[KongchangAI]
PaperHomophily-aware Agentic Recommendation via dual Profiles

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)

Nov 4

HARP recasts LLM recommendation as a four-stage agentic communication process: BUILD, TRANSMIT, UNDERSTAND, SYNC.

Unverified50%
Nov 4

HARP is a training-free, three-operator extension framework attached to a frozen LLM backbone that does not update model parameters.

Unverified50%
Nov 4

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).

Unverified50%
Nov 4

On four public benchmarks (Yelp, Books, MovieTV, Goodreads), HARP improves Hit@1 by 6.9% to 24.0% over the strongest published baseline.

Unverified50%
Nov 4

In HARP's experiments, the Holm-corrected p-values on every dataset were all less than 1e-9.

Unverified50%
Nov 4

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.

Unverified50%
Nov 4

HARP's reflection process introduces a trade-off and may introduce extra overhead or noise in some scenarios.

Unverified50%
Nov 4

HARP's code is open-sourced on GitHub (https://github.com/Jove-Zhen/HARP).

Unverified50%
Nov 4

HARP's rule-extracted item profile operator requires no additional LLM calls, generating self-descriptions for items at zero cost.

Unverified50%
Nov 4

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.

Unverified50%

2 more timeline events

All Facts (11)

Unverified

HARP's single-pass reflection update achieves in-context profile evolution without any gradient steps.

50%
Unverified

HARP is a training-free, three-operator extension framework attached to a frozen LLM backbone that does not update model parameters.

50%
Unverified

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).

50%
Unverified

On four public benchmarks (Yelp, Books, MovieTV, Goodreads), HARP improves Hit@1 by 6.9% to 24.0% over the strongest published baseline.

50%
Unverified

In HARP's experiments, the Holm-corrected p-values on every dataset were all less than 1e-9.

50%
Unverified

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.

50%
Unverified

HARP's reflection process introduces a trade-off and may introduce extra overhead or noise in some scenarios.

50%
Unverified

HARP recasts LLM recommendation as a four-stage agentic communication process: BUILD, TRANSMIT, UNDERSTAND, SYNC.

50%
Unverified

HARP's rule-extracted item profile operator requires no additional LLM calls, generating self-descriptions for items at zero cost.

50%
Unverified

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.

50%
Unverified

HARP's code is open-sourced on GitHub (https://github.com/Jove-Zhen/HARP).

50%

Source Articles