You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$2,400 left to reach the funding goal.
I’m Habeeb Abdulfatah, an undergraduate studying mechatronics engineering in Nigeria. Our paper, "Evaluating Mental Health LLM Responses to Localized African English", was accepted as a poster at the Global South AI Affinity Event at NeurIPS 2026 in Sydney, and I am the lead author.
While my conference registration costs is already covered, covering an international flight to Australia, lodging, a visa, and meals is impossible on my current student budget. I’m requesting $2,400 in travel support to fly to Australia and present our work in person at the NeurIPS 2026 conference.
Why this research matters for AI safety: Frontier AI safety evals and automated monitors (LLM-as-a-judge) assume models generalize cleanly across diverse inputs. But what happens when safe behavior breaks down under natural linguistic shifts, and semantic meaning seems identical?
Working alongside my co-authors Namirah Rasul and Ewura Ama Sam, we tested MentaLLaMA-chat-7B on 300 controlled prompt variants (100 clinician-grounded scenarios from MentalChat16K held constant across Standard English, Nigerian Pidgin, and Ghanaian English). We uncovered two critical issues:
Aggregates hide tail-risk collapse: While composite quality dropped from 3.94 in Standard English to 3.79 in Pidgin (p = 0.0027), the mean hides what happens in the tail. In the worst scenarios, identical semantic meaning written in Pidgin flipped supportive replies into outright refusals, boilerplate dismissals, or corrective scolding.
Insensitive automated judges: Even with these qualitative failures in a high-stakes setting, standard automated safety rubrics scored 5/5 across all 300 outputs. Automated oversight pipelines can produce dangerous false positives by completely missing out-of-distribution behavioral failure modes.
OpenReview: https://openreview.net/forum?id=ICZEGRAVCn
The core research is finished and accepted. The goal here is to get to Sydney so I can present the findings, gather feedback, and turn this into broader safety benchmark work:
Present the poster at NeurIPS: Share our results directly with researchers focused on out-of-distribution robustness, scalable oversight, and automated red-teaming.
Stress-test LLM-as-a-judge rubrics: Discuss methods with evaluation researchers on how to design automated evaluators that detect tail degradation instead of averaging it away.
Scale to frontier safety benchmarks: I want to adapt this controlled semantic-invariance setup into an open-source evaluation suite (compatible with platforms like UK AISI’s Inspect AI) to test whether frontier models exhibit sandbagging or alignment breakdown under dialectal shifts.
Support fieldbuilding: As co-founder and research lead of AI Safety Hub Nigeria, attending in person allows me to bring back technical connections, open problems, and collaboration opportunities to our local fellows.
My conference registration is already covered. All requested funds go directly toward necessary travel costs:
Round-trip flight (Nigeria – Sydney): $1,500
Accommodation (5 nights in Sydney): $650
Visa: $150
Meals $100
Total Requested: $2,400
Minimum funding goal ($1,650): $1,650 covers my flight and visa, so I can physically reach Sydney. If I only hit the minimum, I will try my best to cover food, transit, and shared-budget hostel lodging out of pocket. If any pending grants come through, I will reduce this goal immediately.
I am the first author, collaborating with co-authors Namirah Rasul and Ewura Ama Sam. I study mechatronics engineering at Ahmadu Bello University in Nigeria.
Over the past three years, my focus has been on empirical LLM evaluations and building technical AI safety capacity:
Evaluations & Research: Built SandbagBench-Lite (a black-box harness for measuring strategic underperformance in LLMs) and worked as an AI Safety & Evaluation Research Intern at HeyLina (designing conversational evaluation rubrics and LLM-as-a-judge frameworks). Authored Forecasting AGI: A Granular, CHC-Based Approach and received recognition in the ENAIS AI Safety Action Competition for research on behavioral adaptation (Punctuation Paranoia).
Fellowships: 2024 Non-Trivial Fellow, Cooperative AI Research Fellowship finalist, and alumnus of the TReND-CaMinA ML & Computational Neuroscience program in Kenya.
Community Leadership: Co-founder and Technical Lead of AI Safety Hub Nigeria, President of Effective Altruism ABU, and Policy Lead for PauseAI Nigeria.
Without travel funding, I cannot afford the flight to Australia on a student budget. The paper remains in the official proceedings, but we lose the chance to present the poster in person, get technical feedback from leading AI evals researchers, and build collaborations. The other main variable is visa processing, which I am handling by submitting the application immediately using my official NeurIPS acceptance documentation.
$0 so far. Outside of my registration being taken care of, I haven't raised anything for travel yet. I’ve put in applications for a BlueDot Rapid Grant and NeurIPS travel support, and I'll adjust the target here straight away if either comes through.