Research map / Multi-agent systems and games / LLM agents
Orion: Steering Personalized Web Agents via Global-Micro Profiling and Adaptive Intent Tracking
Die Hu, Jingguo Ge, Weitao Tang, He Kong, Liangxiong Li, Bingzhen Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Orion, a personalized LLM-based web agent that models user intent via global-micro profiling and adaptive intent tracking.
Problem
LLM-based web agents have limited personalization because user intent is ambiguous and dynamic, making it hard to model diverse interests and track intent changes over time.
Approach
Orion combines global-micro user profiling with adaptive intent tracking so that a web agent can model diverse user interests and follow evolving intent. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add online GRPO RL after SFT (ablation) | Orion · Ablation avg over three tasks | result accuracy (%): 34.7 -> 44.3 | Improved |
| add global-micro profile instead of single profile (ablation) | Orion · Ablation avg over three tasks | result accuracy (%): 30.8 -> 44.3 | Improved |
| add dynamic profile update during interaction (ablation) | Orion · Ablation avg over three tasks | result accuracy (%): 31.6 -> 44.3 | Improved |
| add proactive disambiguation action (ablation) | Orion · Ablation avg over three tasks | result accuracy (%): 32.1 -> 44.3 | Improved |
| add proactive disambiguation action (ablation) | Orion · Ablation avg over three tasks | average steps: 6.4 -> 3.83 | Improved |
| use GRPO instead of PPO for online RL | Qwen2.5-7B-Instruct SFT · Personalized Recommendation | result accuracy (%): 10.3 -> 19.8 | Improved |
| add online RL (PPO) after SFT (ablation) | Qwen2.5-7B-Instruct SFT · Personalized Recommendation | result accuracy (%): 8.2 -> 10.3 | Improved |
| use PUMA profile prompting vs ChatShop on GPT-4.1 | GPT-4.1 · Product Search | result accuracy (%): 38.0 -> 42.8 | Improved |
| increase clustering threshold alpha away from 15 (ablation) | Orion · sensitivity analysis | result accuracy | Mixed |
| increase context size lc beyond 6 (ablation) | Orion · sensitivity analysis | result accuracy | No clear effect |
Closest papers by meaning
- Persona2Web: Benchmarking Personalized Web Agents for Contextual Reasoning with User History (ICML 2026 · similarity 0.89)
- Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 2025 · similarity 0.85)
- WebPilot: A Versatile and Autonomous Multi-Agent System for Web Task Execution with Strategic Exploration (AAAI 2025 · similarity 0.85)
- Enhancing Personalized Multi-Turn Dialogue with Curiosity Reward (NeurIPS 2025 · similarity 0.85)
- Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.85)
- REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
