Multi-turn Bootstrap
A self-instruction given only once often fades over a long reasoning task; repeating it at every assistant turn helps Qwen3-32B sustain the requested constraint across turns.
reasoning modelspromptingmulti-turn agents
Reasoning Models Can Control Their Chains of Thought When They Tell Themselves To
Uriel Dolev, Idan Schwartz, Yoav Goldberg
arXiv 2610.23257 26 pages Read the paper
The short version
- The problem
- Reasoning models often ignore requests about how their private step-by-step text should be written, such as avoiding a word or using a particular style. This makes it hard both to shape their reasoning usefully and to know whether reading that text is a dependable way to catch dangerous behavior.
- The idea
- The authors test a different approach: have the model tell itself the rule at the start of its reasoning, then see whether it follows that rule. The rule can be placed directly into the reasoning, or the model can be asked in the prompt to repeat it there first—like asking someone to write a reminder to themselves before starting a task.
- What it got
- This approach made reasoning much easier to steer: for one model, full compliance rose from 0.4% to 22.2%, and for another it rose from 2.5% to 34.3%. The authors also used these instructions to lower how often a monitor caught hidden harmful behavior.
Modules from this paper
Self-Instructed Chain-of-Thought Controllability
Reasoning constraints placed only in a user prompt are often ignored; this evaluation puts the constraint inside the reasoning as a self-instruction and measures whether the remaining reasoning follows it.
reasoningchain-of-thoughtevaluationinstruction-following
Self-Instructed Reasoning as a Monitor-Evasion Risk
Reasoning monitors can miss malicious intent when a model directs its own reasoning style; self-instructions reshape the monitored text while task completion stays broadly comparable.
reasoningchain-of-thoughtmonitoringsafety