Hassid et al. argue LLM post-training is just massive supervised learning

In a new position paper on arXiv, researchers Michael Hassid, Yossi Adi, and Roy Schwartz from the Hebrew University of Jerusalem argue that the current LLM training paradigm—relying heavily on post-training phases like supervised fine-tuning (SFT) and reinforcement learning (RL)—is effectively a return to the supervised learning approach of the BERT era. They trace the evolution of LLMs and note that today's models are increasingly tailored to specific benchmarks, mirroring early days when task performance depended on in-distribution dataset fitting.

To support their claim, the authors conducted experiments comparing pre-trained models to randomly initialized ones, fine-tuning both on modern reasoning datasets and evaluating them on competitive math and code benchmarks. They found that models post-trained from scratch achieved highly non-trivial performance, suggesting that current post-training functions primarily as a distribution-fitting mechanism rather than genuine generalization. The paper concludes by advocating for a shift toward training procedures where models 'learn how to learn', moving beyond extensive post-training for predefined behaviors to develop more generally capable systems.