HARP lifts Hit@1 in LLM-based recommendation by 6.9%-24.0%
Original title: HARP: Training-Free Dual-Profile Agentic Communication for LLM-Based Recommendation
Training-free dual-profile agentic communication framework improves recommendation accuracy over strong baselines with p < 1e-9 on four datasets.
HARP recasts LLM-based recommendation as agentic communication with four stages: BUILD, TRANSMIT, UNDERSTAND, and SYNC. It addresses one-sided item profiles, prompt-trapped signals, and frozen user descriptions. HARP adds three operators to a frozen LLM backbone: symmetric item profile, homophily-aware profile channel, and reflective update. On Yelp, Books, MovieTV, and Goodreads, HARP lifts Hit@1 over the strongest baseline by 6.9%-24.0% with family-wise Holm-corrected p < 1e-9.
- Three-operator addition to frozen LLM backbone
- Rule-extracted symmetric item profile with zero extra LLM calls
- Profile-level homophily H(phi_u, phi_i) consumed at retrieval
- Single-pass reflective update for in-context profile evolution
- Family-wise Holm-corrected p < 1e-9 on every dataset
From the original article
1 min read at arxiv.orgIn LLM-based recommendation, memory acts as a user-level agent in a memory-augmented backbone. Yet users and items remain asymmetric: only the user is given a structured profile; it enters a bandwidth-limited reranker prompt; and it is frozen after warmup.
By Zhen Tao · arxiv.org
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.