The Art of (Artificial) Reasoning

Scaling laws suggest that “more is more” — brute-force scaling of data and compute leads to stronger AI capabilities. However, despite rapid progress on benchmarks, state-of-the-art models still exhibit "jagged intelligence," indicating that current scaling approaches may have limitations in terms of sustainability and robustness. Additionally, while the volume of papers on arXiv continues to grow rapidly, our scientific understanding of artificial intelligence hasn't kept pace with engineering advances, and the current literature presents seemingly contradictory findings that can be difficult to reconcile. In this talk, I will discuss key insights into the strengths and limitations of LLMs, examine when reinforcement learning succeeds or struggles in reasoning tasks, and explore methods for enhancing reasoning capabilities in smaller language models to help them close the gap against their larger counterparts in specific domains.