Feedback Descent: Open-Ended Text Optimization via Pairwise Comparison

The recent publication of 'Feedback Descent' on arXiv presents a groundbreaking approach to optimizing text artifacts, such as prompts and molecules, through structured textual feedback. Unlike traditional methods that rely on scalar rewards, Feedback Descent utilizes detailed critiques, significantly widening the information bottleneck in preference learning. This innovative framework allows for directed optimization in text space, enabling targeted edits without altering model weights. Evaluations across diverse domains demonstrate its superiority over state-of-the-art methods, including GEPA and GRPO, particularly in the DOCKSTRING molecule discovery benchmark, where it identified novel drug-like molecules that exceed the 99.9th percentile of a database containing over 260,000 compounds. This advancement not only showcases the potential of structured feedback in AI but also highlights a significant leap in the capabilities of text optimization techniques.