Archive · 24 Jul 2025
AI Study: Longer Reasoning Time Can Degrade Model Performance
Imagine you're trying to solve a puzzle, but taking too long can make it harder.
That's what this study found about AI.
Researchers discovered that when AI models are given more time to think, they sometimes get worse at their tasks.
This happened with different AI models, like those from Anthropic and OpenAI.
Instead of getting smarter, the AI got confused or focused on the wrong things.
It's like trying to study for a test, but spending too much time worrying about things that don't matter.
Anthropic study reveals that extended AI reasoning time can degrade performance.
Experiments included models from Anthropic, OpenAI, and DeepSeek, showing declining accuracy in reasoning tasks.
Claude models became sensitive to irrelevant information, while OpenAI models overfit to problem types.
Longer reasoning didn't improve deductive puzzle results, leading to confusion and reduced precision.
The findings challenge the assumption that more compute always enhances AI capabilities.
- Who
- Researchers from Anthropic conducted the study, testing various AI models.
- What
- A study showed that giving AI more time to 'think' can make it perform worse.
- When
- The study was recently conducted.
- Where
- The study was conducted in the context of AI model research.
- Why
- To challenge the assumption that more compute time always improves AI performance and explore new AI safety issues.
Traditional View
Study Findings
Effect of Compute Time
Traditional View
More compute time enhances AI performance
Study Findings
Extended reasoning time can diminish performance
- Study Focus
- Impact of extended reasoning time on AI performance
- Key Finding
- Extended reasoning can degrade AI accuracy
- Models Tested
- Anthropic's Claude, OpenAI's o-series, DeepSeek, and others
- Tasks Affected
- Reasoning, deductive puzzles, and regression tasks
- Implication
- Companies may need to reevaluate compute time for AI models.