The benchmarks evaluate LM agent on SWE/Computer-use tasks across different operating systems.