The 39th Annual Conference on Neural Information Processing Systems (NeurIPS) was capped off by a series of firsts for the InstaDeep team, with researchers leading the way in exchanging ideas and showcasing innovation within the AI community during our visit to the San Diego Convention Center from the 02-07 December.
With five workshops, three accepted papers, including a spotlight session (selected from among just 3% of the over 21,575 submissions this year) and excitingly, our first ever oral presentation at NeurIPS (representing one of just 0.3% of submissions), our busy team made the most of the opportunity to connect, collaborate and share the latest in AI across the week long event.
Bringing African AI expertise to center stage in our first oral presentation at NeurIPS
Breaking the Performance Ceiling in Reinforcement Learning requires Inference Strategies, led by Felix Chalumeau, Daniel Rajaonarivelomanantsoa and Ruan de Kock drew huge attention at this year’s conference, with the trio taking the stage to share what they’ve learned with an rapt audience..
Representing the most extensive investigation into inference-time strategies to date, and drawing on more than 60,000 experiments, the paper demonstrates their potential to unlock significant improvements, showing that inference-time methods are not just optional refinements but key elements of performance for reinforcement learning (RL).
But it’s more than just a first for InstaDeep. Daniel is the first researcher from Madagascar to earn this distinction, representing a huge step forward for our support for the development of AI in the region and a significant step toward promoting Africa as a hub for excellence in AI.
• Learn more in this video
• See more from Daniel on Inference Strategies.
• Read the full paper here.
Papers, posters and meeting people
On posters, in person and online, NeurIPS was a chance to open conversations on our latest research, with two other papers accepted by the conference this year.
Oryx: a Scalable Sequence Model for Many-Agent Coordination in Offline MARL
Oryx is a new sequential model for offline Multi-agent reinforcement learning (MARL) designed to achieve effective, multi-step coordination among many agents in complex environments and represents a significant step forward in offline MARL. Adapting InstaDeep’s Sable retention-based architecture combined with Implicit Constraint Q-Learning (ICQ), the model provides a strong foundation for moving MARL beyond simulation toward impactful real-world applications.
• You can learn more about Oryx from one of the key authors, Oumayma Mahjoub, in this video.
• Read the full paper here.
MEMENTO: Memory-Enhanced Neural Solvers for Routing Problems
Selected as a Spotlight paper at NeurIPS 2025, MEMENTO has the potential to change how we address combinatorial optimisation (CO) problems, among the hardest challenges in computer science , and with real-world importance in logistics, transport, and energy systems.
MEMENTO introduces memory mechanisms that make RL solvers better at using previous attempts to inform new decisions, enabling AI systems to make better use of available compute, achieving state-of-the-art results on 11 of 12 benchmarks and demonstrating an impressive zero-shot combination with our diversity-based solver, COMPASS.
• Learn more from lead author, Felix Chalumeau in this video.
• Read the full paper here.
Connecting with top researchers from around the world
In the spirit of sharing our expertise from the cutting-edges of each field, our team also found the time during a busy conference to run workshops on Genomics interpretability, ML Interatomic Potentials, Digital Biology and more, helping to inspire the researchers who’ll be sharing a new generation of research at NeurIPS in the years to come. And that wasn’t all, our team also connected with attendees on MLIPAudit, our open-source library for benchmarking machine-learned interatomic potentials (MLIPs) across molecular, liquid, and biomolecular systems, designed to achieve consistent comparison through evaluations of accuracy, stability and scaling and with open leaderboards for the fair comparison.
Beyond ranking models, MLIPAudit reveals what makes them reliable in the first place, insight that is critical for developing dependable tools in molecular discovery and biomolecular simulation. Visit our GitHub and learn more on PyPi, here.
NeurIPS this year was a chance to connect with AI research on new levels at an event that’s growing in importance as well as scale. It’s a chance to look ahead at what’s next for a rapidly maturing AI field, and what we saw at the event (and shared) has left us excited for the future, especially for what 2026 means for the development of AI.
