Bridging Exploration and Persona: Information-Directed Dueling for Test-Time Agent Personalization
We present ID-TAP, a scalable, IDS-inspired neural dueling-bandit framework for large discrete persona spaces.
ProofPaper ↗
Key points
- 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.
- ID-TAP combines a query-conditioned neural utility model with a lightweight Bradley-Terry explorer, and transfers comparison evidence across related queries through similarity-weighted warm starts.
- These results support information-directed challenger selection as an effective design for test-time persona adaptation.
Sources (1)
- [1]Bridging Exploration and Persona: Information-Directed Dueling for Test-Time Agent PersonalizationarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 02:40 AM
We present ID-TAP, a scalable, IDS-inspired neural dueling-bandit framework for large discrete persona spaces.
Effective test-time personalization requires LLM agents to adapt to query-dependent user preferences from lightweight interaction.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 16, 2026[Research] 2,788 tokens/s on a single RTX 5090 with a diffusion LLM — Introducing ORL (fully open source)
- Oct 15, 2026open source framework that makes your AI agents generate UI elements
- Oct 15, 2026Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint
- Oct 15, 2026NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
- Oct 15, 2026PriorityFEC: Deadline-Aware Streaming FEC for KV-Cache Handoff in Edge Multi-Agent Systems
- Oct 15, 2026Two Sample Sizes in Test-Time LLM Ensembling: Buffered ERM and Minimax Theory