PrefBench: Evaluating Zero-Shot LLM Agents in Hidden-Preference Personalized Pricing Negotiations
The paper presents PrefBench, a benchmark for evaluating zero-shot LLM agents in personalized pricing negotiations with hidden buyer preferences. It highlights the challenges faced by sellers in achieving profitable outcomes despite high deal rates. The findings indicate that while LLMs can comply with structured protocols, their profit performance remains subpar compared to simpler heuristics.
- ▪PrefBench is a simulator-based benchmark designed for hidden-preference personalized pricing negotiations.
- ▪The study evaluates zero-shot LLM sellers against heuristic references over 7,500 episodes.
- ▪Despite achieving deal rates above 0.99, the LLMs' average profit is only slightly above random baseline levels.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computer Science and Game Theory arXiv:2605.22855 (cs) [Submitted on 19 May 2026] Title:PrefBench: Evaluating Zero-Shot LLM Agents in Hidden-Preference Personalized Pricing Negotiations Authors:Yingjie Lei View a PDF of the paper titled PrefBench: Evaluating Zero-Shot LLM Agents in Hidden-Preference Personalized Pricing Negotiations, by Yingjie Lei View PDF Abstract:Personalized pricing negotiations are a challenging testbed for LLM agents because successful interaction does not guarantee profitable decision making. A seller may produce valid actions and close many deals while still pricing poorly when buyer willingness to pay and bargaining traits remain hidden.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
