SFT vs RL Is the Wrong Question in Post-Training
A paper posted to arXiv reframes how the field should think about post-training. The central claim: supervised fine-tuning and reinforcement learning both reweight a pretrained model's reference distribution. The real question is whether that reweighting stays inside behaviors the model could already reach, or expands that reachable set.
The authors introduce the concept of 'accessible support', the set of behaviors a model can practically produce under finite compute budgets. Training that shifts probability mass within that support is capability elicitation. Training that changes the support itself is capability creation. The free-energy framing treats demonstration signals and reward signals as two versions of the same mechanism, not fundamentally different operations.
This matters for anyone building on top of foundation models. Labs, operators, and evaluators have used SFT versus RL as a proxy for 'imitation versus discovery.' That proxy is now contested. The paper argues search, interaction, tool use, and new information are the real levers for capability creation. Watch how safety evaluators and post-training teams update their benchmarks in response.