A single prefilled token can substantially improve control over AI reasoning traces

The study reveals an entropy bottleneck in the first few reasoning tokens and shows that simple interventions can improve compliance by 34–41 percentage points for additive and stylistic control, but less for suppression.

Top University

Hugh Van Deventer · Dmitrii Troitskii · Laura Ruis

Harvard University · Cambridge Boston Alignment Initiative · MIT

Research Digest·2 November 2026·2 min read

Illustration: A single prefilled token can substantially improve control over AI reasoning traces

In one line

Chain-of-thought controllability is limited by an entropy collapse in the first few tokens after the thinking tag, and can be recovered by prefilling or banning frequent openers.

Before you read this

An idea this paper builds on

runtime monitoringAudits an AI agent's live actions, internal states, or effects against constraints or policies to catch errors and attacks.110 papers →

Read it as

Some AIs show their reasoning step-by-step to let humans verify honesty. This oversight only works if the AI cannot easily fake its thinking. Current AIs struggle to control their reasoning, but why was unclear.

The researchers found that the very first words after the AI starts thinking are highly formulaic. Two interventions, prefilling the opening or banning common starting words, boosted the AI's compliance by 34 to 41 percentage points.

This means the main hurdle is breaking from a templated start, not a general inability. As models improve, they may overcome this, making monitoring harder. Additionally, adding and suppressing information act as separate skills, so evaluations should treat them independently.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

chain of thought runtime monitoring token prefilling in-context learning