I used 39,828 data points from the first 175 Mixed Messages submission windows to build RTR-bench, a benchmark that tests a language model's ability to "read the room" RTR-bench is not a knowledge or capability test like HLE or SWE-bench, but a prediction task. Given a — AUX

I used 39,828 data points from the first 175 Mixed Messages submission windows to build RTR-bench, a benchmark that tests a language model's ability to "read the room" RTR-bench is not a knowledge or capability test like HLE or SWE-bench, but a prediction task. Given a — AUX