August 5, 2025

Highlights from ACL 2025

ACL 2025 in Vienna was one of the most exciting conferences I have attended in recent years. The community continues to grapple with fundamental questions about what large language models truly understand and how we can make them more reliable. Here are the themes and papers that stood out to me.

A major theme was the faithfulness of model reasoning. Several papers proposed methods to verify whether chain-of-thought explanations actually reflect the model's decision process. This connects directly to our own work on semantic entailment-based faithfulness, which we will present at ACL 2026.

I was encouraged to see strong progress on multilingual NLP, particularly for languages that have historically been underrepresented. New datasets for African and Southeast Asian languages, combined with better cross-lingual transfer methods, are making NLP more globally inclusive. Our own MultiLingua toolkit was cited in several of these papers.