Study: AI Still Unable to Recursively Self-Improve

An experiment conducted by MIT researchers has demonstrated that modern AI models are still unable to achieve recursive self-improvement. In the study, AI agents attempted to replicate scientific research but encountered fundamental challenges. They lacked the creativity required to generate novel ideas and quickly abandoned promising hypotheses based on limited data.

AI agents also struggled to adapt to failures: instead of revising their methodology, they merely narrowed their conclusions or added disclaimers. Additionally, they inefficiently used available resources, such as computational power and time, while ignoring feedback from auxiliary models or external evaluation tools. Notably, the agents did not engage in reward hacking—manipulating data to achieve results—which was the only positive outcome.

The researchers attribute these limitations to the training paradigms of current models. While modern algorithms excel at tasks with automatically verifiable success, scientific research demands an open-ended approach that remains beyond AI’s current capabilities. The team is now testing Anthropic’s advanced Mythos model, though results from these trials are pending.

Despite its findings, the study has limitations: it covered only two scientific projects, and the original authors knew they were evaluating AI-generated texts. Nonetheless, the results cast doubt on claims by companies like Anthropic and OpenAI about the imminent era of recursive AI self-improvement.