A new research paper maps the reasoning limits of current AI models, showing where performance degrades predictably with complexity.
The authors propose a benchmark that isolates multi-step logical reasoning from pattern recognition.
Findings are expected to guide next-generation model architectures and training methodologies.
