TasteVal pushes AI research taste into the recursive self-improvement debate

A new benchmark from P-Zero Research is putting a sharper measurement on one of the hardest questions in artificial intelligence: can models decide which experiments to run, learn from the results and improve the research process itself? TasteVal measures that narrower capability — experimental research taste — rather than claiming that machines can independently choose the most important scientific problems. ¹²

The timing matters because Anthropic co-founder Jack Clark has moved his estimate for recursive self-improvement (RSI) forward to 2027. Clark's prediction is still a forecast, not a result demonstrated by TasteVal, and it comes from a prominent figure at a company whose model performed strongly in the benchmark. The distinction is important: a model becoming better at a constrained research task is not the same as an AI system designing and training a stronger successor without meaningful human direction. ²

What TasteVal actually measures

P-Zero Research defines research taste as the ability to identify worthwhile questions, design experiments and interpret their results. TasteVal tests only the experimental part. Each of its eight tasks supplies a fixed research goal, while the model chooses experiments and uses their outcomes to guide the next step. A separate coding agent runs the work on a single Nvidia H100, limiting the influence of raw programming ability. ¹

The benchmark compares models with the best attempt from 24 human experts who had recently worked at organisations including OpenAI, Google DeepMind, Nvidia and Microsoft Research. The headline result is a 2.30x compute multiplier for Anthropic's Opus 5.5: under P-Zero's definition, it reached expert-level performance using less than half the experimental compute. The reported uncertainty is wide, however, with a 95 per cent confidence interval from 1.15x to 4.37x. ¹

P-Zero also reports that experimental research taste has roughly doubled every three months since December 2025. That is a striking curve, but it is not a guarantee of continued progress. The tasks are relatively quick and cheap to verify, which may favour the kinds of research work that AI labs already know how to optimise. The human comparison also does not include the world's top researchers, so the benchmark may understate the best human baseline. ¹

Why the 2027 prediction needs caution

TasteVal makes a useful contribution because it measures a capability that has often been discussed more vaguely. It shows that models can make increasingly effective choices inside a defined experimental loop. It does not show that they can select a long-term research agenda, recognise which unanswered question is most valuable, or judge whether a surprising result deserves a new line of investigation.

Those gaps are exactly where the recursive self-improvement argument becomes much more demanding. Clark's 2027 estimate is therefore best understood as a response to an encouraging measurement, not as a date stamped on the arrival of self-improving AI. The benchmark narrows part of the distance; it does not erase the rest.

Our opinion

TasteVal is more interesting than another leaderboard because it tests decisions rather than just answers. The result should make AI researchers take model-led experimentation seriously, especially when a system can reach a useful result with far less compute than a human team. But the benchmark's limitations are doing as much work as its headline number. Until models can choose the right problems and explain why those problems matter, “research taste” remains a promising component of scientific autonomy rather than proof of recursive self-improvement.

For now, the responsible reading is neither panic nor dismissal. The experimental loop is getting better quickly, according to P-Zero's data, while the broader research process remains substantially human-shaped. Future benchmarks will need to test whether that missing judgement can be measured too.