P‑Zero Research has released TasteVal, a benchmark designed to evaluate how well AI systems exhibit “research taste” — the ability to pick interesting problems, design experiments, and interpret experimental outcomes. TasteVal comprises eight tasks intended to represent frontier AI development, including pretraining data curation, pretraining and fine‑tuning of language models, preference modeling, and robustness to adversarial prompts.
The benchmark focuses on the experimental side of research taste: given a fixed research problem, it measures how effectively a model iteratively proposes experiments and draws conclusions from their results. Model outputs are compared to a human baseline drawn from 24 experts (at least two per task) who have recently worked at organizations including OpenAI, Google DeepMind, NVIDIA, and Microsoft Research.
Results: rapid growth and Opus 5.5
According to TasteVal, the experimental research taste of frontier models has doubled roughly every 3.0 months since December 2025 (95% confidence interval 1.7–5.0 months). The top performing model in the benchmark, Opus 5.5, significantly exceeds the human baseline. The reported compute multiplier for Opus 5.5 is 2.30× (95% CI 1.15–4.37×) relative to the expert baseline, while operating at roughly 1/30 of the expert baseliners' average cost per attempt.
The authors note that research taste appeared to grow much more slowly in prior years, with the slope changing around late 2025 and early 2026.
Why this matters: creativity, RSI, and timelines
The TasteVal findings are discussed in the context of recursive self‑improvement (RSI). Previous analyses identified creativity — the ability to generate novel, effective research directions and experimental designs — as a key variable for how quickly AI systems could develop capacities to improve themselves. Earlier in the year the authors observed that AI systems had not yet demonstrated such creativity in a clearly transformative way, although some results accelerating mathematical research were suggestive.
With benchmarks like TasteVal and other recent advances, the authors now see stronger signs that models are becoming better at AI research itself. Based on this acceleration, the author updates their judged timeline: full RSI now appears more likely in 2027, whereas prior to these developments 2028 was considered the more likely date.
Conclusions and further reading
TasteVal indicates that frontier models are increasingly capable of selecting research problems, designing iterative experiments, and interpreting results — and doing so more efficiently. This rapid improvement in experiment design and interpretation has implications for both the opportunities and the risks associated with advanced AI, particularly concerning the possibility of RSI. Stakeholders tracking AI progress may want to monitor these benchmarks and related work closely.
For more detail on TasteVal and its methodology, see the P‑Zero Research introduction.
