New paper proposes framework to distinguish AI capability elicitation from creation

A research paper submitted to arXiv on 8 May 2026 proposes a new framework for understanding how post-training methods affect large language models ¹. Titled "On Distinguishing Capability Elicitation from Capability Creation in Post-Training: A Free-Energy Perspective," the paper was authored by Yuhao Li and Shengchao Liu ¹.

The authors argue that post-training research should distinguish between capability elicitation and capability creation ¹. To operationalize this distinction, they introduce the concept of accessible support: the set of behaviors a model can practically produce under finite budgets ¹.

Under this framework, post-training that reweights behaviors within this support constitutes capability elicitation, while changing the support itself represents capability creation ¹. This distinction addresses what the authors describe as an overly coarse treatment of supervised fine-tuning as imitation and reinforcement learning as discovery ¹.

The paper develops its argument through a free-energy view of post-training ¹. Both supervised fine-tuning and reinforcement learning can be understood as reweighting a pretrained reference distribution, differing only in their external signals: demonstration signals define low-energy behavior for supervised fine-tuning, while reward signals do so for reinforcement learning ¹.

When updates remain close to the base model, the paper argues, the primary effect is local reweighting rather than capability creation ¹. Within this framework, the central question shifts from whether post-training is framed as supervised fine-tuning or reinforcement learning to whether it reweights behaviors already within reach or expands the model's reachable behavioral space through search, interaction, tool use, or the incorporation of new information ¹.

The paper is categorized under Artificial Intelligence, Statistical Mechanics, and Machine Learning ¹. It carries the arXiv identifier arXiv:2605.08368 and has been assigned the DOI 10.48550/arXiv.2605.08368 ¹. The first version was submitted on 8 May 2026 at 18:23:25 UTC, with a file size of 55 KB ¹.

The paper is licensed under Creative Commons BY 4.0 ¹. Support for arXiv is gratefully acknowledged from the Simons Foundation, member institutions, and all contributors ¹.