Research map / Multi-agent systems and games / LLM agents
HARP: Training-Free Dual-Profile Agentic Communication for LLM-Based Recommendation
Zhen Tao, Xun Zhou, Xuhui Chen, Ziyue Qiao, Qingqiang Sun
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 HARP, a training-free dual-profile agentic communication framework for LLM-based recommendation that gives items as well as users structured profiles and lets them communicate.
Problem
LLM-based recommendation systems treat users and items asymmetrically: only users get a structured memory profile, that profile is consumed within a bandwidth-limited reranker prompt, and it is frozen after warmup.
Approach
Recasts LLM-based recommendation as agentic communication between user-level and item-level agents, each with a structured profile, so both sides are symmetric participants; the approach is training-free and avoids the frozen, prompt-bottlenecked profile of prior memory-augmented methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add retrieval-side homophily gate with both profiles (ablation) | HARP · Yelp 1,000-user subsplit | Hit@1: .578 -> .635 | Improved |
| inject user profile into reranker prompt only (ablation) | Memory backbone · Yelp 1,000-user subsplit | Hit@1: .580 -> .578 | No clear effect |
| add item profile without retrieval gate (prompt injection) (ablation) | Memory backbone · Yelp 1,000-user subsplit | Hit@1: .580 -> .580 | No clear effect |
| replace retrieval-side gate with prompt-side injection (ablation) | HARP · Yelp 1,000-user subsplit | Hit@1: .635 -> .580 | Hurt |
| add prompt injection and reflective Sync on top of gate (ablation) | HARP · Yelp 1,000-user subsplit | Hit@1: .635 -> .643 | Improved |
| add prompt injection and Sync (ablation) | HARP · Books subsplit | Hit@1: -0.4 pp | Hurt |
| add prompt injection and Sync (ablation) | HARP · Goodreads subsplit | Hit@1: +1.3 pp | Improved |
| add HARP to memory backbone (ablation) | HARP · Yelp/MovieTV/Books/Goodreads subsplits | Hit@1: +6.3/+3.4/+1.9/+1.9 pp | Improved |
| use HARP instead of MemRec | HARP · Yelp | Hit@1: .487 -> .604 | Improved |
Closest papers by meaning
- AgentRecBench: Benchmarking LLM Agent-based Personalized Recommender Systems (NeurIPS 2025 · similarity 0.90)
- AgentSelect: Benchmark for Narrative Query-to-Agent Recommendation (ICML 2026 · similarity 0.88)
- ConvRec-R1: Training LLM-based Conversational Recommender Systems with Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Re2LLM: Reflective Reinforcement Large Language Model for Session-based Recommendation (AAAI 2025 · similarity 0.88)
- Objective Shaping with Hard Negatives: Windowed Partial AUC Optimization for RL-based LLM Recommenders (NeurIPS 2026 · similarity 0.88)
- Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · similarity 0.87)
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.
