HARP
A training-free LLM recommendation framework that reformulates recommendation as an agent communication process. Through three operators (symmetric item profiling, profile-level homogeneity gating, and single-pass reflective update), it addresses the asymmetry between user and item profiles and enhances recommendation on a frozen LLM backbone.
Timeline (last 90 days)
HARP recasts LLM recommendation as a four-stage agentic communication process: BUILD, TRANSMIT, UNDERSTAND, SYNC.
HARP is a training-free, three-operator extension framework attached to a frozen LLM backbone that does not update model parameters.
HARP's three core operators respectively correspond to: rule-extracted symmetric item profiles (solving the one-way BUILD), channel-decomposable profile-level homogeneity H(φ_u, φ_i) consumed at retrieval time (breaking the prompt trap), and a single-pass reflection update (completing the SYNC stage).
On four public benchmarks (Yelp, Books, MovieTV, Goodreads), HARP improves Hit@1 by 6.9% to 24.0% over the strongest published baseline.
In HARP's experiments, the Holm-corrected p-values on every dataset were all less than 1e-9.
Ablation experiments and per-case attribution studies show that HARP's three operators contribute complementary gains, with the retrieval-side gating being the main source.
HARP's reflection process introduces a trade-off and may introduce extra overhead or noise in some scenarios.
HARP's code is open-sourced on GitHub (https://github.com/Jove-Zhen/HARP).
HARP's rule-extracted item profile operator requires no additional LLM calls, generating self-descriptions for items at zero cost.
HARP's homogeneity function H(φ_u, φ_i) measures the degree of match between user profiles and item profiles and gates candidate items at the retrieval stage, letting profile signals participate directly in recall.
2 more timeline events
All Facts (11)
HARP's single-pass reflection update achieves in-context profile evolution without any gradient steps.
50%UnverifiedHARP is a training-free, three-operator extension framework attached to a frozen LLM backbone that does not update model parameters.
50%UnverifiedHARP's three core operators respectively correspond to: rule-extracted symmetric item profiles (solving the one-way BUILD), channel-decomposable profile-level homogeneity H(φ_u, φ_i) consumed at retrieval time (breaking the prompt trap), and a single-pass reflection update (completing the SYNC stage).
50%UnverifiedOn four public benchmarks (Yelp, Books, MovieTV, Goodreads), HARP improves Hit@1 by 6.9% to 24.0% over the strongest published baseline.
50%UnverifiedIn HARP's experiments, the Holm-corrected p-values on every dataset were all less than 1e-9.
50%UnverifiedAblation experiments and per-case attribution studies show that HARP's three operators contribute complementary gains, with the retrieval-side gating being the main source.
50%UnverifiedHARP's reflection process introduces a trade-off and may introduce extra overhead or noise in some scenarios.
50%UnverifiedHARP recasts LLM recommendation as a four-stage agentic communication process: BUILD, TRANSMIT, UNDERSTAND, SYNC.
50%UnverifiedHARP's rule-extracted item profile operator requires no additional LLM calls, generating self-descriptions for items at zero cost.
50%UnverifiedHARP's homogeneity function H(φ_u, φ_i) measures the degree of match between user profiles and item profiles and gates candidate items at the retrieval stage, letting profile signals participate directly in recall.
50%UnverifiedHARP's code is open-sourced on GitHub (https://github.com/Jove-Zhen/HARP).
50%