Reproduce “Aligning Dialogue Agents with Global Feedback via Large Language Model Multimodal Reward Decomposition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dong Won Lee, Hae Won Park, Cynthia Breazeal, Louis-Philippe Morency
Authors introduce LLM-GELI, which uses a frozen pretrained language model to assign pseudo-rewards to agent turns from dialogue and a session rating; its multimodal version also receives language descriptions of listener behavior. A text-only reward model learns these labels and guides PPO fine-tuning. On CANDOR, multimodal LLM-GELI gets 21.67 ± 1.89% Interestingness selections versus GELI’s 19.33 ± 1.70%, a difference not established as significant.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.