Karolina Stańczak
banner
karstanczak.bsky.social
Karolina Stańczak
@karstanczak.bsky.social
#NLP Postdoc at Mila - Quebec AI Institute and McGill University | Former PhD @ University of Copenhagen (CopeNLU)

🌐 karstanczak.github.io
Reposted by Karolina Stańczak
6/ 🤝 Thanks to our steering committees and co-organizers for their hard work in making the VLMs4All Workshop possible!

@meharbhatia.bsky.social @rabiul.bsky.social @spandanagella.bsky.social @sivareddyg.bsky.social @svansteenkiste.bsky.social @karstanczak.bsky.social
March 14, 2025 at 3:55 PM
5/ Read our full paper here: arxiv.org/abs/2503.00069
Let’s discuss! How should AI align with society? 🤝💡
Societal Alignment Frameworks Can Improve LLM Alignment
Recent progress in large language models (LLMs) has focused on producing responses that meet human expectations and align with shared values - a process coined alignment. However, aligning LLMs remain...
arxiv.org
March 4, 2025 at 4:08 PM
4/ We also discuss the role of participatory alignment, where diverse stakeholders help shape AI behavior rather than deferring solely to designers.
March 4, 2025 at 4:08 PM
3/ Instead of perfecting rigid alignment objectives, we explore how LLMs can navigate uncertainty—a feature, not a flaw!
March 4, 2025 at 4:08 PM
2/ We propose leveraging societal alignment frameworks to guide LLM alignment:
🔹 Social alignment: Modeling norms, values & cultural competence
🔹 Economic alignment: Fair reward mechanisms & collective decision-making
🔹 Contractual alignment: Legal principles for LLMs
March 4, 2025 at 4:08 PM
1/ LLM alignment remains a challenge because human values are complex, dynamic, and often conflict with narrow optimization goals.

Existing methods like RLHF struggle with misspecified objectives.
March 4, 2025 at 4:08 PM