- The paper introduces the Ordinal Semantic Anchoring (OSA) framework, preserving detailed rating semantics in LLM-based recommender systems.
- It maps user-item interactions into the LLM latent space using a two-layer MLP and fixed token embeddings as semantic anchors.
- Empirical validation shows enhanced overall Hit@1 and up to 44.9% improvement in fine-grained preference discrimination on multiple datasets.
Injecting Ordinal Semantics into LLM-Based Recommenders: The OSA Framework
Motivation and Limitations of Prior Work
LLMs have recently been leveraged in recommender systems, often through hybrid strategies that integrate collaborative filtering (CF) signals using learned projections or representation mapping. While these hybrid CF-LLM frameworks outperform their purely collaborative or text-based counterparts, they commonly reduce explicit user ratings to binary or positive-only feedback. This simplification discards the ordinal semantics intrinsic to rating scales, which encode nuanced levels of user preference intensity. The result is a diminished capacity to infer fine-grained distinctions among candidate items—an area where LLMs' semantic modeling abilities are especially relevant, but underutilized.
The Ordinal Semantic Anchoring (OSA) Method
The proposed Ordinal Semantic Anchoring (OSA) framework addresses the above limitation by explicitly modeling preference strength using interaction-level anchoring in the LLM latent space. OSA operates as follows:
- Interaction-Level Representation: Each user-item interaction in the CF component is encoded as a concatenation of user and item embeddings. Crucially, each interaction is labeled with its observed rating, preserving the granularity of preference.
- Learnable Projection: Interaction representations are mapped into the LLM's feature space via a two-layer MLP, enabling their integration as contextual vectors in downstream prompts.
- Ordinal Semantic Anchors: Each possible rating (e.g., 1 through 5) is associated with its respective numeric token embedding from the pretrained LLM. These fixed embeddings serve as semantic anchors. The inherent ordinal structure in the LLM's token space is leveraged to structurally organize projected collaborative signals.
- Strength-Aware Alignment Objective: The alignment loss explicitly draws projected CF interaction representations toward their corresponding semantic anchor, with additional weighting applied to interactions representing more extreme (strong like/dislike) ratings. This ensures that the LLM's latent organization respects the ordinal preference intensity, rather than collapsing all positive interactions into a single latent region.
- Prompt Engineering: During inference and training, the system presents user histories and candidate set information using hybrid prompts that explicitly include both items and their associated ratings, further reinforcing ordinal grounding in the input text.
Empirical Validation and Numerical Results
OSA was evaluated on three explicit-feedback datasets: MovieLens-1M, Amazon Scientific, and Amazon Video Games. The next-item prediction task was used, and evaluation focused on Hit@1 as well as a pairwise preference task designed to directly assess fine-grained discrimination.
- General Recommendation Performance: Across all datasets, OSA achieved superior Hit@1 compared to both collaborative-only and existing hybrid CF-LLM baselines. For example, on MovieLens-1M, OSA obtained 0.6014 Hit@1, exceeding LLaRA (the best prior hybrid method) by a significant margin.
- Fine-Grained Preference Discrimination: For pairwise judgments involving subtle and strong preference differences (e.g., 1 vs. 2 or 4 vs. 5), OSA outperformed prior methods with relative gains of up to 44.9% in strong preference discrimination accuracy. Methods that binarized ratings, or failed to maintain ordinal structure, approached random-guess baselines in these evaluations.
- Ablation Results: Removing OSA's alignment loss, its strength-aware weighting, or the use of LLM-based semantic anchors each materially degraded both overall and pairwise performance. These results confirm that preserving rating-level semantics in both representation and alignment is essential to OSA's efficacy.
Practical and Theoretical Implications
OSA establishes that explicit preservation of ordinal semantics offers clear benefits in LLM-based recommendation pipelines, especially when recommendations must discriminate between closely-rated alternatives. The framework leverages the natural ordinal structure present in LLM token embeddings, allowing affordable, parameter-efficient adaptation via LoRA without bespoke architectural changes to the LLM itself.
Practically, this approach facilitates the use of LLM-based recommenders in domains where fine-grained user intent needs to be surfaced, such as personalized entertainment, expert technical content, and preference-sensitive e-commerce. Theoretically, OSA opens avenues to further explore how LLMs' pretrained semantic hierarchies can be systematically integrated with structured, domain-specific signals (such as ordinal ratings) in other AI tasks.
Future Directions
Potential extensions include the development of semantic anchors for implicit feedback, enabling similar ordinal modeling even in environments lacking explicit rating data. There is also scope for applying analogous anchoring strategies to continuous or categorical preference signals, or in multi-faceted recommendation scenarios where multiple preference dimensions must be integrated.
Conclusion
OSA provides a principled mechanism for encoding and leveraging ordinal semantics in LLM-based recommender systems, overcoming a key limitation in prior hybrid approaches. Empirical results demonstrate substantial improvements in both overall recommendation quality and the modeling of nuanced user preferences. The framework offers a template for further integration of structured preference signals with pretrained LLM semantics, enhancing both the interpretability and specificity of AI-driven recommendations.
