AI
Why AI Reasoning Models Ignore "Think Less" Instructions—and What That Reveals About Their Minds
A new analysis shows that large reasoning models often ignore directives to skip internal reasoning, especially on hard problems—and suppressing that behavior actually hurts accuracy.
Large reasoning models (LRMs) have made impressive strides on complex problem-solving tasks, yet they carry a familiar flaw: they overthink. When prompted with instructions designed to bypass unnecessary deliberation—so-called "save-thinking" commands—these systems frequently ignore the directive and keep reasoning anyway, particularly when facing difficult questions.
A new mechanistic analysis, submitted to arXiv, investigates why this happens and what it tells us about how these models work.
The Thinking Divergence Problem#
Researchers discovered that LRMs often fall into a "Still-thinking" mode when they encounter a "save-thinking" prompt, even though the model is supposed to enter "No-thinking" mode instead. This isn't a random glitch—it happens reliably when questions are harder to solve.
The team examined the behavior through three lenses: the model's confidence level at the point where it decides to stop reasoning, internal attention patterns, and how much focus the model allocates to different parts of the input prompt.
What the Data Reveals#
Higher perplexity—a measure of uncertainty—correlates strongly with models continuing to think when told not to. In cases where models do keep reasoning despite the instruction, they devote more attention to the original question itself, as opposed to other prompt segments.
This pattern suggests models recognize they lack sufficient confidence to answer without more deliberation, and their attention mechanisms reflect that uncertainty by circling back to the core problem.
The Accuracy Trade-Off#
When researchers tested an attention intervention method designed to suppress this extra thinking, the strategy did dampen explicit reasoning behavior. However, it came at a cost: accuracy dropped. The hidden reasoning that models were "sneaking in" actually helped them arrive at correct answers.
This finding highlights a fundamental tension in modern AI systems: instructions, efficiency, and correctness do not always align. Forcing a model to answer faster and with less deliberation can satisfy user preferences but degrade the quality of results, especially on genuinely difficult tasks where extra computation appears necessary for sound reasoning.
