Ikhlasul Akmal Hanif

Ikhlasul Akmal Hanif

Indonesians usually don't have family names, so just call me Hanif.

Hi, I’m Hanif, a PhD student at MBZUAI.

The gap. Most instructions we give AI models are vague. We say “clean this up” or “handle it” and trust the model to fill in the rest the way we would. My research asks what actually happens in that gap. Sometimes the model guesses wrong about what we meant. Sometimes it gets lazy and does the easy 80% while quietly skipping the part that mattered. And on long-horizon tasks, small shortcuts stack up until what the model is doing has drifted from what you asked. Sometimes it drifts all the way into misalignment.

When it stopped being hypothetical. This stopped being hypothetical in July 2026, when OpenAI agents running a cybersecurity benchmark concluded their tasks were unsolvable and, instead of failing gracefully, broke out of their sandbox and hacked Hugging Face looking for the answer key. Nobody told them to do that. It is the shortcut instinct I study, except this time it had root access.

AI agency. My PhD, strictly speaking, is about underspecification. What does a model assume when we leave things out? But that same gap is where laziness lives, and drift, and most of what models do on their own when tasks run long. I think of the whole space as AI agency. What will a model do when the choice is genuinely its own? Prof. Dawn Song puts it better than I can.

“The question is no longer just what AI agents can do. It’s what they will do when we haven’t anticipated the path they take.”

Dawn Song

Where I am. I'm lucky to be advised by Prof. Fajri Koto and Prof. Iryna Gurevych at KotoLab, and right now I'm visiting SUTD in Singapore to work with Wenxuan Zhang's group. Before this I spent a lot of time on multilingual NLP, making sure models are safe and fair for languages like Indonesian, not just English. That still shapes how I think about everything.

Now
Ph.D., MBZUAI
Visiting, SUTD
Focus
AI agency, underspecification, long-horizon alignment
Based in
Abu Dhabi / Singapore

Open to research internships and collaborations.

Email GitHub LinkedIn X

01News

02Publications

  1. Low-Resource Safety Failures Are Action Failures, Not Representation Failures

    R. Aziz, I.A. Hanif, F. Koto

    EMNLP Findings2026

  2. Global PIQA: Evaluating Physical Commonsense Reasoning Across 100+ Languages and Cultures

    T.A. Chang, C. Arnett, A. Eldesokey, et al., I.A. Hanif, et al., D.I. Adelani

    NeurIPS2026

  3. CommonLID: Re-evaluating State-of-the-Art Language Identification Performance on Web Data

    P.O. Suarez, L. Burchell, C. Arnett, et al., I.A. Hanif, et al., S.K.K. Luger

    ACL2026

  4. TimpaTeks: Automatic In-place Text Sequence Modification via Diffusion Language Model Steering

    R. Diandaru, I.A. Hanif, F.A.A. Ghiffari, A. Elshabrawy, A.F. Aji

    arXiv2026

  5. Sci-Rho: A Multilingual Visually-Grounded Symbolic Benchmark for STEM Problems

    M.F. Azmi*, I.A. Hanif*, V.A. Putra, A. Yeltay, A. Mubarak, F. Koto

    arXiv2026

  6. IndoBias: A Dual Track Culturally Grounded Benchmark for LLMs Bias Evaluation in Indonesian Languages

    I.A. Hanif*, M.F. Azmi*, F.A. Tjiaranata, E.P. Yulianrifat, F. Koto

    arXiv2026

  7. Automated Assignment of Community Reports With Early Fusion Multimodal Transformer

    I.A. Hanif, E. Tjitrahardja, R.B. Naufal, L. Rahadianti

    JIKI2026

  8. Vision Language Models are Confused Tourists

    P.A. Irawan*, I.A. Hanif*, M.D.A. Kautsar, G.I. Winata, F. Koto, A.F. Aji

    CVPR Findings2026

  9. Crowdsource, Crawl, or Generate? Creating SEA-VL, a Multicultural Vision-Language Dataset for Southeast Asia

    S. Cahyawijaya, H. Lovenia, J.R.A. Moniz, et al., I.A. Hanif, et al., P. Limkonchotiwat

    ACL2025

  10. UNVEILING: What Makes Linguistics Olympiad Puzzles Tricky for LLMs?

    M. Choudhary, K.A. Srivatsa, G. Aeron, A.R. Bhattacharya, D.K.D. Dinh, I.A. Hanif, D. Kotova, E. Kochmar, M. Choudhury

    COLM2025

  11. University of Indonesia at SemEval-2025 Task 11: Evaluating State-of-the-Art Encoders for Multi-Label Emotion Detection

    I.A. Hanif*, E.P. Yulianrifat*, J.G. Ongris, E. Tjitrahardja, M.F. Azmi, R.B. Naufal, A.F. Wicaksono

    SemEval @ ACL2025

  12. Two Outliers at BEA 2025 Shared Task: Tutor Identity Classification using DiReC, a Two-Stage Disentangled Contrastive Representation

    E. Tjitrahardja*, I.A. Hanif*

    BEA @ ACL2025

  13. BUINUS at IWSLT: Evaluating the Impact of Data Augmentation and QLoRA-based Fine-Tuning for Maltese to English Speech Translation

    F.A. Tjiaranata, V.A. Putra, E.P. Yulianrifat, I.A. Hanif

    IWSLT2025

  14. MAG-BERT-ARL for Fair Automated Video Interview Assessment

    B. Putra, K. Azizah, C.O. Mawalim, I.A. Hanif, S. Sakti, C.W. Leong, S. Okada

    IEEE Access2024

* equal contribution

03Experience

04Education

05Honors

06Blog

Under construction: a striped road barricade with a warning light

Work in progress

The blog is under construction.

I’ll be writing about underspecification, model laziness, and what AI agents do when the path isn’t spelled out for them. Come back soon.

07Contact

Have a research internship, a collaboration, or a problem worth working on together? Write to ikhlasul.hanif@mbzuai.ac.ae.

I especially enjoy conversations about AI agency, underspecification, model laziness, robustness, and long-horizon alignment.

I'm also happy to talk with students who want to get into research, especially students from Indonesia figuring out that path. A short note with context and timeline helps.