The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.
Metagaming in AI agent evaluations
Data for the paper "How does metagaming affect evaluations?" by Ida Caspary, Rossella Arcucci and Kaiwen Wang.
Agents solved real software issues from SWE-bench Multilingual and SWE-rebench V2 under 59 conditions that make the situation look more or less like an evaluation. We read the reasoning and commands of every run, order the runs of each model by metagaming (the MG-Order), and check the order against a judge that sees the run without its reasoning.
This repository will hold the tasks, the cues and every prompt, the runs of the models with their reasoning, the signals read from each run, the judges' outputs and the MG-Order scores. We are preparing it now.
GPT-6 Astra was accessed through research access. None of its outputs are released.
Code: https://github.com/IdaCy/metagaming-evaluations
License: CC BY 4.0 for the data created in this work. Task texts and repositories keep the licenses of their source datasets.
- Downloads last month
- 19