How STAR Enables Training‑Free Recommendations with Large Language Models

1. Introduction

Recent advances in large language models (LLMs) have motivated their use in recommendation systems (LLM4Rec). Existing approaches fall into three main categories:

LLM as a Feature Encoder

LLMs generate semantic embeddings from item metadata (title, description, category, brand, price, etc.) and user profiles. These embeddings are either discretized (e.g., vector quantization) and fed to downstream generators, used to initialise sequence‑model embeddings, or directly employed to compute item‑user similarity. While effective, such pipelines require additional training and reduce model generality.

LLM as a Scoring and Ranking Function

Natural‑language prompts are used to let LLMs infer user preferences from interaction histories. Pure LLM‑only solutions often lag behind models fine‑tuned on collaborative user‑item data, leading to hybrid methods that combine interaction‑based fine‑tuning with LLM semantic understanding—at the cost of extra training.

LLM as a Ranker for Information Retrieval

LLMs serve as zero‑shot rankers for document retrieval, sometimes outperforming supervised cross‑encoders. Prompt designs are classified as point‑wise (evaluate each document independently), pair‑wise (compare two items), or list‑wise (compare multiple items), each with distinct computational trade‑offs.

Most LLM4Rec techniques still rely on downstream fine‑tuning, incurring training overhead. The authors therefore propose STAR (Simple Training‑free Approach for Recommendation), a two‑stage framework—retrieval and ranking—that operates without any additional model training.

2. Method

The overall STAR architecture is illustrated below.

2.1 Retrieval Stage

The retrieval component scores unseen items for a user based on the user's historical behavior sequence, integrating semantic and collaborative signals without any fine‑tuning.

2.1.1 Semantic Relation

Item textual fields (title, description, category, brand, sales rank, price, etc.) are fed to an LLM via a prompt. The LLM’s embedding API returns a dense vector for each item. All item embeddings are pre‑computed offline. During inference, the cosine similarity between each item in the user’s recent sequence (e.g., items #1‑#3) and a candidate item (e.g., item #4) is calculated, yielding a semantic similarity score s.

2.1.2 Collaborative Relation

A binary user‑item interaction matrix M (rows = items, columns = users, entry = 1 if interaction occurred) is constructed from historical data. Item‑item cosine similarity is computed on the rows of M, producing a collaborative co‑occurrence score c for every item pair.

2.1.3 Score Fusion Rule

For each candidate item, STAR combines four signals:

Semantic similarity s Collaborative co‑occurrence c User rating score r (e.g., explicit rating of previously watched items)

Time‑decay factor t, an exponential decay based on the position of the historical item in the sequence (more recent items receive larger weight)

The final retrieval score is a weighted sum: score = λ₁·s + λ₂·c + λ₃·r·t where λ₁, λ₂, λ₃ are hyper‑parameters that balance the contributions. The top‑K items with the highest scores are passed to the ranking stage.

2.2 Ranking Stage

2.2.1 Ranking Strategies

STAR evaluates three ranking strategies on the retrieved list:

Point‑wise: each item is scored independently using a prompt that incorporates the user’s sequence; ties are broken by the retrieval score.

Pair‑wise: a sliding window of size 2 compares adjacent items; if the lower‑scored item appears before a higher‑scored one, their positions are swapped.

List‑wise: a sliding window of size W (e.g., 5) moves with stride S, comparing all items within the window jointly. Pair‑wise is a special case with W=2.

2.2.2 Item Prompt Construction

The ranking prompt concatenates the item’s metadata with two additional signals:

Popularity information: total interaction count of the item in the dataset (e.g., “Number of users who purchased this item: 1234”).

Co‑occurrence information: number of users who interacted with both the candidate item and a specific historical item (e.g., “Users who bought both this item and historical item #1: 57”).

3. Experiments

3.1 Overall Performance

STAR is compared against several baselines on standard recommendation benchmarks. The figure below shows that STAR achieves competitive or superior hit‑rate and NDCG metrics while requiring no additional model training.

3.2 Hyper‑parameter Analysis

Retrieval‑stage hyper‑parameters (weights λ, top‑K size) are varied to assess sensitivity. Results indicate that moderate values of λ₁ and λ₂ balance semantic and collaborative signals effectively.

Ranking‑stage experiments explore different window sizes W and strides S. Larger windows improve list‑wise performance up to a point, after which computational cost outweighs gains.

3.3 Ablation Studies

Two ablations are performed:

Removing the rating term r from the retrieval score to quantify its contribution.

Omitting popularity and co‑occurrence information from the ranking prompt to evaluate their impact.