Do LLMs have personalities?

This paper introduces LMLPA, a framework for assessing linguistic personalities of LLMs using adapted Big Five inventories + AI raters, validated for reliability and consistency.

📄 Read in Computational Linguistics: direct.mit.edu/coli/article...

Evaluating NLG remains challenging — even with LLMs.

This survey reviews LLM-based evaluation methods: prompting, fine-tuning, and human–LLM collaboration, and outlines key open problems.

📄 Read in Computational Linguistics: direct.mit.edu/coli/article...

LLMs can generate fluent text — but can they twist your tongue?

TwisterLister generates phoneme-aware tongue twisters, with a 17k-example dataset and phonologically constrained decoding.

📄 Read the paper in Computational Linguistics: direct.mit.edu/coli/article...

Can LLMs learn possible human languages but struggle with impossible ones?

Kallini et al. claimed evidence, but Hunter argues their key experiment conflates factors — the core question remains open.

📄 Read the article in Computational Linguistics: direct.mit.edu/coli/article...

Even state-of-the-art MT systems struggle with ambiguous words — especially rare senses.

DiBiMT is a new fully human-curated benchmark across 8 language pairs to study disambiguation bias in MT.

📄 Read the paper in Computational Linguistics: direct.mit.edu/coli/article...

🚨 As LLMs generate more of the text we see daily, detecting AI-generated text is more critical than ever.

New survey from the University of Macau explores cutting-edge detection methods, key challenges, and future directions to ensure responsible AI use.

Read: direct.mit.edu/coli/article...

🔍 Neural semantic parsers are great – but often just parrot surface text.
Researchers from the University of Groningen propose a taxonomical parser that builds deeper, hierarchical meaning representations.

Read: direct.mit.edu/coli/article...

🚀 New from Queen Mary University of London: A unified framework for evaluating synthetic data generation from user-generated text! 📱💬

It tackles style, meaning, and privacy – paving the way for safe, shareable, high-quality synthetic language data.

Read: direct.mit.edu/coli/article...

How do humans understand novel sentence meanings? A new dataset (STS3k) from UniMelb & Hitotsubashi researchers tests compositionality in language models. Findings: hybrids of syntax + vectors better match human judgments than top transformers.🧠📚

Read: direct.mit.edu/coli/article...

Researchers from the University of Edinburgh, the University of Zurich, Supertext, and Microsoft introduce ACES, a challenging benchmark spanning 146 language pairs, in their 'Machine Translation Meta Evaluation through Translation Accuracy Challenge Sets'.

Read: direct.mit.edu/coli/article...

Researchers from Georgetown University, LMU Munich, Abo Akademi, and Indiana University present “eRST: A Signaled Graph Theory of Discourse Relations and Organization”.

Read: direct.mit.edu/coli/article...