People more influenced by advice from AI than fellow humans, UBC study finds

Theory of Mind can make LLMs better at dialogue: more strategic, goal-oriented, enabling long-horizon adaptation!

In our new paper, we introduce ToMA, a dialogue lookahead training framework that enables the LLMs to generate mental states that are maximally useful for achieving dialogue goals.🧵👇

@veredshwartz.bsky.social author of ‘Lost in Automatic Translation’, explores the opportunities and pitfalls of language technologies, from automatic translation tools to large language models (LLMs) like ChatGPT. 🤖

Read the full interview 🔗 https://cup.org/4n8ilt4

Outlook spell checker is suggesting that I write "I attached me slides" instead of "I attached my slides". #AI_fail

I just published Are we expecting too much of LLMs? An analysis of prompts in NLP papers medium.com/p/are-we-exp...

"Lost in Automatic Translation" is finally out!

Get it here: lostinautomatictranslation.com

*SEM 2025 final call for ARR COMMITMENTS! Final chance to commit to *SEM 2025, co-located with EMNLP in Suzhou!! ARR COMMITMENT deadline is August 22, 2025. Give your strong-but-overlooked EMNLP paper the second chance it deserves! 😉 starsem2025.github.io/cfp

#NLP #NLProc

Questions if Google Translate and ChatGPT will help us master foreign languages and better understand one another, or make language learning obsolete.

Lost in Automatic Translation by @veredshwartz.bsky.social, Coming Soon

cup.org/45lmrI2

#LangSky #Linguistics 🐦🐦 #AI #LLMs #ArtificalIntelligence

I'm not sure yet if I'll make it to #IJCAI2025 because of the planned Air Canada strike, but if you're attending, you can order my book with the IJCAI discount (30% off and free shipping) through mid-September:
www.cambridge.org/us/universit...

Congratulations to Shruthi Chockkalingam, Seyed Hossein Alavi, Raymond Ng & @veredshwartz.bsky.social for receiving an Outstanding Paper Award at the Social Influence in Conversations (SICon) Workshop at the 2025 Association for Computational Linguistics Conference! #ACL2025

I just finished listening to the audiobook of "Empire of AI" by @karenhao.bsky.social (highly recommend!). The copyright statement in the credits explicitly mentions it's not allowed to use the book for training AI. I love it and will ask my producer to do the same 🙂

“My main concern at the moment is the widespread adoption, reliance and trust in LLMs, despite a very serious technical limitation — the hallucination problem.”

@isaacphannay.bsky.social interviews "Lost in Automatic Translation" author @veredshwartz.bsky.social.

generic description that doesn't mention specific books or podcasts they've narrated, hundreds of audio samples. I hate living in a world in which I'm constantly wondering if things are real or AI-generated. 4/4

The interesting thing is that the narration is actually good enough, at least for my short sample. I suspected because of other things: AI generated profile pictures that don't yield results in Google lens, no website or contact details, 3/4

Can't say I'm surprised that people are trying to profit from selling AI narration. But I'm a bit taken aback about the fact that they pose as human narrators. This is a scam. 2/4

I'm using ACX to produce an audiobook (my publisher only does print and ebooks). I posted a text sample for narrators to audition. 80% of the auditions I got were from fake, AI narrators. 1/4

I just published "This book will not be used to train LLMs":
lostinautomatictranslation.com/blog/this-book-will-not-be-used-to-train-llms-1

Unexpected events grab human attention & push AI models beyond their training data. 🤔

How well do VLMs reason about these critical, novel scenarios?

🦢 Introducing Black Swan @ #CVPR2025 (Poster: Sun 10:30am #304)

Co-lead: @sahithya.bsky.social
w/ @veredshwartz.bsky.social

🔗 blackswan.cs.ubc.ca

Congrats, Sarah! 🎉

Starting in a few minutes - we have an exciting line up of talks, posters, and a panel! Join us at room 104E. 🌍 #CVPR2025

🗓️ Save the date! It's official: The VLMs4All Workshop at #CVPR2025 will be held on June 12th!

Get ready for a full day of speakers, posters, and a panel discussion on making VLMs more geo-diverse and culturally aware 🌐

Check out the schedule below!

I recently got an email that said "I enjoyed reading your paper [...]" mentioning the first paper on the publication page on my website, which is not yet public.

The fact that ChatGPT learned to respond with "I take full responsibility" when being accused, but not the implications of actually taking responsibility, is an excellent example for learning form vs meaning (or at least semantics vs pragmatics).

Simple example for how errors can creep into our papers through LLM use: I had a statistic for 2020. I googled about the same statistic for 2023. AI overview tells me the statistic for 2023 and provides a link to support the claim. The link is to a 2023 article citing the 2020 statistic.