Skip to content
Nakoda AI News
The Brief
Markets Are Starting to Price In How Companies Govern Their Own AIIndependent AI Assurance Is Becoming Its Own Category of Institutional TrustThe Productivity Gain From AI in Auditing Is Real. So Is the New Bottleneck It CreatesWhat Investors Actually Look for When They Ask "How Mature Is Your AI Operation"The Customer Service Bot That Remembers Too Much, and What It's Doing to TrustCFOs Are Being Asked to Justify AI Spend the Way They Justify Everything ElseData Never Stops at a Border, But AI Governance Rules DoWhen the Model Is Wrong, Who Actually Carries the Liability?
Research

New benchmark finds models still struggle with long-horizon planning

Researchers tested a dozen frontier models on tasks requiring 20+ sequential steps.

By Nakoda Newsroom

·5 min read

Prefer Nakoda AI News on Google

A new benchmark released this week measures how well AI models handle tasks that require planning more than twenty steps ahead.

Across a dozen frontier models tested, average success rates dropped sharply after the tenth step, with most models failing to recover from early mistakes.

The researchers say the results suggest current training methods reward short-term coherence over long-term plan correction.

TagsResearchEngineeringTechnology

Written by

Nakoda Newsroom

Independent journalism at the intersection of AI, business and society. Part of the Nakoda AI ecosystem.

Follow