Bridging Exploration and Persona: Information-Directed Dueling for Test-Time Agent Personalization
ID-TAP posted the strongest aggregate results among tested baselines with just ten rounds of preference comparisons.
The framework adapts an LLM agent’s persona by comparing pairs of candidate responses for a given query. It combines a neural utility model with a lightweight Bradley–Terry explorer to select challengers, carrying comparison evidence across similar queries. The evaluation covered seven user-conditioned generation benchmarks using reference- and model-based preference simulators. The reported advantage reflects simulated preferences, rather than a demonstrated result with live users.
HF Daily Papers' note
The framework adapts an LLM agent’s persona by comparing pairs of candidate responses for a given query. It combines a neural utility model with a lightweight Bradley–Terry explorer to select challengers, carrying comparison evidence across similar queries. The evaluation covered seven user-conditioned generation benchmarks using reference- and model-based preference simulators. The reported advantage reflects simulated preferences, rather than a demonstrated result with live users.
HF Daily Papers' note
score 4