Reproduce “Naturally Occurring Feedback is Common, Extractable and Useful” 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
Shachar Don-Yehiya, Leshem Choshen, Omri Abend
Researchers demonstrated that naturally occurring explicit feedback from user interactions with large language models is prevalent and extractable. The automatically collected feedback dataset, when used for finetuning and preference training, improved LLM alignment with human preferences, achieving human evaluation win rates up to 81.5% against pretrained models.
Key insights
- Naturally occurring explicit feedback is common, found in approximately 30% of multi-turn human-LLM conversations, and its prevalence has significantly increased with newer, more sophisticated LLMs.
- An LLM-based method can successfully extract a large-scale dataset of natural feedback, demonstrating its feasibility despite moderate extraction precision.
- A detailed, specific taxonomy for feedback categories is crucial for effective automatic extraction, outperforming more general or limited categorization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.