Prompt Recommender Systems

• Prompt Recommender Systems are algorithmic frameworks that construct, recommend, and optimize structured inputs to guide model behavior in recommendation tasks.
• They leverage techniques like parameter-efficient tuning, in-context learning, and meta-learning to enhance personalization, fairness, and cross-domain transfer.
• PRS have demonstrated significant improvements in cold-start settings and real-time applications, achieving measurable gains in click-through, dwell, and ranking metrics.

Prompt Recommender Systems (PRS) are algorithmic and interaction frameworks in which prompts—structured inputs or instructions—are constructed, recommended, and optimized either to steer foundation models (especially LLMs, LMs) for downstream recommendation tasks, or to augment user-facing applications by surfacing contextually relevant prompt suggestions. PRS generalize beyond classical recommendation by considering both "prompts for models" (parameter-efficient tuning, context control, personalization, fairness, and domain adaptation) and "prompt-as-user-item" (human-facing suggestions in co-creative AI). This article surveys the key principles, model architectures, algorithmic strategies, empirical results, and deployment scenarios of PRS, drawing on systematizations and novel methods published between 2022–2026.

1. Definition and Taxonomy of Prompt Recommender Systems

Prompt Recommender Systems comprise two interrelated paradigms:

• Prompt-based recommender models: Traditional recommendation objectives (ranking, CTR prediction, rating regression, slot filling, conversation) are reformulated as prompt-based input–output mappings for LLMs or multi-modal models. A prompt either replaces explicit retraining (zero/few-shot), constitutes the main interface for adaptation (soft/hard prefix, template, or input chunk), or enables flexible parameter-efficient tuning.
• Prompt-for-the-user approaches: Prompts themselves are treated as first-class recommendable items, with interactive systems recommending follow-up prompts to end-users to scaffold ideation, task planning, or exploration (Kim et al., 22 Jan 2026).

An authoritative taxonomy for PRS architectures (consolidating survey (Liu et al., 2023), empirical frameworks (Xu et al., 2024), and application-specific works) is summarized as:

Training Paradigm	Prompt Type	PLM Weights Tuned?	Representative Works
Fixed-PTM Prompt Tuning	Soft & hard	No	(Wu et al., 2022, Wu et al., 2023)
Fixed-Prompt PTM Tuning	Hard only	Yes	(Liu et al., 2023)
Tuning-Free Prompting	Hard only	No	(Liu et al., 2023, Li et al., 2023)
Prompt+PTM Tuning	Soft & hybrid	Yes	(Jiang et al., 2024)

Prompts may be discrete (fixed text templates), continuous (learned prefix or input embeddings), graph-based (neighbor prompts for GNN), user/item-specific (meta-learned), or composite (task/domain decomposed, fairness-controlled, etc.).

2. Prompt Construction and Optimization Strategies

2.1. Prompt-as-Template for In-Context Learning

Prompt-based recommenders linearize structured user/item/context information into natural language (NL) strings or embedding sequences. This is seen in PromptRec, which transforms each user–item pair into a template (structured as user profile, item profile, masked sentiment sentence), then asks the LM to select a sentiment word ('good' or 'bad') as a supervised classification task (Wu et al., 2023).

2.2. Prompt Generation and Decomposition

Prompts can be decomposed into task and domain sub-prompts, with explicit cross-domain adaptation via prompt pre-training (e.g., task prompt S_T, fixed from source data; domain prompt S_D, set by keyword extraction on target data) (Wu et al., 2023). In item cold-start, pinnacle feedback is selected as the set of top-k highest-value user interactions (as measured by dwell time, explicit preference), and encoded as soft prompts by small per-item networks (Jiang et al., 2024).

2.3. Data-Centric and Parameter-Efficient Tuning

Data-centric PRS optimize relevance by (a) curating refined pre-training corpora rich in relevant user–item–response contexts, (b) maximizing mutual information between schema and real-world corpora, and (c) only tuning prompt or adapter parameters while freezing the backbone (parameter- and compute-efficient) (Wu et al., 2023, Jiang et al., 2024). Meta-learning further enables rapid per-user adaptation by treating each user as a task and optimizing soft prompt embeddings via Reptile or MAML-style episodic learning, achieving real-time (<300 ms) personalization on cold-start users (Zhao et al., 22 Jul 2025).

2.4. Interactive and User-Facing Prompt Suggestion

PRS for user interaction (e.g., PromptHelper (Kim et al., 22 Jan 2026)) treat prompts as recommendable artifacts, suggesting follow-up prompts in diverse semantic categories with explicit context grounding. Candidate prompts are generated by prompting the LLM with the user's last turn and conversation context, then filtered for diversity by taxonomy category.

3. Core Model Architectures and Theoretical Perspectives

3.1. Prompt-Tuned Transformers and LMs

Architectures leverage transformer backbones (e.g., BERT, GPT, T5), often frozen, with prompts injected as prefix tokens (soft/continuous), contextualized template segments, or keyword-decomposed vectors. In contrastive graph prompt-tuning, per-user graph prompts (hard and soft) are inserted into GNN aggregator layers, tuned on a cross-domain target while the GNN backbone is fixed (Yi et al., 2023).

3.2. Unified Conversational and Multi-task Models

Prompt learning unifies multi-task objectives (e.g., recommendation and conversation in CRS) by treating each subtask as sequence generation over a single LM backbone, augmented with fused knowledge embeddings, soft prompts, and dynamic context (Wang et al., 2022). Continual learning versions allocate a per-task prompt as external memory, supporting multi-objective user modeling without catastrophic forgetting (Yang et al., 26 Feb 2025).

3.3. Theoretical Underpinnings

Data-centric PRS are formalized as hierarchical mixture models (e.g., HMM over latent "concepts" θ), with success relying on accurate estimation of conditional factors p(y|X,θ) (outcome given context and concept), p(X|θ) (context likelihood), and p(θ) (concept prior) (Wu et al., 2023). RL-based PRS reframe RL training as a supervised prompt-to-action mapping, substituting state/reward pair prompts for policy/value target estimation (Xin et al., 2022).

4. Task Settings, Evaluation, and Empirical Performance

PRS are evaluated primarily on cold-start settings (system, user, or item cold-start), cross-domain transfer, few/zero-shot adaptation, and multi-task generalization. Metrics include nDCG@K, HitRate@K, AUC, parameter count, latency, and empirical sample efficiency.

• System cold-start: Small LMs enhanced by refined pre-training and TPPT achieve GAUC ≈57–64%, comparable to BERT-large (~57–58%) at <20% of latency (Wu et al., 2023).
• Item cold-start: PROMO outperforms classical and existing prompt-based baselines in Hit@5 and NDCG@10, with up to +16 percentage point gain, and is validated in billion-user production A/B tests (Jiang et al., 2024).
• Cross-domain: Graph-prompt approaches achieve 74% parameter savings and 11.4% Recall@10 uplift in target-domain cold-start, due to strong knowledge transfer via contrastive pretraining and per-user hard prompt construction (Yi et al., 2023).
• Meta-learning prompt personalization: Meta-prompt adaptation delivers superior NDCG@10 and adaptation time (<300 ms), with zero-history personalization exceeding statically tuned or zero-shot LLMs by 8–15% Hit@10 (Zhao et al., 22 Jul 2025).
• Recommendation fairness: Selective fairness PRS, using per-attribute prompt-adapter blocks and adversarial training, achieve substantial reduction in attacker F1 (+12–41% relative) with ≤3% utility loss (Wu et al., 2022).
• Interactive PRS for users: PromptHelper studies show significantly increased exploration and expressiveness, with no observed rise in user workload, when PRS-generated prompt lists supplement standard chatbot interaction (Kim et al., 22 Jan 2026).

5. Prompt Engineering, Design, and Empirical Insights

A systematic analysis of PRS engineering dimensions highlights four critical, empirically validated prompt components (Xu et al., 2024):

• Task Description ($\tau$): Re-ranking, click-through prediction, and multi-candidate selection can be cast as point-wise, pair-wise, or list-wise NL prompts or as action-generation over state/reward contexts (Liu et al., 2023, Xin et al., 2022).
• User Interest Modeling ($\phi(H_u)$): Concatenation of short-term (recent items) and long-term (summary or personalized memory) achieves higher NDCG@10; adding LLM-generated summarizations further improves accuracy (Xu et al., 2024).
• Candidate Construction ($C_u$): Representing candidates as text titles (rather than IDs) and using jointly pre-trained LLM text embeddings increase ranking accuracy (+0.04–0.31 NDCG@10); re-ranking over BPR or pop-recall lists benefits more than over deep sequential baselines (Xu et al., 2024).
• Prompting Strategy ($\sigma$): Role cues (“You are a recommender”), recency emphasis, and tailored chain-of-thought variants yield measurable improvements; generic few-shot ICL typically does not outperform optimized zero-shot prompting.

Batched feedback aggregation and position-aware correction, as in AGP (Wang et al., 4 Apr 2025), stabilize prompt refinement for LLM-driven reranking, achieving >5.6%–20.7% NDCG@10 improvement over base models.

6. Practical Deployment and Applications

PRS frameworks have seen successful large-scale application:

• Commercial deployments: PROMO was operationalized on Kuaishou's short-video platform for billions of users, demonstrating +3%–5% uplifts in click-through, dwell, and engagement metrics for cold-start content (Jiang et al., 2024).
• User-facing PRS: PromptHelper, targeting creative writing, showed that recommending diverse, context-sensitive prompt alternatives increases user engagement and expressiveness in structured studies (Kim et al., 22 Jan 2026).
• Domain-specialized assistants: Context-aware PRS integrating telemetry, knowledge retrieval, and hierarchical ranking achieved >96% prompt usefulness for security analysts, with modular support for new plug-ins or skills (Tang et al., 25 Jun 2025).
• Finance, healthcare, education: Meta-learned personalized prompt-tuning enables real-time (<300 ms) cold-start profiling in financial risk scenarios and is being assessed in low-resource medical and educational recommenders (Zhao et al., 22 Jul 2025).

7. Future Directions, Open Challenges, and Limitations

Open problems and research frontiers include:

• Inference and Latency: Scaling prompt-based LLMs for industrial real-time use remains bottlenecked by context length and parameter size (Xu et al., 2024). Efficient distillation, caching, and adaptive parameter-tuning are required.
• Fairness, Privacy, and Interpretability: While soft prompts support selective fairness and privacy controls, understanding prompt embedding semantics and mitigating bias or leakage in end-to-end LMs are under-explored (Wu et al., 2022).
• Prompt Robustness, Transfer, and Meta-Learning: Domain shift, cross-task transfer, and rapid few-shot adaptation merit additional theoretical and empirical study. Meta-learning and domain-adaptive prompt sampling are promising mitigations (Zhao et al., 22 Jul 2025).
• Prompt Generation and Representation: Optimal choices for prompt length, content (ID, content, feedback), and network architecture depend on data and context (Jiang et al., 2024).
• Compositionality and Task Integration: Unifying recommendation, generation, explanation, and control (diversity, topic, fairness) in a single prompt-based interface calls for further research (Wang et al., 2022, Li et al., 2023).
• User Experience and Interaction: PRS as exploratory aids must balance automation with user agency, minimize cognitive load, and provide actionable, transparent suggestions, especially as prompt complexity and number increase (Kim et al., 22 Jan 2026).

Prompt Recommender Systems represent an overview of parameter-efficient model adaptation, prompt engineering, and interactive suggestion, achieving both strong empirical advances in traditional recommendation metrics and novel capacities for personalization, fairness, and user engagement. The field is evolving rapidly, with emerging multimodal extensions, optimization of RLHF-for-prompt strategies, and deeper integration of prompt recommendation into AI user interfaces (Liu et al., 2023, Xu et al., 2024, Kim et al., 22 Jan 2026).