Germans Savcisens (Savčišens)
@savcisens.com
Postdoc @nunetsi.bsky.social (Northeastern Uni) 🎓 Computational Social Science 👾 ✨ work on stability of belief in LLMs & Human-AI Collaboration 🌿 he/him 🇱🇻 🇺🇦 | www.savcisens.com
Pinned
@nunetsi.bsky.social had a great week at @ic2s2.bsky.social! Looking forward to the next #IC2S2 in Vermont 🔬⛰️
Truthfulness isn’t always binary. Sometimes it’s… neither 🤔 Our Trilemma of Truth paper is headed to the @neuripsconf.bsky.social Mechanistic Interpretability workshop 🚀 Let’s connect in San Diego! 🌴
Preprint: arxiv.org/abs/2506.23921
Code and data: github.com/carlomarxdk/...
Preprint: arxiv.org/abs/2506.23921
Code and data: github.com/carlomarxdk/...
September 23, 2025 at 2:12 PM
Truthfulness isn’t always binary. Sometimes it’s… neither 🤔 Our Trilemma of Truth paper is headed to the @neuripsconf.bsky.social Mechanistic Interpretability workshop 🚀 Let’s connect in San Diego! 🌴
Preprint: arxiv.org/abs/2506.23921
Code and data: github.com/carlomarxdk/...
Preprint: arxiv.org/abs/2506.23921
Code and data: github.com/carlomarxdk/...
Had the pleasure of presenting our work on Three-valued veracity probes for LLMs at #NEMI Workshop! Mechanistic Interpretability has such a great and welcoming community.
If we crossed paths - let’s connect! 🚀
Poster: zenodo.org/records/1690...
Preprint: arxiv.org/abs/2506.23921
If we crossed paths - let’s connect! 🚀
Poster: zenodo.org/records/1690...
Preprint: arxiv.org/abs/2506.23921
August 23, 2025 at 12:47 PM
Had the pleasure of presenting our work on Three-valued veracity probes for LLMs at #NEMI Workshop! Mechanistic Interpretability has such a great and welcoming community.
If we crossed paths - let’s connect! 🚀
Poster: zenodo.org/records/1690...
Preprint: arxiv.org/abs/2506.23921
If we crossed paths - let’s connect! 🚀
Poster: zenodo.org/records/1690...
Preprint: arxiv.org/abs/2506.23921
@nunetsi.bsky.social had a great week at @ic2s2.bsky.social! Looking forward to the next #IC2S2 in Vermont 🔬⛰️
July 25, 2025 at 10:12 PM
@nunetsi.bsky.social had a great week at @ic2s2.bsky.social! Looking forward to the next #IC2S2 in Vermont 🔬⛰️
Reposted by Germans Savcisens (Savčišens)
All the keynote recordings are available now, enjoy! www.youtube.com/playlist?lis...
IC2S2'25 Norrköping - YouTube
This playlist contains all keynotes from IC2S2'25 in Norrköping, Sweden.
www.youtube.com
July 25, 2025 at 7:13 PM
Presented our work on veracity-tracking in LLMs at #IC2S2 today!
Now looking forward to the next few days of great talks and conversations ✨️🎓
Now looking forward to the next few days of great talks and conversations ✨️🎓
July 22, 2025 at 3:44 PM
Presented our work on veracity-tracking in LLMs at #IC2S2 today!
Now looking forward to the next few days of great talks and conversations ✨️🎓
Now looking forward to the next few days of great talks and conversations ✨️🎓
Little wins: our "Trilemma of Truth" dataset just hit 150 downloads. It contains true, false, and neither-valued statements (inspired by the three-valued logic) used to stress-test LLMs for fact-checking, veracity tracking, and uncertainty handling.
Dataset📚: huggingface.co/datasets/car...
Dataset📚: huggingface.co/datasets/car...
carlomarxx/trilemma-of-truth · Datasets at Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
July 21, 2025 at 9:25 AM
Little wins: our "Trilemma of Truth" dataset just hit 150 downloads. It contains true, false, and neither-valued statements (inspired by the three-valued logic) used to stress-test LLMs for fact-checking, veracity tracking, and uncertainty handling.
Dataset📚: huggingface.co/datasets/car...
Dataset📚: huggingface.co/datasets/car...
Perfect weather, charming streets, and a poster so big it almost needed its own boarding pass 🧳✨
Excited to attend #IC2S2 in Norrköping 🇸🇪 Find me at the Poster Session on Tuesday: "Improving Probes that Track Veracity in Large Language Models" (Poster ID: 39) 🧪
Excited to attend #IC2S2 in Norrköping 🇸🇪 Find me at the Poster Session on Tuesday: "Improving Probes that Track Veracity in Large Language Models" (Poster ID: 39) 🧪
July 20, 2025 at 6:51 PM
Perfect weather, charming streets, and a poster so big it almost needed its own boarding pass 🧳✨
Excited to attend #IC2S2 in Norrköping 🇸🇪 Find me at the Poster Session on Tuesday: "Improving Probes that Track Veracity in Large Language Models" (Poster ID: 39) 🧪
Excited to attend #IC2S2 in Norrköping 🇸🇪 Find me at the Poster Session on Tuesday: "Improving Probes that Track Veracity in Large Language Models" (Poster ID: 39) 🧪
I’m presenting a poster on my latest project: “The Trilemma of Truth.”
Drop by to see how LLMs leverage three‑valued logic to model truth 🔢🤖
And hey, if you fancy grabbing a coffee ☕, DM me!
📄 Poster: zenodo.org/records/1605...
📖 Preprint: arxiv.org/abs/2506.23921
Drop by to see how LLMs leverage three‑valued logic to model truth 🔢🤖
And hey, if you fancy grabbing a coffee ☕, DM me!
📄 Poster: zenodo.org/records/1605...
📖 Preprint: arxiv.org/abs/2506.23921
July 18, 2025 at 1:46 AM
I’m presenting a poster on my latest project: “The Trilemma of Truth.”
Drop by to see how LLMs leverage three‑valued logic to model truth 🔢🤖
And hey, if you fancy grabbing a coffee ☕, DM me!
📄 Poster: zenodo.org/records/1605...
📖 Preprint: arxiv.org/abs/2506.23921
Drop by to see how LLMs leverage three‑valued logic to model truth 🔢🤖
And hey, if you fancy grabbing a coffee ☕, DM me!
📄 Poster: zenodo.org/records/1605...
📖 Preprint: arxiv.org/abs/2506.23921
Reposted by Germans Savcisens (Savčišens)
Germans Savcisens, Tina Eliassi-Rad: The Trilemma of Truth in Large Language Models https://arxiv.org/abs/2506.23921 https://arxiv.org/pdf/2506.23921 https://arxiv.org/html/2506.23921
July 1, 2025 at 6:31 AM
Germans Savcisens, Tina Eliassi-Rad: The Trilemma of Truth in Large Language Models https://arxiv.org/abs/2506.23921 https://arxiv.org/pdf/2506.23921 https://arxiv.org/html/2506.23921
🚨 New preprint!
Do LLMs really know what’s true?
In our paper, @eliassi.bsky.social and I introduce sAwMIL: a probing method that distinguishes between true, false, and neither—capturing what LLMs actually “retain.”
We evaluated 16 open models across 3 new datasets.
📄 arxiv.org/abs/2506.23921
Do LLMs really know what’s true?
In our paper, @eliassi.bsky.social and I introduce sAwMIL: a probing method that distinguishes between true, false, and neither—capturing what LLMs actually “retain.”
We evaluated 16 open models across 3 new datasets.
📄 arxiv.org/abs/2506.23921
The Trilemma of Truth in Large Language Models
We often attribute human characteristics to large language models (LLMs) and claim that they "know" certain things. LLMs have an internal probabilistic knowledge that represents information retained d...
arxiv.org
July 1, 2025 at 3:09 AM
🚨 New preprint!
Do LLMs really know what’s true?
In our paper, @eliassi.bsky.social and I introduce sAwMIL: a probing method that distinguishes between true, false, and neither—capturing what LLMs actually “retain.”
We evaluated 16 open models across 3 new datasets.
📄 arxiv.org/abs/2506.23921
Do LLMs really know what’s true?
In our paper, @eliassi.bsky.social and I introduce sAwMIL: a probing method that distinguishes between true, false, and neither—capturing what LLMs actually “retain.”
We evaluated 16 open models across 3 new datasets.
📄 arxiv.org/abs/2506.23921
That happens way too often to me 🥲
Everyday struggle 🧪
plentyofroom.beehiiv.com
plentyofroom.beehiiv.com
May 3, 2025 at 3:20 PM
That happens way too often to me 🥲
What's the coolest guide/source on "Complex Data Visualization"? I am looking for some inspiration to visualize graphs and high dimensional data.
April 14, 2025 at 7:10 PM
What's the coolest guide/source on "Complex Data Visualization"? I am looking for some inspiration to visualize graphs and high dimensional data.
Reposted by Germans Savcisens (Savčišens)
Great read from 2019 about abandoning the use of p-values in a dichotomous way and what we can do instead. More thinking, and less relying on significance tests to decide things for us!
www.nature.com/articles/d41...
www.nature.com/articles/d41...
Scientists rise up against statistical significance
Valentin Amrhein, Sander Greenland, Blake McShane and more than 800 signatories call for an end to hyped claims and the dismissal of possibly crucial effects.
www.nature.com
March 18, 2025 at 7:31 PM
Great read from 2019 about abandoning the use of p-values in a dichotomous way and what we can do instead. More thinking, and less relying on significance tests to decide things for us!
www.nature.com/articles/d41...
www.nature.com/articles/d41...
Reposted by Germans Savcisens (Savčišens)
Sometimes we need a reality check 😉 @serge.belongie.com
March 5, 2025 at 2:37 PM
Sometimes we need a reality check 😉 @serge.belongie.com
Amazing time with the folks from @tint-philosophy.bsky.social at the retreat on Predictability of Human Lives. Great people & discussions, and so much to reflect on—especially around integrative modeling and how neural networks can help us get there. Plus, a relaxing sauna to top it off!
February 8, 2025 at 10:11 AM
Amazing time with the folks from @tint-philosophy.bsky.social at the retreat on Predictability of Human Lives. Great people & discussions, and so much to reflect on—especially around integrative modeling and how neural networks can help us get there. Plus, a relaxing sauna to top it off!
Visiting @mpidr.bsky.social this week—super excited to see what’s happening in Demographic Studies (don’t miss my talk!).
Also, I’ll be in Berlin on Feb 2, Helsinki from Feb 3-6, and Copenhagen from Feb 10-12. Let me know if you’re around and up for a coffee 🧪🔬☕️
Also, I’ll be in Berlin on Feb 2, Helsinki from Feb 3-6, and Copenhagen from Feb 10-12. Let me know if you’re around and up for a coffee 🧪🔬☕️
January 27, 2025 at 2:48 PM
Visiting @mpidr.bsky.social this week—super excited to see what’s happening in Demographic Studies (don’t miss my talk!).
Also, I’ll be in Berlin on Feb 2, Helsinki from Feb 3-6, and Copenhagen from Feb 10-12. Let me know if you’re around and up for a coffee 🧪🔬☕️
Also, I’ll be in Berlin on Feb 2, Helsinki from Feb 3-6, and Copenhagen from Feb 10-12. Let me know if you’re around and up for a coffee 🧪🔬☕️
Reposted by Germans Savcisens (Savčišens)
For those interdisciplinary students/scholars who are having identity crisis, this is for you (from 2018):
"How to survive as an interdisciplinary being"
www.slideshare.net/slideshow/ho...
#NetSciX2025
"How to survive as an interdisciplinary being"
www.slideshare.net/slideshow/ho...
#NetSciX2025
How to survive as an interdisciplinary being
How to survive as an interdisciplinary being - Download as a PDF or view online for free
www.slideshare.net
January 16, 2025 at 6:42 PM
For those interdisciplinary students/scholars who are having identity crisis, this is for you (from 2018):
"How to survive as an interdisciplinary being"
www.slideshare.net/slideshow/ho...
#NetSciX2025
"How to survive as an interdisciplinary being"
www.slideshare.net/slideshow/ho...
#NetSciX2025
Reposted by Germans Savcisens (Savčišens)
New tool to estimate the level of participation in collective action expressed in natural language.
Applied to social media, it can produce large-scale and granular estimates of behavior change wrt collective action.
github.com/ariannap13/e...
@nerdsitu.bsky.social @itu.dk @carlsbergfondet.dk
Applied to social media, it can produce large-scale and granular estimates of behavior change wrt collective action.
github.com/ariannap13/e...
@nerdsitu.bsky.social @itu.dk @carlsbergfondet.dk
Excited to share the tool @lajello.bsky.social & I built to predict social media participation in collective action! It moves beyond keywords, tracking activism stages across topics. See it in action with climate activism on Reddit 🌱
Check it out: arxiv.org/abs/2501.07368
@nerdsitu.bsky.social
Check it out: arxiv.org/abs/2501.07368
@nerdsitu.bsky.social
January 15, 2025 at 2:58 PM
New tool to estimate the level of participation in collective action expressed in natural language.
Applied to social media, it can produce large-scale and granular estimates of behavior change wrt collective action.
github.com/ariannap13/e...
@nerdsitu.bsky.social @itu.dk @carlsbergfondet.dk
Applied to social media, it can produce large-scale and granular estimates of behavior change wrt collective action.
github.com/ariannap13/e...
@nerdsitu.bsky.social @itu.dk @carlsbergfondet.dk
I once asked ChatGPT how it thinks my life would look like in 20 years. And "Visionary Multidimensional Social Scientist" sounds like a great job title 😅 I guess it captured my love for the "His Dark Materials" trilogy.
January 4, 2025 at 2:54 PM
I once asked ChatGPT how it thinks my life would look like in 20 years. And "Visionary Multidimensional Social Scientist" sounds like a great job title 😅 I guess it captured my love for the "His Dark Materials" trilogy.
Reposted by Germans Savcisens (Savčišens)
📢 @savcisens.com discusses a recent study that shows that LLMs exhibit social identity biases similar to humans. www.nature.com/articles/s43...
🔓https://rdcu.be/d5owe
🔓https://rdcu.be/d5owe
Large language models act as if they are part of a group - Nature Computational Science
An extensive audit of large language models reveals that numerous models mirror the ‘us versus them’ thinking seen in human behavior. These social prejudices are likely captured from the biased conten...
www.nature.com
January 2, 2025 at 2:58 PM
📢 @savcisens.com discusses a recent study that shows that LLMs exhibit social identity biases similar to humans. www.nature.com/articles/s43...
🔓https://rdcu.be/d5owe
🔓https://rdcu.be/d5owe
Happy to write this News & Views piece on the recent audit showing LLMs picking up "us versus them" biases: www.nature.com/articles/s43... (Read-only version: rdcu.be/d5ovo)
Check out the amazing (original) paper here: www.nature.com/articles/s43...
Check out the amazing (original) paper here: www.nature.com/articles/s43...
Large language models act as if they are part of a group - Nature Computational Science
An extensive audit of large language models reveals that numerous models mirror the ‘us versus them’ thinking seen in human behavior. These social prejudices are likely captured from the biased conten...
www.nature.com
January 2, 2025 at 2:11 PM
Happy to write this News & Views piece on the recent audit showing LLMs picking up "us versus them" biases: www.nature.com/articles/s43... (Read-only version: rdcu.be/d5ovo)
Check out the amazing (original) paper here: www.nature.com/articles/s43...
Check out the amazing (original) paper here: www.nature.com/articles/s43...
Reposted by Germans Savcisens (Savčišens)
I don't like the way many CS papers are written, even the supposedly good ones, but these tips are very generically applicable and useful. Just ignore the bit about the acronyms...
For those who missed this post on the-network-that-is-not-to-be-named, I made public my "secrets" for writing a good CVPR paper (or any scientific paper). I've compiled these tips of many years. It's long but hopefully it helps people write better papers. perceiving-systems.blog/en/post/writ...
Writing a good scientific paper
perceiving-systems.blog
November 23, 2024 at 3:40 PM
I don't like the way many CS papers are written, even the supposedly good ones, but these tips are very generically applicable and useful. Just ignore the bit about the acronyms...
If you want to follow what happens with #ML, #AI, #LLM research in Denmark - here is a great Starter Pack 😋 #researchers
Yay! I have created a Danish Machine Learning peoples starter pack! Feel free to comment below if you want to add yourself or others ☺️.
I aim to include people working in ML in Denmark.
I aim to include people working in ML in Denmark.
November 20, 2024 at 8:48 PM
If you want to follow what happens with #ML, #AI, #LLM research in Denmark - here is a great Starter Pack 😋 #researchers
#YouTube has so far come up with the best application of GenAI (in recent months): summarization of videos and the ability to chat with them -- a great way to check the contents of 60+ min talks.
November 20, 2024 at 3:38 PM
#YouTube has so far come up with the best application of GenAI (in recent months): summarization of videos and the ability to chat with them -- a great way to check the contents of 60+ min talks.