Interpreting and advancing ARC-AGI 2 — probing the limits of abstract reasoning, compositional generalisation, and genuine novelty-handling in frontier AI systems.
ARC-AGI 2 represents one of the most demanding evaluations of abstract reasoning in AI: tasks designed explicitly so that memorisation and pattern-matching are insufficient. Success requires genuine compositional generalisation — the ability to combine known primitives in ways never encountered during training.
3 project leads and 6 researchers tackling abstract reasoning and AGI evaluation through 2026–27.
Applications for Viveka 2.0 will open shortly. Stay tuned for details on how to join the team for the 2026–2027 tenure.