Ph.D. student at @UMichCSE, adviced by @radamihalcea. Learning and working on AI. Wants to get lost in the mountains with a good book and a cup of coffee.
Ann Arbor, MI
Joined February 2014
- Tweets20
- Following58
- Followers27
- Likes13
Can we fix it? Fine-tuning on corrected examples lifts resistance to false norms to 85–93%, even for countries never seen in training.
Inside the model, the false claim's influence drops sharply.
Everything is open:
📄 Paper: arxiv.org/abs/2610.27202
💻 Code: github.com/MichiganNLP/cultu…
🤗 Data & models: huggingface.co/collections/s…
So is the knowledge lost? No.
Looking inside the models, the country's norm is still there. Blocking the false claim inside the model brings the correct answer back in most cases.
The false claim overrides what the model knows rather than erasing it.
If a prompt gets a cultural norm wrong, will an LLM correct it or go along with it?
Usually it goes along, even when it knows better.
New paper with @FurkanDanismann, @AnganaBorah2, @honglaklee, Dehan Kong, @lajanugen, Moontae Lee, @ZhijingJin, and @radamihalcea
🤖 Made with AI
We built CultureLens: 9,261 scenarios across 75 countries, from the Cultural Atlas.
Each scenario can include the country, a value, the true norm, or a contradictory one, so we can test which evidence a model actually uses.
Models: Llama-3.1-8B, Gemma-3-12B, Qwen-3.5-27B.
We built CultureLens: 9,261 scenarios across 75 countries, from the Cultural Atlas.
Each scenario can include the country, a value, the true norm, or a contradictory one, so we can test which evidence a model actually uses.
Models: Llama-3.1-8B, Gemma-3-12B, Qwen-3.5-27B.
Samee Arif retweeted
🚨 Excited to share that our paper "One Example Is Enough to Pass Fairness Benchmarks: Rethinking Fairness Evaluation for Aligned LLMs" has been accepted to EMNLP 2026 (Main)! 🎉
Samee Arif retweeted
🌎 A dataset says it covers Spanish. But whose Spanish?
Spanish is spoken across Spain and much of Latin America, from Mexico to Argentina and beyond. Yet an NLP dataset labeled simply “Spanish” may contain data representing only a small subset of those countries.
Samee Arif retweeted
Happy to share that this work has been accepted to EMNLP 2026 Main! 🤗
Excited to share our new preprint:
Whose Norms? Disentangling Cultural and Personal Alignment in Large Language Models
We ask: when cultural norms and personal preferences conflict, whose norms do LLMs follow? 🌍⚖️
Check out our work here:
📄 Paper: arxiv.org/abs/2606.07877
🤗 Dataset: huggingface.co/datasets/Anga…
💻 Code: github.com/MichiganNLP/pact_…
🌐 Website: lit.eecs.umich.edu/pact_cult… (1/n)
Samee Arif retweeted
🌍⚡ New paper: The Language–Energy Divide
Does the language you ask in change how much energy an LLM burns to answer? 👉 Yes — and the gap is enormous.
We measured LLM inference energy across 122 languages and found a systemic divide:
⚡ Up to 8.3× more energy per token across languages — same model, same task
🔋 Up to 179× more total energy per request set: English (17.6 kJ) vs. Pashto (3,147 kJ)
📉 A double penalty — the languages that cost the most to serve are also the ones answered least accurately
Excited to share that our paper has been accepted to EMNLP!
TL;DR: We introduce KIDBench to evaluate whether LLM responses to children are age-appropriate, emotionally supportive, culturally aligned, and robust across multi-turn conversations.
Paper: arxiv.org/abs/2605.25510
A small reflection on how the universe, through its patterns, gravity, stardust, light, uncertainty, time, and laws, can quietly teach us something about ourselves.
Link: dumplings.sameearif.com/post…
Part II of my CPU/GPU blog.
This one looks inside a core, how tiny switches become a system that can fetch, understand, and run instructions.
Link: dumplings.sameearif.com/post…
I just published a guide to graduate applications, built from advice and experiences shared by students across different application cycles.
Thank you to everyone who shared their insights and made this possible!
Link: dumplings.sameearif.com/post…
Loss is not only something we feel. It is something the mind slowly learns to live around.
A small essay on memory, attachment, grief, absence, and the strange science of being human.
Blog: dumplings.sameearif.com/post…
Excited to share that our paper has been accepted to COLM 2026!
LLM safety isn’t only about the final prompt. Path matters.
We introduce ICD, a trajectory-based jailbreak strategy showing how small continuations can progressively weaken safety.
Paper: arxiv.org/abs/2604.25921
I’ve started a small blog, Dumplings.
The first post explores Planck’s constant visually, and the second reflects on my first year of Ph.D.
Read About Dumplings to see what the blog is meant to be. I’m planning a visual Calculus series next.
dumplings.sameearif.com/