Extent of new versus repurposed reasoning capabilities in thinking language models

Determine to what extent thinking language models learn entirely new reasoning capabilities during post-training versus repurpose pre-existing capabilities and representations already present in their base models acquired during pre-training, in order to explain their superior performance on reasoning tasks.

Background

The paper examines why reasoning-focused ("thinking") LLMs outperform their base counterparts on benchmarks such as GSM8K and MATH500. It proposes evidence that base models already possess many reasoning mechanisms and that post-training (e.g., RLVR) primarily teaches when to deploy them. This question—new capability acquisition versus repurposing existing ones—frames the central contribution, motivating an unsupervised taxonomy of reasoning mechanisms and a hybrid steering approach to activate latent reasoning behaviors in base models.

Clarifying this distinction has practical implications for training efficiency and interpretability. If thinking models mostly learn orchestration rather than novel mechanisms, future methods might target activation-space interventions or timing policies, reducing reliance on extensive fine-tuning.