AI Evaluates AI Researchers: Gemini and Claude Agree, but GPT-5.4 Does Not

Autonomous research systems are already generating articles, but how can we determine which one is better? A group of researchers proposed the first quantitative benchmark: four frameworks (Sakana AI v1 and v2, CycleResearcher, Data-to-Paper) were given the same set of 15 research proposals and generated 60 papers. Three independent LLM reviewers—GPT-5.4, Gemini, and Claude—based on four criteria: originality, scientific rigor, clarity, and significance. The results are published on arXiv.

The main finding is encouraging: Gemini and Claude demonstrate nearly perfect agreement with each other (ρ = 0.907, p < 0.001) and an extremely strong correlation with the synthesized evaluation (ρ = 0.961). This means that automated peer review using LLMs can function as a scalable and consistent evaluation system—at least when reviewers assess articles in the same way.

But GPT-5.4 throws a wrench into the picture. Its correlation with other models is only ρ ≈ 0.32. The authors cautiously explain this by noting that the model “uses different evaluation criteria.” In practice, this means that if you choose GPT-5.4 as the sole reviewer, the system rankings will look fundamentally different than when using Gemini or Claude. For a benchmark aspiring to be a standard, this is a serious flaw.

Here’s another inconvenient fact. The research proposals used to test all four frameworks were published by the commercial company FARS—which also provided the 15 reference articles. The result: FARS’s articles score 2.14–2.47 out of 5, while all others score between 1.00 and 1.87. In the evaluations by Gemini and Claude, FARS’s lead is more than double. When the benchmark organizer is also the winner, questions about the independence of the evaluation hang in the air—even if the LLM reviewers are formally independent.

What does this mean in practice? Multi-model LLM review is a useful tool, but only if you verify agreement among reviewers and do not blindly trust a single model. GPT-5.4, Gemini, and Claude don’t just give different scores—they appear to optimize for different notions of quality. Before incorporating an automated reviewer into your pipeline, it’s worth determining exactly what “flavor” of evaluation each model produces, and whether your benchmark might actually reflect a single vendor’s preferences rather than true quality.

Source: cs.AI updates on arXiv.org