reinforcement learning from human feedback
Reinforcement learning from human feedback (RLHF) is an approach where human insights or evaluations are used to guide and improve the learning of an AI model. This method helps models align their outcomes with human preferences, often enhancing interpretability and usefulness.
- Better Estimation of the Kullback--Leibler Divergence Between Language Models
- Beyond Expectations: Quantile-Guided Alignment for Risk-Calibrated Language Models
- Enhancing Personalized Multi-Turn Dialogue with Curiosity Reward
- Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback
- Fine-Tuning Discrete Diffusion Models with Policy Gradient Methods
- Generative RLHF-V: Learning Principles from Multi-modal Human Preference
- HelpSteer3-Preference: Open Human-Annotated Preference Data across Diverse Tasks and Languages
- Learning “Partner-Aware” Collaborators in Multi-Party Collaboration
- Mitigating Reward Over-optimization in Direct Alignment Algorithms with Importance Sampling
- Provably Efficient Online RLHF with One-Pass Reward Modeling
- ResponseRank: Data-Efficient Reward Modeling through Preference Strength Learning
- Sharp Analysis for KL-Regularized Contextual Bandits and RLHF