Sara Candussio

PhD student @UniTrieste | Reasoning and interpretability in LLMs

[Sara Candussio @SaraCandussio]

Does entropy actually track what's important in a chain of thought? Our paper (accepted to the BlackBoxNLP repr. track) says mostly no, and we found the spurious cause behind why it looked like it did 🧵

[jasmine @jasminecoded_]

I hate sf halloween, what do you mean you’re going as survivorship bias

[Aflah 🍉🕊️ @Aflah02101]

emnlp’26

[BlackboxNLP @BlackboxNLP]

BlackboxNLP 2026 is starting now! 🎉 Opening remarks are underway in Room P3, with keynotes from Sheridan Feucht and Ivan Titov up next. Come join us!

[Sara Candussio @SaraCandussio]

See you today at @BlackboxNLP, 16:00-17:00!

[Sara Candussio @SaraCandussio]

Does entropy actually track what's important in a chain of thought? Our paper (accepted to the BlackBoxNLP repr. track) says mostly no, and we found the spurious cause behind why it looked like it did 🧵

[Sean Yang @aicheye]

these are the high quality niche bangers i scroll for. i built this timeline brick by brick

[Graham Neubig @gneubig]

I'm glad someone appreciated my google slides default template + memes + emojis presentation style as an anti-culture fashion statement.

[Sara Candussio @SaraCandussio]

Thanks to my co-author @daniel_sc4 and our supervisors Luca Bortolussi, Elisabetta Fersini, Malvina Nissim, with a special contribution from @gsarti_ 🥰

[Sara Candussio @SaraCandussio]

Does entropy actually track what's important in a chain of thought? Our paper (accepted to the BlackBoxNLP repr. track) says mostly no, and we found the spurious cause behind why it looked like it did 🧵

[Sara Candussio @SaraCandussio]

TL;DR: What looked like an "entropy effect" on math tasks was really a numeric-token effect hiding behind it.

📄 paper: https://t.co/6sWhFYq9tt
💻 code: https://t.co/BcK7uU2WhO

Huge thanks to the reviewers at BlackBoxNLP Reproducibility track! 🙏

[Sara Candussio @SaraCandussio]

Heading to EMNLP to present our work with @daniel_sc4 and @gsarti_! See you at BlackBoxNLP on October 29, 16:00–17:00 🤠

[Sara Candussio @SaraCandussio]

Does entropy actually track what's important in a chain of thought? Our paper (accepted to the BlackBoxNLP repr. track) says mostly no, and we found the spurious cause behind why it looked like it did 🧵

[AILC_NLP @AILC_NLP]

All CLiC-it proceedings, from 2014 to 2025, are now available on ACL Anthology! 🎉
https://t.co/7NBI9g4CyZ
The 2026 proceedings are coming soon!
We've also added the 2014 and 2016 proceedings of EVALITA:
https://t.co/6fkt70eYnq
Thank you @aclanthology!

[Hilde Kuehne @HildeKuehne]

For everybody who plans to submit some random AI slop paper that you didn’t even bother to understand or even read … just don’t. Withdraw as long as you can. Nobody is going to give you a job or PhD position for this, but you will never get this out of your CV. And people will figure you out. We will ask you about what you wrote and this will be the end of the interview. You need to understand that you are closing the door to the career you might actually seek by doing this. Stop it as long as you can!

[Sara Candussio @SaraCandussio]

It was fun!

[CLiC-it Conference @CLiC_it_conf]

🗣️ Sara Candussio presenting:

“Don’t Say It!”: Constraints, Compliance, and Communication when Language Models Play Taboo

Authors: Sara Candussio, Francesca Padovani, Daniel Scalena and Malvina Nissim

[Sara Candussio @SaraCandussio]

@CLiC_it_conf 😈 and RogueAI, an interactive game and its turn-by-turn data analysis aimed at shedding light on how well humans are able to spot which of two LLM agents is deceptive and which is honest
https://t.co/RzSjB7e780

[Sara Candussio @SaraCandussio]

In Palermo for @CLiC_it_conf, a return to my roots (this was my very first conference! 🥹).
At this edition I'm presenting two works together with some collaborators, details below:

[Sara Candussio @SaraCandussio]

@CLiC_it_conf 🙈 "Don't Say It!", testing Language Models' ability to follow linguistic constraints, using Taboo as a testbed
https://t.co/wU0d6DjVuS