DxHF: Providing High-Quality Human Feedback for LLM Alignment with Interactive Decomposition
summary
DxHF decomposes long responses into decomposed claims, links similar statements, encodes claim relevance, and keeps raw text visible; an LLM-driven workflow performs decomposition, ranking, and linking. A crowdsourcing study (N=160) reports a 5% accuracy gain on preference labels and an average 18s slower judgment time, with the biggest improvement when annotator certainty is low. Simulated annotators show decomposition helps under low rationality; the UI reduces cognitive load but may introduce decomposition-induced bias.
AI-written summary · about
chapters (6)
- 0:00Introduction to alignment
- 1:50Decomposition principle for feedback
- 3:17DXHF interface design
- 5:45Evaluation studies
- 8:59User feedback and trade-offs
- 10:02Takeaways and future work
related talks
33mDesigning LLM Applications for Massive Datasets
15mLLMs for Data-Driven Decision Making in Product Development
47mGive Your LLMs a Left Brain
56mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
25mEngineering Better Evals: Scalable LLM Evaluation Pipelines That Work
22mPractical Strategies for Implementing Domain-Specific LLMs
39mBeyond the Prompt: Evaluating, Testing, and Securing LLM Applications
38mThe Art of Compression: Crafting Insightful Summaries with LLMs