CTRL-Rec: Controlling Recommender Systems With Natural Language

TL;DR

CTRL-Rec tackles the lack of fine-grained user control in recommender systems by injecting natural language requests at retrieval time. It introduces a distillation-based approach where an LLM-specified value $v(u,i,r)$ is approximated by a dot product $v(u,i,r) \approx f(u,r)^T g(i)$, enabling real-time control via a dual-encoder and a scoring function $score_{CTRL-Rec}(u,i,r) = score_{base}(u,i) + w_{control} v(u,i,r)$. The method achieves scalable control with only a single LLM embedding per request, demonstrated on MovieLens with diverse requests and validated in a Letterboxd user study. The results show improved controllability and satisfaction without sacrificing engagement, highlighting practical potential for balancing explicit user preferences with engagement-driven recommendations.

Abstract

When users are dissatisfied with recommendations from a recommender system, they often lack fine-grained controls for changing them. Large language models (LLMs) offer a solution by allowing users to guide their recommendations through natural language requests (e.g., "I want to see respectful posts with a different perspective than mine"). We propose a method, CTRL-Rec, that allows for natural language control of traditional recommender systems in real-time with computational efficiency. Specifically, at training time, we use an LLM to simulate whether users would approve of items based on their language requests, and we train embedding models that approximate such simulated judgments. We then integrate these user-request-based predictions into the standard weighting of signals that traditional recommender systems optimize. At deployment time, we require only a single LLM embedding computation per user request, allowing for real-time control of recommendations. In experiments with the MovieLens dataset, our method consistently allows for fine-grained control across a diversity of requests. In a study with 19 Letterboxd users, we find that CTRL-Rec was positively received by users and significantly enhanced users' sense of control and satisfaction with recommendations compared to traditional controls.

Figures (13)

• Figure 1: Our approach differs from standard engagement recommenders by allowing to optimize for both revealed preferences (through engagement signals) and stated preferences (through natural language requests). This allows users to explicitly control their recommendations while maintaining the benefits of engagement-based recommendation.
• Figure 2: CTRL-Rec has the potential to be able to handle many kinds of user requests, unlocking many novel forms of user control.
• Figure 3: Overview of our method. We train a distilled model to approximate LLM judgements of whether a user would like a specific item based on their natural language request. This obviates the need for LLM calls (apart from a single request embedding operation) at test time.
• Figure 4: Qualitative examples of CTRL-Rec recommendations for three diverse natural language requests ($w_\text{control} = 0.995$ & no engagement history).
• Figure 5: CTRL-Rec reachability experiment results. (a) Traditional filters can already greatly reduce distance to the target feed, but CTRL-Rec meaningfully improves performance further across all distance bins. (b) CTRL-Rec achieves up to $\sim$45% distance reduction for source-target pairs with higher distance, and 30% for low distances. The percentage overlap metric shows similar trends, with one outlier showing decreased overlap for high-distance pairs.