Newest papers

ProteinGeoBench: Frontier Structural-Biology Agents Are Fluent at Lookup but Lost in Geometry

Matt Turk, Brandon L. Neel, Diana M. Lopez, Zachary Daniel Smith, Grace N. Johnecheck, James Brown Golden III · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26829 · cited by 0 works (OpenAlex) · open-access PDF

German

Wir fragen, wie viel von einem Benchmark-Score diese Lücke widerspiegelt und wie viel die Bewertungsaufgabekonfiguration um das Modell widerspiegelt. Wir studieren ProteinGeoBench, eine Suite von 131 Protein-Struktur-Aufgaben, mit 4,249 Rollouts von vier Grenzmodellen und 4,486 weiteren Tests über 26 Modelle. Drei der vier Grenzmodelle liegen innerhalb von 0,031 voneinander (0,785 bis 0,816) ohne Unterschied zwischen ihnen gelöst, während die vierte eine Bewertung von 0,707 erzielt und alle drei Lücke dazu haben, Null auszuschließen. Drei der vier machen es besser bei der Metadata-Suche als bei der numerischen Berechnung. Das vierte ist etwas besser bei der numerischen Berechnung. Die numerischen Aufgaben fallen

Machine translation (Meta M2M100); the English original follows.

Abstract

Frontier agents retrieve structural metadata well and compute geometry from coordinates less reliably. We ask how much of a benchmark score reflects that gap and how much reflects the evaluation configuration around the model. We study ProteinGeoBench, a suite of 131 protein-structure tasks, using 4,249 rollouts from four frontier models and 4,486 further trials across 26 models. Three of the four frontier models fall within 0.031 of one another (0.785 to 0.816) with no difference between them resolved, while the fourth scores 0.707 and all three gaps to it exclude zero. Three of the four do better on metadata lookup than on numeric calculation. The fourth is slightly better on numeric. Numeric tasks are 28 of the 131 but half of the twelve tasks that all four models fail. The ranking also depends on the evaluation budget: with one run per task, 23.9% of bootstrap replicates put a different model first. On a selected 15-task slice, removing network access lowers scores more than blocking eight structural-database domains, but the two contrasts were graded in different passes and cannot be cleanly decomposed. Swapping in each model's own command-line agent changes scores by -0.071 to +0.086, a comparison too small to rule out moderate harness effects. Across 26 models, numeric performance rises with non-numeric performance rather than staying flat. The experiments use different task subsets, so they cannot rank model, harness, access and task composition against one another. Evaluations should report results by task family, along with the number of runs, the harness and the access conditions.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Task (project management), Benchmark (surveying), Computer science, Metadata, Ranking (information retrieval), Frontier

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange