The paper "Linguistic Calibration of Long-Form Generations" by Neil Band, Xuechen Li, Tengyu Ma, and Tatsunori Hashimoto from Stanford University addresses the issue of language models (LMs) confidently hallucinating and making suboptimal decisions when users rely on their outputs. The authors propose a new concept called "linguistic calibration" for long-form generations, which involves the LM verbally conveying the probability that its claims are correct. This is achieved through a training framework that first bootstraps the LM to emit long-form generations with confidence statements using supervised finetuning, followed by reinforcement learning (RL) to optimize the LM's policy for generating calibrated confidence statements. The key contributions of the paper include: 1. **Definition of Linguistic Calibration**: The authors define linguistic calibration for long-form generations as enabling users to make calibrated probabilistic predictions based on the LM's generations. 2. **Training Framework**: They develop a training framework that uses supervised finetuning to bootstrap the LM to emit long-form generations with confidence statements and then optimizes these generations using RL to reward calibrated answers to related questions. 3. **Evaluation**: The authors evaluate their method on the Llama 2.7B model, finding that it significantly improves calibration compared to strong baselines with comparable accuracy. The results also show zero-shot transfer to out-of-domain scientific question-answering datasets and a held-out person biography generation task. The paper demonstrates that long-form generations can be end-to-end calibrated by constructing an objective in the space of the predictions that users make in downstream decision-making tasks. This approach leverages the connection between calibration and decision theory, providing a practical solution to the challenge of LM hallucinations and improving the reliability of LM outputs in real-world applications.The paper "Linguistic Calibration of Long-Form Generations" by Neil Band, Xuechen Li, Tengyu Ma, and Tatsunori Hashimoto from Stanford University addresses the issue of language models (LMs) confidently hallucinating and making suboptimal decisions when users rely on their outputs. The authors propose a new concept called "linguistic calibration" for long-form generations, which involves the LM verbally conveying the probability that its claims are correct. This is achieved through a training framework that first bootstraps the LM to emit long-form generations with confidence statements using supervised finetuning, followed by reinforcement learning (RL) to optimize the LM's policy for generating calibrated confidence statements. The key contributions of the paper include: 1. **Definition of Linguistic Calibration**: The authors define linguistic calibration for long-form generations as enabling users to make calibrated probabilistic predictions based on the LM's generations. 2. **Training Framework**: They develop a training framework that uses supervised finetuning to bootstrap the LM to emit long-form generations with confidence statements and then optimizes these generations using RL to reward calibrated answers to related questions. 3. **Evaluation**: The authors evaluate their method on the Llama 2.7B model, finding that it significantly improves calibration compared to strong baselines with comparable accuracy. The results also show zero-shot transfer to out-of-domain scientific question-answering datasets and a held-out person biography generation task. The paper demonstrates that long-form generations can be end-to-end calibrated by constructing an objective in the space of the predictions that users make in downstream decision-making tasks. This approach leverages the connection between calibration and decision theory, providing a practical solution to the challenge of LM hallucinations and improving the reliability of LM outputs in real-world applications.