[Submitted on 8 Oct 2025 (v1), last revised 7 Jul 2026 (this version, v4)] · arXiv.org

View PDF HTML (experimental)

Abstract:What do thinking language models learn during training that their base models lack? We first present an unsupervised method that discovers a model's reasoning behaviors by training small Sparse Autoencoders on sentence-level activations of reasoning traces, yielding interpretable reasoning taxonomies. Building on this, we introduce constructive model diffing, which aims to reconstruct the base-to-fine-tuned difference from interpretable components: reasoning mechanisms (category vectors that can induce a reasoning behavior in the base model) and reasoning heuristics (a classifier determining when a mechanism should fire). Across nine base/thinking pairs (four RL-trained, four SFT-distilled, one mixed), two independent findings agree: category vectors in the base model converge to far lower loss for taxonomies derived from purely RL-trained models, and hybrid models recover roughly 76% of the RL base-to-thinking gap but only 11% of the SFT gap. This indicates RL primarily teaches heuristics for orchestrating pre-existing base mechanisms, whereas SFT-distillation installs new ones, offering a new lens on what training paradigms teach, with implications for efficient reasoning-model development.
Comments: Accepted as a Spotlight at the International Conference on Machine Learning 2026
Subjects: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Cite as: arXiv:2510.07364 [cs.AI]
  (or arXiv:2510.07364v4 [cs.AI] for this version)
  https://doi.org/10.48550/arXiv.2510.07364

arXiv-issued DOI via DataCite

Submission history

From: Constantin Venhoff [view email]
[v1] Wed, 8 Oct 2025 17:58:28 UTC (543 KB)
[v2] Fri, 10 Oct 2025 17:48:16 UTC (544 KB)
[v3] Wed, 22 Oct 2025 16:02:22 UTC (544 KB)
[v4] Tue, 7 Jul 2026 14:19:58 UTC (3,623 KB)

Read the original on arxiv.org ↗