开启时会模糊预览图，关闭后正常显示
The 2025 reward hacking hall of fame award goes to GPT-5.1 for calling the calculator tool to calculate 1+1 on 5% of prod traffic. Because on many prompts using the calculator was superficially rewarded (as a "search") during RL. 🤗
agent safety @openai | previously: @AISecurityInst @AnthropicAI @nyuniversity @SussexUni
The 2025 reward hacking hall of fame award goes to GPT-5.1 for calling the calculator tool to calculate 1+1 on 5% of prod traffic. Because on many prompts using the calculator was superficially rewarded (as a "search") during RL. 🤗
agent safety @openai | previously: @AISecurityInst @AnthropicAI @nyuniversity @SussexUni
The 2025 reward hacking hall of fame award goes to GPT-5.1 for calling the calculator tool to calculate 1+1 on 5% of prod traffic. Because on many prompts using the calculator was superficially rewarded (as a "search") during RL. 🤗
agent safety @openai | previously: @AISecurityInst @AnthropicAI @nyuniversity @SussexUni