ID-TAP personalizes LLM agents from binary response comparisons
Original title: Bridging Exploration and Persona: Information-Directed Dueling for Test-Time Agent Personalization
Within a fixed ten-round comparison budget, ID-TAP led aggregate results on seven user-conditioned generation benchmarks.
ID-TAP is an information-directed neural dueling-bandit method for test-time persona selection. It treats binary comparisons between candidate responses as the online signal and scores a large discrete persona space with a query-conditioned neural utility model plus a lightweight Bradley-Terry explorer. Comparison evidence transfers across related queries through similarity-weighted warm starts. On seven user-conditioned generation benchmarks, under reference- and model-based preference simulators, it had the strongest aggregate performance among evaluated baselines inside a ten-round budget.
- Persona choice uses binary comparisons between candidate responses.
- Evaluation covered seven user-conditioned generation benchmarks.
- The comparison budget was fixed at ten rounds.
- A Bradley-Terry explorer scores hypothetical challengers with efficient updates.
From the original article
1 min read at huggingface.co
Effective test-time personalization requires LLM agents to adapt to query-dependent user preferences from lightweight interaction. We study persona selection with binary comparisons between candidate responses as the online signal. We present ID-TAP, a scalable, IDS-inspired neural dueling-bandit framework for large discrete persona spaces.
By Jie Mao, Pingchen Lu, Xiangyi Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.