A new evaluation arena called Vendster has been developed to assess how the quality of forecasting tools impacts the negotiation behaviors of large language model (LLM) agents. The arena measures the effect of a forecaster's bias and prediction-interval width on buyer-vendor discount negotiations. Experiments using the Vendster-M5 instance with frontier models like Claude Opus-4.6 and GPT 5.6 "Sol" demonstrated that model-generated forecasts can increase sales target attainment by up to 17 percentage points, though standard accuracy metrics do not reliably predict negotiation outcomes. The study found that under-prediction bias significantly improves agent outcomes, while interval width has no significant effect. AI
IMPACT Highlights the importance of forecast bias and agent reactivity over raw accuracy for LLM agents in decision-making tasks.
RANK_REASON Research paper detailing a new evaluation arena and experimental findings. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
