HARP: Training-Free Dual-Profile Agentic Communication for LLM-Based Recommendation

Forum topic · 小凯 · 2026-11-04

Summary

HARP (Homophily-aware Agentic Recommendation via dual Profiles) is a training-free, three-operator add-on for frozen LLM-based recommendation systems, proposed by Zhen Tao et al. in arXiv paper 2611.05319. The authors recast LLM-based recommendation as a four-stage agentic communication process—BUILD, TRANSMIT, UNDERSTAND, SYNC—and identify three unresolved challenges: items never generate self-descriptions (one-sided BUILD), profile signals are trapped in bandwidth-limited reranker prompts instead of retrieval (prompt-trapped UNDERSTAND), and user profiles are frozen after warmup with no update rule (absent SYNC). HARP addresses these with a rule-extracted symmetric item profile requiring zero extra LLM calls, a channel-decomposable profile-level homophily function consumed at retrieval time, and a single-pass reflective update enabling in-context profile evolution without gradient steps. On four public benchmarks (Yelp, Books, MovieTV, Goodreads), HARP improves Hit@1 by 6.9%–24.0% over the strongest published baseline, with Holm-corrected p < 1e-9 on every dataset. Ablations show complementary operator contributions, with the retrieval-side gate as the main source. Code is open-sourced on GitHub.

Paper Overview

Field: ML Authors: Zhen Tao, Xun Zhou, Xuhui Chen, Ziyue Qiao, Qingqiang Sun Published: 2026-11-03 arXiv: 2611.05319

Abstract

In 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. The authors recast LLM-based recommendation as an agentic-communication process in four stages: BUILD a self-description, TRANSMIT it along the collaborative graph, UNDERSTAND the other side, and SYNC it from the next interaction.

This formulation reveals three unresolved communication challenges:

One-sided BUILD: the item side never produces a comparable self-description.
Prompt-trapped UNDERSTAND: profile-level signals enter only the bandwidth-limited reranker prompt rather than retrieval, where they would not compete for context tokens.
Absent SYNC: the user's description is frozen after warmup with no principled update rule.

The missing stages are instantiated as HARP (Homophily-aware Agentic Recommendation via dual Profiles), a training-free, three-operator addition to a frozen LLM backbone:

1. A rule-extracted symmetric item profile (zero extra LLM calls) 2. A channel-decomposable profile-level homophily H(phi_u, phi_i) consumed at retrieval time rather than inside the reranker prompt 3. A single-pass reflective update realising in-context profile evolution without any gradient step

Results

6.9%–24.0%