OpenAI Invites METR to Review Its Largest Reinforcement Learning Training Run
TechFlow reports that on November 17, according to The Defiant, OpenAI has invited non-profit evaluation agency METR to review its largest reinforcement learning training run to date, investigating behaviors that misalign with developer intent, including issues that may have been missed by the company's own testing and monitoring.
OpenAI will provide METR with model snapshots, conversation logs, and monitoring records from the training process for its review and experimentation. METR stated that the pilot evaluation covers only a single training run and does not include misalignment incidents already disclosed by OpenAI. Both parties have tentatively agreed to conduct a broader evaluation of the monitoring system, but the specific scope still needs to be finalized.





