Alham Fikri Aji

🎉Happy to share our recent collaborative effort on building a culturally diverse, multilingual visual QA dataset! CVQA consists of over 9,000 questions across 28 countries, covering 26 languages (with more to be added!) 🌐cvqa-benchmark.org 📜arxiv.org/pdf/2406.05967

🎉Excited to share new work with undergrads from @FASILKOM_UI Multilingual interpretability is often done at the neuron level; but they are polysemantic We show that using a self-autoencoder first yields more monosemantic, interpretable features More: arxiv.org/pdf/2507.11230

is there a superhero name that actually sounds like good name to give to a child?

Personal take: With AI handles more of our thinking, cognitive decline could become a concern We may soon see a rise in brain gyms or mental fitness businesses designed to keep our minds sharp

I'll be hiring a couple of Ph.D. students at CMU (via LTI or MLD) in the upcoming cycle! If you are interested in joining my group, please read the FAQ before reaching out to me via email :) docs.google.com/document/d/12V…

📢New dataset alert! A new long-form video benchmark assessing multimodal LLMs' understanding of Theory of Mind in realistic social interactions

🚨 Deadline Extended 🚨 The MELT Workshop Track submission deadline is now June 30, 2025 (AoE). We still welcome new & unpublished work: abstracts (≤2 p) | shorts (≤5 p) | longs (≤9 p) 📌 Non-archival 🔗 Submit here → openreview.net/group?id=colmw… #MELTWorkshop2025 #COLM2025

Supposedly 87% of Indonesian high school graduates have math skills equivalent to a 3rd grader This will probably get worse in the coming years, with more students relying on LLMs to do their homework without actually learning anything

Working on multilingual and culturally-aware NLP and want to attend @COLM_conf? Consider submitting to the MELT Workshop 2025! Submissions are non-archival

There are many interesting questions How people anthropomorphize LLMs, correlation w/ their individual traits, demographics, LLM understanding, perceived trust, etc2. just a random thought, but this whole AI-LLM interaction thing is fascinating

🤔Ever wonder why LLMs give inconsistent answers in different languages? In our paper, we identify two failure points in the multilingual factual recall process and propose fixes that guide LLMs to the "right path." This can boost performance by 35% in the weakest language! 📈

🧵 Multilingual safety training/eval is now standard practice, but a critical question remains: Is multilingual safety actually solved? Our new survey with @Cohere_Labs answers this and dives deep into: - Language gap in safety research - Future priority areas Thread 👇

Research isn’t always about chasing SOTA. It’s about understanding why things work (or don’t) and pushing the boundaries of our knowledge Negative results like this are still valuable and essential for the community Looking forward to the next banger!

Amidst the evaluation/reproducibility crisis for reasoning LLMs, it's great to see *concurrent independent work (with different models & benchmarks) aligns with our findings*! We reported the same fundamental trade-off: language forcing leads to ✅ compliance, ❌ accuracy!

Happy to share that we have 8 papers accepted at ACL! Grateful for the amazing team of students and collaborators. Hope to see you in Vienna soon!

Thanks for covering our work!

Multilingual vision is important! We have developed a multilingual VQA benchmark almost a year ago - CVQA arxiv.org/abs/2406.05967 in NeurIPS.

After publishing 4 EMNLP papers last year, 2 as first author, and now this. His citation has grown 10x in the past year. He's on track to surpass 1,000,000 citations in 5 years! An amazing future lies ahead for Farid.

Reasoning language models are primarily trained on English data, but do they generalize well to multilingual settings in various domains? We show that test-time scaling can improve their zero-shot crosslingual reasoning performance! 🔥