Nexus Labs encountered a significant issue during their DPO training for booking agents, where the LLM used as a preference judge exhibited high self-disagreement (up to 28%), leading to a 4-point drop in production accuracy despite seemingly clean training metrics. The team resolved this by implementing a consensus-based judging system using three different LLMs (GPT-4o-mini, Claude Sonnet, and Gemini 2.5 Pro) with a 2-of-3 majority rule, which improved judge consistency and restored production accuracy, albeit at a higher cost and latency. AI
IMPACT Highlights the critical need for robust evaluation and consensus mechanisms in LLM training to prevent noisy signals from degrading model performance.
RANK_REASON The item details a technical issue and solution in an LLM training pipeline, which is a form of research into LLM application and improvement. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
