Inverse Scaling in Test-Time Compute

Allocating more test-time computation often improves large reasoning models, but the relationship is not universally monotonic. This work constructs tasks in which longer reasoning lowers accuracy, covering counting with distractors, regression with spurious features, constraint-tracking deductions, and advanced-AI-risk evaluations.

The experiments identify several mechanisms behind inverse scaling. Depending on the model family and task, extended reasoning can increase distraction, encourage overfitting to the problem framing, replace reasonable priors with spurious correlations, or make it harder to maintain focus through a complex deduction.

The results show why evaluation at a single reasoning budget can conceal important behaviour. Measuring performance across reasoning lengths is necessary both to characterise model capability and to detect undesirable patterns that additional computation may reinforce.

Paper: https://openreview.net/forum?id=NXgyHW1c7M