In 2020, Giambattista Parascandolo gave a presentation on using GPT for reasoning during his interview for a faculty position at MIT. According to a report by Jiqizhixin, most professors on the interview committee dismissed this direction as "nonsense." The presentation's theme was how to enable artificial neural networks to break through their training distribution and achieve more human-like generalization and planning abilities.
Parascandolo argued that humans can recombine existing knowledge, identify key invariants, build abstract models, and perform long-horizon planning, while artificial neural networks still have significant room for improvement in these areas. The presentation concluded with three future research directions: open-ended reasoning in neural networks, unexplored degrees of freedom in artificial neural networks, and using language as a reasoning medium in reinforcement learning to improve sample efficiency.
The most critical concept was "open-ended reasoning," which Parascandolo defined as: the model can invest more time and computation to continuously revise its answers; the harder the problem, the more steps the model should think, and the extra computation should translate into better results. This already sounds very much like today's test-time compute scaling.