Fine-tuning experiments on CoT controllability

p:has(> img) { margin-bottom: 0; } .content img { margin: 0.75em 0; } Kei Nishimura-Gasparian is an Astra fellow and was the primary contributor to this work. Neev Parikh provided mentorship and feedback. Summary: We find that a small amount of fine-tuning on instruction following in the CoT generalizes to meaningful increases in CoT controllability on an out-of-distribution set of tasks (CoTControl eval suite). We fine-tune four reasoning models on small datasets (240 examples or ~100K-300K tokens of fine-