AgentRecommender: LLM Agents Enable Customizable Recommender Systems on the User Side

1 Introduction

• Platform objectives such as purchases and watch time can conflict with user interests, and platform systems have been linked to filter bubbles [52] and unfairness [12] [64].Platforms may not actively address such problems when they do not harm their own interests.Source: §1 · p. 1
• User-side recommenders let users specify desired functionality, but users lack direct access to platform databases and other users’ logs.Such systems can rebalance political coverage or make recommendations less limited to popular items.Source: §1 · p. 1
• Conventional user-side recommenders [42] [44] require item labels, which may be unavailable, while preparing a classifier to infer them burdens ordinary users.For example, a site may not disclose view counts or provide political tags.Source: §1 · pp. 1–2
• AgentRecommender uses LLM agents’ world knowledge and, when needed, search or other tools to estimate item attributes and build user-side recommenders without additional data.The authors say it can provide desired functionality in regimes where conventional user-side recommenders do not work.Source: §1 · p. 2

2 Problem Setting

• The provider supplies related items for an accessed item, but its recommendations need not satisfy the user’s desired constraints.The provider system may be Amazon’s “Customers Who Bought This Item Also Bought” or YouTube’s “Up next.”Source: §2 · p. 2
• Users define an arbitrary attribute space and request proportions, balance, or coverage across its categories, while item attributes remain hidden during construction.Examples include genre, era, popularity, price range, and political leaning.Source: §2 · p. 2
• Given oracle access to provider recommendations, natural-language constraints, and budget B, the system outputs Q: I → I^K, which may query the provider at most B times per recommendation.Attributes are used only at evaluation time.Source: §2 · p. 2
• The paper focuses on user-side item-to-item recommendation, while stating that user-to-item recommendation can be constructed in the same manner.The user-to-item extension is outside the simplified setting considered here.Source: §2 · p. 2

3 Proposed Method: AgentRecommender

• The LLM repeatedly expands promising items through the provider’s recommendations, building a candidate pool from which it selects a constraint-satisfying list.The process is described as actively exploring the recommendation network.Source: §3 · pp. 2–3
• Starting from the provider’s recommendations for the accessed item, the system repeats up to B times: choose an unexpanded item and add its provider recommendations to the list.The LLM chooses which item to expand based on the requirements.Source: §3 · p. 3
• The LLM receives item names, estimates attributes using internal knowledge or tools such as search, and chooses items to explore toward the constraints.For example, it may expand a horror movie when the current list is skewed toward action movies.Source: §3 · pp. 3–4
• Because added items are at most B hops from the starting item, they are expected to remain related; the final selection can also be restricted to items sufficiently related to that item.This relation requirement can be included in the prompt.Source: §3 · p. 4
• The approach can also use user-to-item recommendations as its initial list, or query a locally cached recommendation database to reduce online latency and service load.The cache may be built from recommendations obtained during ordinary service use.Source: §3 · p. 4

4 Experiments

• Experiments allowed only black-box access to provider recommender Pprov; its embeddings and BPR-based design could not be exploited, with K = 10, B = 8, and 100 randomly selected users.Evaluation measured the rate of meeting a requirement that every category appear at least τ times and the average number of unique categories when maximizing category variety.Source: §4 · p. 4

4.1 MovieLens 1M

• 68 percent of MovieLens cases met the balanced old/new criterion, and AgentRecommender included 1.6 times as many genres as the provider and iAgent.The old/new criterion required five recommendations from each release-year category.Source: §4.1 · p. 5

4.2 LastFM

• 1.7 times as many genres were included in LastFM recommendation lists as with the provider and iAgent.The evaluation counted distinct genres among the ten recommended artists.Source: §4.2 · p. 5

4.3 Amazon Home and Kitchens

• Four times the provider and iAgent success rate was achieved for balanced low- and high-price recommendations on Amazon Home and Kitchens.The criterion required five products below $19.99 and five at or above that threshold.Source: §4.3 · pp. 5–6

4.4 Budget Sensitivity

• At B = 1, AgentRecommender already greatly outperformed the provider and iAgent; performance generally improved as the budget increased.Amazon was an exception: price estimation from product names was difficult, and success plateaued around 0.6.Source: Figure 2 · p. 6, §4.4 · p. 6

• User-side realization adds desired service functionality while users continue using the service, with examples including user-side search, privacy protection, and watermarking.Users have little power to change a service directly, and platform-locked data and content make building a replacement difficult.Source: §5 · p. 6
• Across movies, music, and e-commerce, AgentRecommender is the first user-side recommender system described here that needs no labels and can discover new, useful items.PrivateWalk, PrivateRank, Consul, and RecCycle require labels, while iAgent reorders presented items and essentially cannot discover new ones.Source: §5 · p. 7
• Unlike user-side methods, most LLM-agent recommender methods assume provider-side operation; user-side realization needs neither provider implementation nor provider endpoints.TEARS conveys preferences in natural language but runs on provider servers.Source: §5 · p. 7
• This study focuses on item fairness and also considers recommendation diversity, while the method can apply to user fairness when the official system is unfair to its user.The passage relates diversity to fairness by treating topic as the sensitive attribute.Source: §5 · p. 7
• Steerable recommenders require provider implementation, whereas arbitrary users can build user-side recommenders even when the official system is not steerable.The paper presents user-side systems as expanding the reach of steerable recommendation.Source: §5 · p. 7

• Across movies, music, and e-commerce, AgentRecommender created recommender systems that satisfied requirements unmet by service providers’ recommenders.Source: §6 · p. 8
• AgentRecommender explores items satisfying user requirements while using the provider’s recommender, drawing on LLM agents’ internal knowledge and search capability instead of additional labels.Source: §6 · p. 8
• The method operates within users’ privileges, avoiding inaccessible provider databases and other users’ log data, and makes recommender systems easier for ordinary users to build.Source: §6 · p. 8