Newest papers

OmicsBench: a long-horizon multi-turn benchmark of AI agents on research-study-scale bioinformatics

Gideon Shaked, Siddharth Nagisetty · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.21074 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

Los biólogos computacionales cada vez más delegan análisis sustanciales a los agentes de IA en sesiones de rango largo, multi-turn, pero la mayoría de los puntos de referencia de las capacidades de bioinformática de los agentes plantean una única tarea corta con una promesa de disparo único. Presentamos OmicsBench, un punto de referencia de bioinformática de rango largo y multi-turn para los agentes de IA. Cada una de sus 97 tareas enlaza las subtascas dependientes en un gráfico acíclico dirigido, para 662 subtascasas subyacentes en total. Cada subtasca se presenta como una vuelta de usuario y se clasifica deterministicamente contra una clave de respuesta, con feedback opcional en contexto sobre respuestas incorrectas. A través de 12 modelos en tres pipes, las medias

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

Abstract

Computational biologists increasingly delegate substantial analyses to AI agents in long-horizon, multi-turn sessions, yet most benchmarks of agents' bioinformatics capabilities pose a single short task with a single-shot prompt. We present OmicsBench, a long-horizon, multi-turn bioinformatics benchmark for AI agents. Each of its 97 tasks links dependent subtasks into a directed acyclic graph, for 662 underlying subtasks in all. Each subtask is posed as one user turn and graded deterministically against an answer key, with optional in-context feedback on incorrect answers. Across 12 models on three harnesses, the slowest agent averages 3.6 hours per task, and no combination comes near saturating the benchmark, with mean scores of 0.43 to 0.57 out of a possible 1 without in-context feedback and 0.54 to 0.67 with it. Multi-turn evaluation also uncovers behavior that single-shot prompts cannot, as agents score lower on subtasks that depend on earlier subtasks, and in-context feedback helps agents recover some of that loss. Because deterministic scoring gives no qualitative account of how an agent failed, we also introduce an unsupervised failure mode discovery pipeline that describes, embeds, and clusters agents' failed attempts to surface emergent failure modes. Applied to OmicsBench, this pipeline exposes qualitative shortcomings in how agents carry out bioinformatics analyses, ranging from agents applying their own statistical filters in place of the specified ones to agents writing whole answers from memory.

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

Topics: Delegate, Benchmark (surveying), Pipeline (software), Computer science, Task (project management), Artificial intelligence

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