A Unified Language Model for Large Scale Search, Recommendation, and Reasoning
Authors: Marco De Nadai, Edoardo D'Amico, Max Lefarov, Alexandre Tamborrino, Divita Vohra, Mark VanMiddlesworth, Shawn Lin, Jacqueline Wood, Jan Stypka, Eliza Klyce, Keshi Dai, Timothy Christopher Heath, Martin D. Gould, Yves Raimond, Sandeep Ghael, Tony Jebara, Andreas Damianou, Vladan Radosavljevic, Paul N. Bennett, Mounia Lalmas, Praveen Chandar
Organizations: Spotify Denmark, Spain, Germany, France, United States, United Kingdom
LLMs are increasingly applied to recommendation, retrieval, and reasoning, yet deploying a single end-to-end model that can jointly support these behaviors over large, heterogeneous catalogs remains challenging. Such systems must generate unambiguous references to real items, handle multiple entity types, and operate under strict latency and reliability constraints requirements that are difficult to satisfy with text-only generation. While tool-augmented recommender systems address parts of this problem, they introduce orchestration complexity and limit end-to-end optimization. We view this setting as an instance of a broader research problem: how to adapt LLMs to reason jointly over multiple-domain entities, users, and language in a fully self-contained manner. To this end, we introduce NEO, a framework that adapts a pre-trained decoder-only LLM into a tool-free, catalog-grounded generator. NEO represents items as SIDs and trains a single model to interleave natural language and typed item identifiers within a shared sequence. Text prompts control the task, target entity type, and output format (IDs, text, or mixed), while constrained decoding guarantees catalog-valid item generation without restricting free-form text. We refer to this instruction-conditioned controllability as language-steerability. We treat SIDs as a distinct modality and study design choices for integrating discrete entity representations into LLMs via staged alignment and instruction tuning. We evaluate NEO at scale on a real-world catalog of over 10M items across multiple media types and discovery tasks, including recommendation, search, and user understanding. In offline experiments, NEO consistently outperforms strong task-specific baselines and exhibits cross-task transfer, demonstrating a practical path toward consolidating large-scale discovery capabilities into a single language-steerable generative model.
Figures & tables
Figure 1. NEO adapts pre-trained LLMs to speak multi-item-type catalogs (e.g. audiobooks and podcasts) through SIDs (e.g. 2 tokens long tuples), and to output both text and SIDs.
Capability ↓ / System →
Generative SID Retrieval ( Qu et al., 2024 ; Rajput et al., 2024 )
Tool-based LLMs ( Doh et al., 2025a ; Friedman et al., 2023 ; Lyu et al., 2024 ; Wang et al., 2024d )
PLUM ( He et al., 2025 )
TalkPlay ( Doh et al., 2025b )
NEO (this work)
Discrete item pointers
✓
✗
✓
✓
✓
Mixed text + item output
✗
▲
✗
✓
✓
Mixed text + item input
▲
✓
✓
✓
✓
User behaviour in input
✓
▲
✓
▲
✓
Language-steerable behavior
✗
✓
▲
✓
✓
Multiple item types
▲
▲
✗
✗
✓
Table 1. Comparison of NEO with prior recommendation paradigms. While several approaches support individual capabilities (e.g., discrete item IDs or LLM-based generation), NEO is the first to combine typed item addressing, mixed text–item outputs, language-steerable behavior, and multi-task discovery within a single tool-calling-free generative model.
Figure 2. We illustrate our approach as the first three stages in a more general four-stage pipeline. In the general case, in Stage 1, semantic representations are learned—such as semantic IDs (SIDs)—to capture entities, their structure, and relationships. In Stage 2, domain grounding is performed via one or both of (i) adaptation over user-behavior sequences and (ii) domain alignment between SID/text pairs, to learn a shared latent space between entities and natural language. This enables the model to use text and SIDs interchangeably and to learn sequential, domain-relevant concepts. Stage 3 teaches the model to answer questions, follow commands, and generate grounded outputs through multi-task instruction-following. We prioritize diversity of tasks over depth within any single task. In Stage 4, we may leverage task-specific datasets and objectives related to the targeted use case (e.g., topic discovery and business constraints) via task-specific fine-tuning, reinforcement learning, and related techniques. We indicate typical choices for which parameters are kept frozen (ice) or trainable (fire) for each stage.
Method
Δ HR@10
NEO (Baseline)
-
A - Atomic item IDs
-59.7%
Semantic IDs
B - LSH quantizer
-51.2%
C - No augmented data
-2.9%
D - Collaborative filtering
-25.6%
Table 2. Ablation for multiple types of item representation.
Inference strategy
Δ HR@K
Δ latency
k=10
k=30
k=10
NEO (Baseline)
-
-
-
A - Beam Search w/o constrains
-1.80%
-1.93%
-2.92%
B - Top- p sampling
-21.97%
-32.41%
-7.91%
Table 3. Accuracy/latency trade-off at inference time.
Method
HR@K ↑
NDCG@K ↑
k=10
k=30
k=10
k=30
NEO (Baseline)
-
-
-
-
A - w/o Domain Grounding
{\color[rgb]{0.6992,0.1328,0.1328}-6\%}
{\color[rgb]{0.6992,0.1328,0.1328}-7\%}
{\color[rgb]{0.6992,0.1328,0.1328}-8\%}
{\color[rgb]{0.6992,0.1328,0.1328}-8\%}
B - Everything in two-stages
{\color[rgb]{0.6992,0.1328,0.1328}-8\%}
{\color[rgb]{0.6992,0.1328,0.1328}-7\%}
{\color[rgb]{0.6992,0.1328,0.1328}-10\%}
{\color[rgb]{0.6992,0.1328,0.1328}-10\%}
C - Random init
{\color[rgb]{0.6992,0.1328,0.1328}-44\%}
{\color[rgb]{0.6992,0.1328,0.1328}-44\%}
{\color[rgb]{0.6992,0.1328,0.1328}-44\%}
{\color[rgb]{0.6992,0.1328,0.1328}-43\%}
D - Continuos pretraining
{\color[rgb]{0.6992,0.1328,0.1328}-3\%}
{\color[rgb]{0.6992,0.1328,0.1328}-3\%}
{\color[rgb]{0.6992,0.1328,0.1328}-2\%}
{\color[rgb]{0.6992,0.1328,0.1328}-3\%}
Table 4. Comparison of training strategies alternative to NEO.
Method
HR@K ↑
NDCG@K ↑
k=10
k=30
k=10
k=30
Episode Recommendation
Baseline
-
-
-
-
NEO - mono
{\color[rgb]{0.1328,0.5469,0.1328}+57\%}
{\color[rgb]{0.1328,0.5469,0.1328}+41\%}
{\color[rgb]{0.1328,0.5469,0.1328}+80\%}
{\color[rgb]{0.1328,0.5469,0.1328}+60\%}
NEO - multi
{\color[rgb]{0.1328,0.5469,0.1328}+58\%}
{\color[rgb]{0.1328,0.5469,0.1328}+40\%}
{\color[rgb]{0.1328,0.5469,0.1328}+80\%}
{\color[rgb]{0.1328,0.5469,0.1328}+59\%}
Show Recommendations
Table 5. Multi-task performance
Variation
Artists
Episodes
Interest profiling
Coverage
3.79
3.47
Groundedness
3.81
3.50
Recsplanation
Faithfulness item
4.08
3.28
Faithfulness history
4.16
2.79
Deceptiveness
4.74
4.52
Table 6. LLM as a Judge result for user analysis and recsplanation tasks. Scores from 0 to 5
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3. Recsplanation result for an episode request. This randomly selected result shows that the model can interleave recommendations (SIDs) and explanation (text) fluently.
Figure 4. Qualitative example of user understanding for an episode, from the SID history of an user.
Hyperparameters
Stage 2
Stage 3
Learning rate
1.0×10−3
1×10−4
LR scheduler
Cosine
Cosine
Weight decay
0.01
0.01
Gradient clip
1.0
1.0
Optimizer
AdamW ( β1=0.9,β2=0.95 )
Warm-up steps
300
400
Appendix
Table 7. Detailed hyperparameters of NEO applied to Qwen
User type
HR@K ↑
NDCG@K ↑
k=10
k=30
k=10
k=30
Coldstart
{\color[rgb]{0.6992,0.1328,0.1328}-48\%}
{\color[rgb]{0.6992,0.1328,0.1328}-57\%}
{\color[rgb]{0.6992,0.1328,0.1328}-34\%}
{\color[rgb]{0.6992,0.1328,0.1328}-40\%}
Warmstart
{\color[rgb]{0.1328,0.5469,0.1328}+57\%}
{\color[rgb]{0.1328,0.5469,0.1328}+40\%}
{\color[rgb]{0.1328,0.5469,0.1328}+70\%}
{\color[rgb]{0.1328,0.5469,0.1328}+60\%}
Appendix
Table 8. Episode recommendations detailed results. The LLM has room for improvement in coldstart users, which represent ∼ 2% of our data.
Variation
HR@K ↑
NDCG@K ↑
k=10
k=30
k=10
k=30
Item type
Artists
{\color[rgb]{0.1328,0.5469,0.1328}+36\%}
{\color[rgb]{0.1328,0.5469,0.1328}+35\%}
{\color[rgb]{0.1328,0.5469,0.1328}+38\%}
{\color[rgb]{0.1328,0.5469,0.1328}+38\%}
Audiobooks
{\color[rgb]{0.1328,0.5469,0.1328}+26\%}
{\color[rgb]{0.1328,0.5469,0.1328}+27\%}
{\color[rgb]{0.1328,0.5469,0.1328}+29\%}
{\color[rgb]{0.1328,0.5469,0.1328}+30\%}
Episodes
{\color[rgb]{0.1328,0.5469,0.1328}+40\%}
{\color[rgb]{0.1328,0.5469,0.1328}+16\%}
{\color[rgb]{0.1328,0.5469,0.1328}+73\%}
{\color[rgb]{0.1328,0.5469,0.1328}+59\%}
Query complexity
Appendix
Table 9. Text-based retrieval/search performance by item type and query complexity.
Figure 5. Prompt template used for LLM as a Judge for the user analysis task.
Figure 6. Prompt template used for LLM as a Judge for the recsplanation task.
Figure 7. Prompt template used for text-based retrieval/search.
Figure 8. Prompt template used for episode recommendation.
Figure 9. Prompt template used for show recommendation.
Figure 10. Prompt template used for audiobook recommendation.
Large language model-based recommendation uses language models to personalize ranking, retrieval, and conversational suggestions from text and behavior.
The paradigm shift from item-centric ranking to answer-centric synthesis is redefining the role of search engines. While recent industrial progress has applied generative techniques to closed-set item ranking in e-commerce, research and deployment of open-ended generative search on large content platforms remain limited. This setting introduces challenges, including robustness to noisy retrieval, non-negotiable safety guarantees, and alignment with diverse user needs. In this work, we introduce SearchLLM, the first large language model (LLM) for open-ended generative search. We design a hierarchical, multi-dimensional reward system that separates bottom-line constraints, including factual grounding, basic answer quality and format compliance, from behavior optimization objectives that promote robustness to noisy retrieval and alignment with user needs. Concretely, our reward model evaluates responses conditioned on the user query, session history, and retrieved evidence set, combining rule-based checks with human-calibrated LLM judges to produce an interpretable score vector over these dimensions. We introduce a Gated Aggregation Strategy to derive the training reward for optimizing SearchLLM with Group Relative Policy Optimization (GRPO). We deploy SearchLLM in the AI search entry of RedNote. Offline evaluations and online A/B tests show improved generation quality and user engagement, increasing Valid Consumption Rate by 1.03% and reducing Re-search Rate by 2.81%, while upholding strict safety and reliability standards.
Wei Wu, Peilun Zhou, Liyi Chen +6
School of Artificial Intelligence and Data Science, University of Science and Technology of China, Hefei, China · Xiaohongshu Inc., Beijing, China · Thrust of Artificial Intelligence, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China +1
Modern LLMs excel at reasoning and instruction following, enabling users to express complex and diverse information needs. However, conventional retrievers largely rely on surface-level matching between queries and documents, resulting in a growing gap between how users express their needs and how retrievers interpret them. In this paper, we present GEM, a generative embedding model that augments retrieval through its own knowledge by explicitly reasoning about user intent and relevance criteria. GEM unifies generation and embedding within a single model: it first reasons over the query, then appends an embedding token to encode the enriched context for retrieval. Evaluated on reasoning-intensive and instruction-following retrieval tasks, GEM demonstrates the effectiveness of its reasoning-augmented retrieval, outperforming its non-reasoning variant and matching baselines using substantially larger models. Furthermore, GEM's generative nature allows test-time compute scaling via prompting to further enhance retrieval performance. Our code is available at: https://anonymous.4open.science/r/GEM.
Recent advances in generative recommendation have leveraged pretrained LLMs by formulating sequential recommendation as autoregressive generation over a unified token space comprising language tokens and itemic identifiers, where each item is represented by a compact sequence of discrete tokens, namely Semantic IDs (SIDs). This SID-based formulation enables efficient decoding over large-scale item corpora and provides a natural interface for LLM-based recommenders to leverage rich world knowledge. Meanwhile, breakthroughs in LLM reasoning motivate reasoning-enhanced recommendation, yet effective reasoning over SIDs remains underexplored and challenging. Itemic tokens are not natively meaningful to LLMs; moreover, recommendation-oriented SID reasoning is hard to evaluate, making high-quality supervision scarce. To address these challenges, we propose SIDReasoner, a two-stage framework that elicits reasoning over SIDs by strengthening SID--language alignment to unlock transferable LLM reasoning, rather than relying on large amounts of recommendation-specific reasoning traces. Concretely, SIDReasoner first enhances SID-language alignment via multi-task training on an enriched SID-centered corpus synthesized by a stronger teacher model, grounding itemic tokens in diverse semantic and behavioral contexts. Building on this enhanced alignment, SIDReasoner further improves recommendation reasoning through outcome-driven reinforced optimization, which guides the model toward effective reasoning trajectories without requiring explicit reasoning annotations. Extensive experiments on three real-world datasets demonstrate the effectiveness of our reasoning-augmented SID-based generative recommendation. Beyond accuracy, the results highlight the broader potential of large reasoning models for generative recommendation, including improved interpretability and cross-domain generalization.
Yingzhi He, Yan Sun, Junfei Tan +6
National University of Singapore Singapore, Singapore · University of Science and Technology of China Hefei, China · Tencent Inc. Shenzhen, China