving the behavior of large language models.

A research paper proposes G-RELIC, a method for relexicalizing clinical documents using a graph-based approach. The method combines LLMs with graphs to optimize entity correspondence and preserve temporal consistency, outperforming state-of-the-art baselines in relational integrity and temporal coherence.

Researchers propose a framework for improving Named Entity Recognition (NER) annotations in low-resource languages, leveraging automated techniques and Large Language Models (LLMs). This work may inform the development of more accurate and robust AI agents, particularly in languages with limited training data.

Researchers proposed POSPAN, a framework for position-constrained span masking in language model pre-training, enhancing span-level masking methods by incorporating position constraint distributions. This improvement was validated through experiments on NLU benchmarks, outperforming existing methods.

Researchers explore the possibility of using graph-based representations to improve the interpretability of Large Language Model (LLM) Natural Language Inference (NLI) systems. They introduce a pipeline that decomposes input text into atomic propositions and represents them as graphs, achieving 89.7% accuracy on the SNLI dataset. The study highlights the trade-off between accuracy and interpretability, with the graph-based approach trailing behind its text-based counterpart by 1.9 points. The authors also demonstrate that combining graph and text modalities can achieve higher accuracy.