Can large language models (LLMs) recognize convincing arguments? Kind of. 😅

We propose evaluating LLMs' persuasiveness capabilities by measuring their ability to recognize whether an argument would convince a user with specific demographics.

📜 arxiv.org/abs/2404.00750

Orchestrated interactions between LLMs, humans, and tools show great promise.

Today, we introduce “Semantic Decoding' - a perspective that views these interactions as optimization and search in the space of semantic tokens (thoughts).

📄 arxiv.org/abs/2403.14562

🚨LLMs outdebate humans when given access to opponents' personal Information!🚨

In a pre-reg study (N=820), participants who debated ChatGPT had 81.7% (p<0.01) higher odds of agreeing with their opponents after the debate (compared to a human baseline)

📜arxiv.org/abs/2403.14380

Cool work from our postdoc Debjit Paul at the intersection of NLP and causality! One thing that have me so excited about LLMs is that they allow us to study counterfactuals via simple interventions—rare in the real world, but possible with LLMs,which we can just reset as we wish

Rich Zemel and I are looking for a postdoc fellow to work on Societal Aspects of AI! Read more details and apply here: csslab.cs.toronto.edu/blog/2024/03/0…

Happy to share the good news that as of today I’m officially tenured — aka promoted from ass. prof. to ass. prof. 🤔
ethrat.ch/en/appointment…

Many thanks to all those who have been supporting me along this exciting path: students, colleagues, and most of all my lovely family. 💋

Wiki Workshop 2024 has its first keynote speaker: the amazing Brent Hecht Brent Hecht, Director of Applied Research at Microsoft. Join us on 20 June 2024 for this free and virtual event — your go-to place for discussing wiki-related research!

We are thrilled to announce that Brent Hecht will be a keynote speaker of #WikiWorkshop2024 ‼️

🔗 wikiworkshop.org/2024/#contribu…

New PNAS paper with Homa Hosseinmardi, Amir Ghasemian, Miguel Rivera-Lanas, Manoel Horta Ribeiro, and Duncan Watts.
Btw, Manoel is on the job market. If you’re hiring assistant professors in CSS, you should try to get him — what an opportunity!

I'm happy to announce that 'low-memory beam search' has been merged into Hugging Face transformers! 🎉

Don't hesitate to try it out if you are struggling with the memory issue of beam search! (Just set 'low_memery=True')

The low-memory mode reduces the memory usage to the same…

Excited about Saibo Geng et al.’s work on sketch-guided constrained decoding—get the power of the strongest available LLMs while being able to control their output format as flexibly as you can do with the openest available LLMs. I think we’ll be using this a lot in our work…