A benchmark from P-Zero Research is measuring whether AI systems can design experiments and interpret their results across eight model-development tasks. Anthropic cofounder Jack Clark says performance has roughly doubled every three months since December 2025, leading him to move his forecast for possible recursive self-improvement from 2028 to 2027.
TasteVal covers tasks involving training-data selection, pretraining, fine-tuning, preference learning and defenses. The research group calls the measured capability “research taste.” Its best tested model reportedly outperformed a comparison group of 24 human experts at much lower cost.
The name is broader than the measurement. Each task already supplies a fixed research goal and a score, so the system does not have to decide which question is worth pursuing. TasteVal evaluates experiment design and interpretation, but omits the selection of promising directions that many scientists consider central to research judgment.
The results therefore do not show that an AI system can improve its own successor without people. They indicate rapid progress in a narrower, structured part of model research. The human comparison is also small and depends heavily on how tasks, time and costs were defined. TasteVal can provide evidence about automated experimentation, while Clark’s 2027 date remains a forecast. Stronger support for recursive improvement would require systems to choose valuable questions, run reliable programs and produce gains that persist beyond a fixed benchmark.