Mechanism behind low chain-of-thought controllability

Identify and characterize the causal mechanisms in training and inference that lead contemporary reasoning models to exhibit low chain-of-thought controllability relative to output controllability.

Background

Across extensive evaluations, the paper finds that models are far less able to control their CoT than their final outputs, and that controllability further decreases with longer reasoning, more RL training, and task difficulty.

Despite these empirical patterns, the authors state that the underlying mechanism causing low CoT controllability is not yet understood, motivating investigation into factors such as training objectives, process supervision, and model architecture.

References

We leave explicit control of CoT lengths under retrospective thinking to future work.

— RetroThinker: Enabling Retrospective Thinking in Speech LLMs  (2609.11864 - Shih et al., 10 Sep 2026) in Section 5, subsection “Main Results”

However, the mechanism behind low controllability is not well understood.

— Reasoning Models Struggle to Control their Chains of Thought  (2603.05706 - Yueh-Han et al., 5 Mar 2026) in Abstract

This establishes the capability, but whether they can also do so without an explicit instruction, or would do so on their own, is not known.

— Reasoning Models Can Control Their Chains of Thought When They Tell Themselves To  (2610.23257 - Dolev et al., 16 Oct 2026) in Section titled “Limitations,” paragraph “Capability, not propensity”