HARP: A Training-Free Dual-Profile Agentic Recommendation Framework with Up to 24% Hit@1 Improvement

HARP reformulates LLM recommendation as four-stage agent communication, using three training-free operators to solve profile asymmetry, improving Hit@1 by 6.9%-24.0%.
HARP addresses the asymmetry between user and item profiles in LLM-based recommender systems by reformulating recommendation as four agent communication stages: BUILD, TRANSMIT, UNDERSTAND, and SYNC, and identifies three challenges: missing item-side self-descriptions, profile signals trapped in re-ranking prompts, and frozen user profiles. It proposes three training-free operators: rule-extracted symmetric item profiles, retrieval-side homophily gating, and single-pass reflective update, attached to a frozen LLM backbone. On Yelp, Books, MovieTV, and Goodreads, Hit@1 improves by 6.9% to 24.0% over the strongest baseline, with statistical significance (Holm-corrected p<1e-9). Ablations show retrieval-side gating contributes the most, while reflection involves trade-offs. Code is open-sourced for easy reproduction and deployment.
The Communication Dilemma in Recommender Systems: The Asymmetry Between Users and Items
In recommendation systems based on large language models (LLMs), memory modules are often viewed as user-level agents embedded in memory-augmented backbones. However, there is a natural asymmetry between users and items: only users have structured profiles, while items lack comparable self-descriptions; user profiles are fed in only through bandwidth-limited re-ranking prompts, not at the retrieval stage; and profiles are frozen after warm-up, lacking a principled update rule.
According to the research paper, this asymmetry constitutes three unresolved communication challenges: the item side never produces comparable self-descriptions (one-way BUILD); profile-level signals enter only bandwidth-limited re-ranking prompts rather than the retrieval stage, causing a "prompt trap" (prompt-trapped UNDERSTAND); and user descriptions are frozen after warm-up with no principled update rule (missing SYNC).

Bandwidth-limited prompt refers to the limited context window capacity of large language models: when user profile information is input as a prompt, it can only be presented with very few tokens and cannot fully convey all the details of user preferences. Therefore, if profile signals are placed only in re-ranking prompts, the information is heavily compressed and cannot influence the retrieval of candidate items, causing the model's understanding of the user to be "trapped" at the prompt level.
HARP's Four Communication Stages and Three Operators
The researchers reformulate LLM recommendation as an agent communication process with four stages: BUILD (construct self-description), TRANSMIT (transmit along the collaboration graph), UNDERSTAND (understand the other party), and SYNC (synchronize from the next interaction). HARP (Homophily-aware Agentic Recommendation via dual Profiles) is precisely a training-free, three-operator extension targeting the missing stages, attached to a frozen LLM backbone.
The three core operators correspond to the above challenges:
- Rule-extracted symmetric item profiles: Generate self-descriptions for items at zero cost without additional LLM calls, solving the one-way BUILD problem.
- Channel-decomposable profile-level homophily H(φ_u, φ_i): Consumed at retrieval time rather than stuffed into re-ranking prompts, breaking through the bandwidth limit and solving the prompt trap.
- Single-pass reflective update: Enables in-context profile evolution without any gradient steps, filling in the SYNC stage.
This design cleverly bypasses fine-tuning costs while allowing profile signals to take effect at the retrieval stage, avoiding competition with context tokens.
Homophily in social networks and recommender systems refers to the phenomenon that similar users tend to interact with similar items. The homophily function ( H(\phi_u, \phi_i) ) in HARP measures the degree of match between user profile ( \phi_u ) and item profile ( \phi_i ), thereby gating candidate items at the retrieval stage so that profile signals directly participate in recall rather than being limited to re-ranking.
Training-free means that no parameters of the large language model are updated and no fine-tuning is performed; recommendation capability is enhanced only through inference-time operator extensions (such as rule extraction and in-context reflection), reducing deployment costs and the risk of catastrophic forgetting.
Experimental Results: Significant Hit@1 Improvement
On four public benchmarks (Yelp, Books, MovieTV, Goodreads), HARP improves Hit@1 by 6.9% to 24.0% over the strongest published baseline, with Holm-corrected p-values below 1e-9 on every dataset, indicating extremely strong statistical significance.
Ablation experiments and case-by-case attribution studies show that the three operators contribute complementary gains, with retrieval-side gating being the main source. The researchers also report the trade-offs introduced by the reflection process.
Interestingly, HARP is completely training-free, and the code is open-sourced (https://github.com/Jove-Zhen/HARP), which is very friendly for practical deployment.
Hit@1 is a commonly used evaluation metric in recommender systems, representing the proportion of samples where the top-ranked item in the model's ranked list is exactly the item the user actually interacted with. Holm correction is a multiple-hypothesis testing correction method used to control the family-wise error rate; when p-values remain below 1e-9 after Holm correction, it indicates that the improvements on each dataset are not accidental.
Multi-Source Perspectives and Limitations
Currently, this work is mainly disseminated through the paper and code repository, and large-scale independent reproduction has not yet been seen. Analysis from a single source points out that retrieval-side gating contributes the most, but the reflection process involves trade-offs and may introduce additional overhead or noise in certain scenarios. More community validation is needed in the future to evaluate its generalization ability.
Overall, HARP introduces the agent communication framework into LLM recommendation, solving the profile asymmetry problem with lightweight operators and providing a new idea for training-free recommender systems.
One-line Summary
HARP reformulates LLM recommendation as four-stage agent communication, using three training-free operators to solve profile asymmetry, improving Hit@1 by 6.9%-24.0%.
Paragraph Summary
HARP addresses the asymmetry between user and item profiles in LLM-based recommender systems by reformulating recommendation as four agent communication stages: BUILD, TRANSMIT, UNDERSTAND, and SYNC, and identifies three challenges: missing item-side self-descriptions, profile signals trapped in re-ranking prompts, and frozen user profiles. To this end, it proposes three training-free operators: rule-extracted symmetric item profiles, retrieval-side homophily gating, and single-pass reflective update, attached to a frozen LLM backbone. On four benchmarks—Yelp, Books, MovieTV, and Goodreads—Hit@1 improves by 6.9% to 24.0% over the strongest baseline, with statistical significance (Holm-corrected p<1e-9). Ablations show that retrieval-side gating contributes the most, while the reflection process involves trade-offs. The code is open-sourced for easy reproduction and deployment.
Related articles

Abnormal Error Rate Spike in Claude Mythos 5.2 and Fable 5.2
Claude Mythos 5.2 and Claude Fable 5.2 show an abnormal spike in request error rates, marked as investigating on the official status page. This article covers the incident overview, practical impact on developers, and recommended responses.

ReAct Runtime Context Trimming: Why Claude Code Acts at 70%
A deep dive into ReAct runtime context trimming: why Claude Code compresses at ~70% usage, why tool calls dominate, why messages must be dropped in pairs, and why codebases don't need RAG.

AI Agent vs LLM: How Cursor and Claude Code Actually Work
An in-depth analysis of the essential difference between AI Agents and LLMs: LLMs generate code suggestions, while Agents modify files and run tests via tool calls. Using a Spring Boot bug fix as an example, it breaks down the workflows, permission control, and verification mechanisms of Cursor and Claude Code.