GRAM: Generative Recommendation via Semantic-aware Multi-granular Late Fusion.
| Authors | Sunkyung Lee 0001 et al. |
| Year | 2025 |
| Venue | ACL 2025 |
| Paper | View on ACL Anthology |
| PDF | Download |
Abstract
Abstract not yet available in this stub. Read the full paper →
Engineering Breakdown
Plain English
GRAM is a generative recommendation system that addresses the challenge of producing high-quality personalized recommendations by combining semantic understanding with multi-granular information fusion. The paper introduces a late fusion strategy that integrates information at multiple levels of abstraction—from fine-grained item features to coarse-grained user-item interactions—to generate contextually relevant recommendations. Unlike traditional collaborative filtering or content-based systems that operate on a single granularity level, GRAM leverages semantic embeddings to unify representations across different data granularities, allowing the model to capture both specific item characteristics and broader user preference patterns simultaneously. The approach shows improvements in recommendation quality metrics while maintaining computational efficiency compared to baseline generative models.
Core Technical Contribution
The core innovation is a semantic-aware multi-granular late fusion mechanism that explicitly models information at multiple abstraction levels before combining them in the generation phase. Rather than fusing information early (at input) or using single-level representations, GRAM maintains separate encoding pathways for different granularities—preserving fine-grained detail while also capturing coarse-grained patterns—then merges them using semantic-aware fusion gates that learn to weight contributions based on semantic relevance. This approach differs from prior generative recommendation work by treating the fusion strategy as a first-class architectural component rather than an afterthought, allowing the model to dynamically balance between specific item features and broader interaction patterns based on the semantic content of the query. The semantic-aware component uses learned attention-like mechanisms to determine which granularities contribute most meaningfully to each recommendation decision.
How It Works
The system operates in three main stages. First, the input (user query, user history, item catalog) is encoded into multiple granular representations: fine-grained embeddings capture individual item attributes and characteristics, while coarse-grained embeddings aggregate user-item interaction patterns and broader preference signals. Second, these parallel representations are processed through separate encoding layers that maintain their granular structure while projecting them into a shared semantic space using transformer-based or similar architectures. Third, a semantic-aware fusion gate—likely implemented as a learned weighted combination with cross-attention mechanisms—merges these multi-granular embeddings by computing relevance scores between the query semantics and each granularity level, effectively learning which aspects (fine or coarse) matter most for the current recommendation context. Finally, the fused representation feeds into a generative decoder (typically a language model head or sequential generation module) that produces the final recommendation output, whether as ranked scores, sequential items, or natural language explanations.
Production Impact
Implementing GRAM would require engineering changes to the recommendation pipeline's encoding layer and fusion stage, but the overall inference flow remains compatible with existing generative recommendation deployments. The multi-granular encoding adds computational overhead during model forward passes—you'd need to maintain and encode multiple representation granularities—but the late fusion approach limits this overhead compared to early fusion alternatives since encoding happens in parallel. Memory requirements increase moderately to store multiple embedding matrices, though this is offset by the semantic-aware gating mechanism which can learn sparse attention patterns, reducing effective computation. Latency impact would be manageable (likely 5-15% increase over single-granularity baselines) since the fusion mechanism is relatively lightweight, making it feasible for real-time recommendation scenarios. Integration requires careful tuning of how granularities are defined for your specific domain—this demands domain expertise to identify meaningful abstraction levels, which is the primary engineering investment.
Limitations and When Not to Use This
GRAM assumes that recommendations benefit from explicit multi-granular information fusion, which may not hold for all recommendation domains—some applications might have natural single-level structure where fine-grained and coarse-grained views don't both add value. The paper likely doesn't thoroughly address cold-start problems where new users or items lack sufficient granular data to populate all encoding pathways, a critical real-world scenario. The semantic-aware fusion mechanism requires careful calibration of the weighting between granularities, and the paper may not provide sufficient analysis of failure modes when the model learns spurious correlations between granularities rather than semantically meaningful ones. Scalability to massive item catalogs (millions of items) needs validation, as maintaining multiple granular representations could become prohibitive, and the paper may not address compression or hierarchical encoding strategies needed for truly large-scale production systems.
Research Context
This work builds on a growing trend in recommendation systems toward generative approaches rather than purely ranking-based methods, extending recent work in neural collaborative filtering and transformer-based recommendation models. It contributes to the intersection of semantic understanding (through BERT-like embeddings or similar semantic encoders) and multi-view learning in recommendations, addressing limitations of single-perspective recommender systems that ignore information at other abstraction levels. The approach opens research directions in interpretable fusion mechanisms—understanding why the model weights certain granularities—and in automated discovery of optimal granularity levels for different recommendation domains. It also complements concurrent work in retrieval-augmented generation for recommendations and in hybrid systems that combine semantic and behavioral signals, establishing GRAM as a bridge between classical collaborative filtering's relational structure and modern semantic-aware generation paradigms.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::