TL;DR
This paper introduces Decoupled Promptable Sequential Recommendation (DPR), a framework that allows users to steer traditional recommendation engines using natural language prompts (e.g., "show me something chaotic" or "no horror"). By modulating the latent user representation directly within the retrieval space, DPR combines the semantic reasoning of LLMs with the efficiency and collaborative precision of classic ID-based recommenders.
What Problem Does This Paper Address?
Modern recommendation systems (RS) are excellent at mining historical patterns—if you’ve watched ten thrillers, they will likely suggest an eleventh. However, they are "structurally blind" to sudden shifts in user intent. If that same user suddenly wants a "lighthearted movie for a family night," a standard model struggles to adapt because it is bound by the inertia of historical data.
Current attempts to fix this using Large Language Models (LLMs) fall into two traps:
- LLM-as-a-Recommender: These models are slow, expensive, and often ignore the rich "collaborative signals" (who else liked what) that make traditional systems so accurate.
- Reranking: Here, an LLM re-orders a list generated by a traditional model. The problem? If the base model didn't "recall" any family movies in its top 100 because of the user's thriller-heavy history, the LLM has nothing relevant to rank.
The authors propose a "Promptable Recommendation" paradigm where natural language instructions steer the retrieval process itself in real-time.
Methods
The authors propose the DPR framework, which is "model-agnostic"—meaning it can be bolted onto existing sequential backbones like SASRec or GRU4Rec. Its architecture consists of three main pillars:
- Signal Fusion Module: Using Multi-Head Cross-Attention (MHCA), the system treats the user’s historical interest as a "Query" and the natural language prompt as the "Key" and "Value." This allows the prompt to mathematically "nudge" the user's representation in the latent space.
- Mixture-of-Experts (MoE) Tower: The authors discovered that "positive steering" (I want X) and "negative constraints" (No Y) create conflicting optimization gradients. To solve this, they use a dual-path MoE architecture that routes requests to specialized "experts" for either amplifying or suppressing specific attributes.
- Three-Stage Training Strategy:
- Stage 1: Standard pre-training of the sequential backbone.
- Stage 2: Coarse alignment using broad categories (genres).
- Stage 3: Deep semantic alignment using LLM-generated descriptive tags (Narrative, Atmosphere, and Appeal) to bridge the gap between specific titles and abstract user desires.
Datasets
The researchers evaluated DPR on two widely used recommendation benchmarks, modified to simulate shifts in user intent:
- MovieLens-1M (ML1M): A classic movie recommendation dataset.
- MIND: A large-scale news recommendation dataset.
To train the "promptable" aspect, they used LLMs to enrich these datasets with "Semantic Augmented Tags," transforming simple genre labels into rich descriptions like "Whimsical and heartwarming childhood nostalgia."
Key Insights and Ideas
- Active Steering vs. Passive Filtering: Unlike a simple "filter" that just hides items, DPR's prompt vector modifies the user representation to "pull" relevant items to the top of the list. This is why DPR's gains in ranking metrics (NDCG) significantly outpaced its gains in retrieval metrics (Recall).
- Disentangling Intent: Positive and negative instructions are fundamentally different operations in vector space. Trying to force one set of parameters to learn both leads to "optimization conflicts," which the MoE architecture successfully resolves.
- Lexical Decoupling: To ensure the model wasn't just memorizing keywords, the authors tested it using "lexically distinct but semantically equivalent" prompts (e.g., training on "Toys" but testing on "Playthings"). The model's success here proves it understands the underlying intent, not just the words.
Key Results
- Dominant Performance: On the ML1M dataset using a SASRec backbone, DPR achieved a 71.84% relative improvement in NDCG@10 for positive steering tasks compared to the strongest filtering baselines.
- Better than LLM Rerankers: DPR outperformed state-of-the-art LLM reranking setups (including those using DeepSeek-V3 and Llama-2) while being significantly faster because it operates in a single end-to-end inference step.
- Zero Compromise: Crucially, adding promptability did not degrade the model's performance on standard sequential recommendation tasks. It remains a top-tier recommender even when the user doesn't provide a prompt.
Limitations and Future Work
- Metadata Dependency: The system's "intelligence" relies heavily on the quality of item metadata. While LLMs can generate this metadata, the approach might be harder to scale in domains where item descriptions are sparse or noisy.
- Semantic Indicator: Currently, the system uses a separate zero-shot classifier (BART) to determine if a prompt is "positive" or "negative" to route it through the MoE. Integrating this classification directly into the prompt embedder could further streamline the architecture.
- Domain Expansion: The study focused on movies and news; future work could explore how well this "steering" works in more complex spaces like e-commerce or short-form video.
Summary updated