Out of over 21,000 submissions to one of the world’s toughest artificial intelligence conferences, one of the standout papers this year came from an African-based institution.
Daniel Rajaonarivelomanantsoa, a 27-year-old PhD student from Madagascar who honed his skills through Google’s AI for Science program at AIMS, has co-authored a paper ranked among the top 0.3% and chosen for an oral presentation at NeurIPS.
The selection process is ruthlessly competitive. This year, there were 21,575 submissions, and only 77 papers (0.3%) were chosen for this elite honor, validating the work as among the most impactful research produced worldwide that year.
It’s a remarkable milestone, both for African research and for the field of AI more broadly. Daniel is the first researcher from Madagascar and the first Africa-based researcher to earn this recognition. His achievement reflects the depth of talent emerging across the continent and the important impact of mentorship and institutions that support young scientists in pursuing ambitious work.
It’s a reminder that influential contributions to global science are coming from places long overlooked and that African researchers are increasingly driving the direction of the field.
This success underscores the essential role diversity plays in AI, where talent that sees the world in a slightly different way is necessary to push the true frontier of knowledge.
The journey: From Madagascar to frontier research
Daniel’s story demonstrates what becomes possible with talent, opportunity, and the right institutional backing. After completing his studies, he took a job in financial data science in Madagascar, but the work, he recalls, quickly felt limiting. “It wasn’t really research. It was repetitive. I knew I wanted to do something deeper,” Daniel says.
With little clarity on what might come next, he discovered the AI for Science Master’s program at the African Institute for Mathematical Sciences (AIMS) South Africa, funded by Google DeepMind. He prepared his application in two weeks and was awarded the scholarship.
The fully residential, year-long program brings together promising students from across the continent and trains them intensively in the mathematical foundations of modern AI research. It welcomed its first cohort in 2023 and is funding 40 full scholarships a year until 2027, giving scholars sustained access to teaching and mentorship from Google DeepMind researchers and engineers.
“AIMS was incredibly intense: we had three courses in three weeks, when most students might only do one, but it taught me to think fast and adapt quickly,” Daniel says. “It gave me the solid foundation I needed to start contributing to research.”
That set the stage for Daniel’s next step: joining InstaDeep, an African-founded deep-tech company with a research track record in reinforcement learning, computational biology and large-scale decision-making systems. InstaDeep’s scientists teach at AIMS, and it was through those interactions that Daniel encountered the research culture that ultimately defined his work.
He joined the company’s Reinforcement Learning team in Cape Town, led by Arnu Pretorius, where he was supervised by Claude Formenak and Felix Chalumeau. He later began his PhD at Stellenbosch University, supported by InstaDeep.
When his student visa required him to return to Madagascar, InstaDeep ensured continuity in his research by providing secure access to their high-performance computing infrastructure and ongoing mentorship, reaffirming their long-term commitment to his development. It has created an unusual setup: a young researcher contributing to a high-performing team from Antananarivo, without a formal lab.
Rethinking AI decision-making
The study—Breaking the Performance Ceiling in Reinforcement Learning requires Inference Strategies—led by Daniel, Felix Chalumeau, and Ruan de Kock of InstaDeep, tackles a persistent challenge within reinforcement learning, a major subfield of AI: In complex real-world tasks, AI systems often hit a performance limit, or “ceiling,” even after extensive training. This happens because the traditional quick method of decision-making, known as the zero-shot approach, relies only on a single, instantaneous response after initial training, which is often not accurate enough.
Daniel’s team demonstrated the limitations of this traditional quick-response approach and highlighted the need for a superior yet often overlooked paradigm: inference-time search.
Instead of relying on that single, instantaneous decision, inference time dedicates a small amount of extra “thinking” time right before the AI acts. This brief window allows the system to quickly explore and evaluate multiple potential outcomes before selecting the single best action.
Using this strategy, the researchers achieved up to a 126% improvement in performance over the zero-shot approach with just 30 seconds of added computation time. This finding highlights that the key to powerful AI is not just training models harder or longer, but teaching them to use their existing knowledge smarter at the moment of decision. This enables far stronger decision-making than was previously possible, moving AI systems from acceptable results to the optimal outcome.
Opening doors
The technical contribution is only part of the story. Daniel is frank about the structural barriers facing researchers in Madagascar and much of Africa. “In Madagascar, many students at my university didn’t even have access to laptops. Compute (processing power and memory needed to run applications and perform tasks) is extremely hard to get. The internet is very expensive,” he says, adding that he knows many “far more talented” people who will never publish because the basic tools of research are simply out of reach.
“This result isn’t about me being exceptional,” he says. “It’s about being given a chance.”
For a community used to watching frontier AI emerge from a handful of global labs, Daniel’s appearance on the NeurIPS stage offers a model of African researchers not merely participating in AI research but starting to push its boundaries. The bottleneck has never been talent; it’s always been access.
Daniel hopes to direct his research towards healthcare, particularly in regions where limited medical infrastructure contributes to preventable deaths. “If AI can help reduce that gap, even a little, that’s where I want to focus,” he says.
His story is a reminder that world-class scientific work can come from anywhere, but only if opportunities are created and talent nurtured.
