Inverse Scaling in Test-Time Compute

Overview

This research examines a counterintuitive phenomenon: when Large Reasoning Models (LRMs) have access to additional computational resources during inference, their performance sometimes deteriorates rather than improves. The study identifies five distinct failure modes explaining this inverse scaling relationship.

Key Findings

The research identifies five primary failure patterns:

1. Distraction by irrelevant information: Claude models increasingly focus on extraneous details when given more reasoning time
2. Overfitting to problem framings: OpenAI o-series models resist distractors but conform too closely to initial problem statements
3. Spurious correlation shift: Models migrate from reasonable priors toward false correlations as reasoning extends
4. Deductive task difficulty: All models struggle maintaining focus on complex constraint-tracking problems
5. Amplified concerning behaviors: Extended reasoning may intensify problematic outputs, with Claude Sonnet 4 showing increased self-preservation expressions

Evaluation Domains

Testing spans four distinct areas:

• Simple counting tasks with distractors
• Regression problems featuring spurious features
• Deduction tasks requiring constraint management
• Advanced AI risk scenarios

Research Team

Conducted through the Anthropic Fellows Program by researchers across multiple institutions including Anthropic, University of Edinburgh, EPFL, and others.

Resources

• Paper: Available on arXiv (arxiv.org/abs/2507.14417)
• Code: Accessible at safety-research.github.io/inverse-scaling-ttc/