2026-10-22T10:52:23Z https://lifeofresearch.com/oai2

oai:lifeofresearch.com:125 2026-10-22T08:42:32Z type:preprint openaire

10.5281/zenodo.23536189 https://lifeofresearch.com/preprints/125 Pranay Mahendrakar 0009-0003-7224-029X SONYTECH Life of Research 2026 reasoning models post-training mechanistic interpretability attention heads model diffing reinforcement learning with verifiable rewards distillation circuit analysis Preprint Creative Commons Attribution 4.0 (c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research Reasoning models are built by post-training a pretrained model with supervised fine-tuning on teacher traces, with reinforcement learning against a verifier, or with both, and a running dispute asks whether this builds new machinery or redirects machinery the base model already had. A 2025 circuit study gave the first side a concrete form: attention heads that are absent from the base model's reasoning circuit and present after post-training, many of them after distillation and supervised fine-tuning, fewer and less stable after reinforcement learning. In this analysis the author reviews and synthesises published results to show that "new" has at least three separable meanings inside a model - a new representation, new wiring that consumes a representation, and a new policy for when existing behaviour is deployed - and that the measurements now in the record each speak to one of them. A head that enters a thresholded attribution circuit is newly important; whether it is newly built is a further question that the comparison as run cannot answer, because the reference model differs from the distilled model's actual parent, the circuit is defined by contrasting the two models' own response styles, and zero-ablation effects in the same study do not separate emergent heads from base-circuit heads or from heads taken from an unrelated task. The strongest cross-model evidence points to a regime split at the level of deployment: base-model steering vectors directed by the thinking model's own gating recover about 76% of the accuracy gap for reinforcement-learned models and about 11% for distilled ones, and a gap-matched pair computed from that study's table preserves the split. At the level of representation, a direction taken from the base model drives backtracking in a distilled model but not in the base. The paper consolidates these values in two tables and a figure, gives a classification procedure, and names five cross-model tests that would show whether a post-training head is new.