MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models
Benchmarks chain-of-thought monitorability across tasks that include hidden objectives, underperformance and encoded communication.
Verify byline in the original ↗