CarryOnBench: Testing LLM Utility Recovery After Safety Refusals

CarryOnBench has been launched by researchers as the inaugural interactive benchmark designed to evaluate if large language models (LLMs) can regain utility when well-meaning users clarify their intentions after an initial safety denial. The study began with 398 queries that appeared harmful but had benign motives, leading to the simulation of 5,970 dialogues by altering user follow-up sequences. Fourteen models were assessed based on intent-aligned utility and safety, resulting in 1,866 unique conversation flows ranging from 4 to 12 turns, which produced 23,880 responses. A novel metric, Ben-Util, employs atomic checklist items to measure how effectively each response meets the user's benign information requirements. Initially, models satisfy only 10.5–37.6% of these needs, revealing a disparity between safety and helpfulness in multi-turn exchanges.