ARES: Alternating Reinforcement Learning and Supervised Fine-Tuning for Enhanced Multi-Modal Chain-of-Thought Reasoning Through Diverse AI…
Averaging log-likelihoods in direct alignmentby Nathan Grinsztajn, Yannis Flet-Berliac, Mohammad Gheshlaghi Azar, Florian Strub, Bill…
The Perils of Optimizing Learned Reward Functions: Low Training Error Does Not Guarantee Low Regretby…
SAIL: Self-Improving Efficient Online Alignment of Large Language Modelsby Mucong Ding, Souradip Chakraborty, Vibhu Agrawal,…
Robust Reinforcement Learning from Corrupted Human Feedbackby Alexander Bukharin, Ilgee Hong, Haoming Jiang, Zichong Li,…
Self-play with Execution Feedback: Improving Instruction-following Capabilities of Large Language Modelsby Guanting Dong, Keming Lu,…
Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Expertsby Haoxiang Wang, Wei Xiong, Tengyang Xie, Han…
Is poisoning a real threat to LLM alignment? Maybe more so than you thinkby Pankayaraj…
WPO: Enhancing RLHF with Weighted Preference Optimizationby Wenxuan Zhou, Ravi Agrawal, Shujian Zhang, Sathish Reddy…
Measuring memorization in RLHF for code completionby Aneesh Pappu, Billy Porter, Ilia Shumailov, Jamie HayesFirst…