Study questions how soon AI can improve its own research
A study led by Peter Kirgis and Sayash Kapoor of Princeton University suggests that AI agents may be farther from recursive self-improvement than some industry forecasts imply. The researchers found that agents could handle many technical tasks involved in AI research, but struggled with the judgment and creativity needed for open-ended investigation.
Using a method called “shadow evaluation,” the team asked Anthropic’s Claude Opus 4.8, running on OpenClaw, to address questions from two unpublished papers submitted to NeurIPS 2026. The agents had six days, $3,000 in Anthropic API credits, computing resources, virtual machines and web access. They reviewed prior work and ran hundreds of experiments, but the papers they produced were rejected by the original authors.
The agents used weak datasets, abandoned promising ideas too quickly, failed to rethink unsuccessful approaches and did not effectively apply feedback or manage time and computing resources. They did not, however, appear to hide or falsify results. The researchers noted that the study covered only two papers and involved subjective evaluation. Kapoor said the findings may challenge short timelines for recursive self-improvement, while leaving open whether progress on narrow, measurable tasks could eventually be sufficient.
Written by Sakhanda Wire AI from reporting by MIT Tech Review. Facts, quotes and full context are in the original article below.