Skip to content

CONCEPT Cited by 1 source

Catalog-aware scoring head

Definition

A catalog-aware scoring head is a neural network output layer that constrains the model's predictions to a known, fixed catalog of items — rather than generating arbitrary tokens from a vocabulary. Each catalog item has a learned embedding eᵢ, and a scoring function combines the model's hidden representation h with these embeddings to produce per-item scores.

Mechanics

  1. The backbone LLM produces a pooled hidden state h from the input prompt.
  2. Each catalog item i is represented by a learned embedding vector eᵢ.
  3. A scoring head φ(h, eᵢ) (dot product or small MLP) produces a scalar score sᵢ.
  4. Softmax over all scores yields a probability distribution converted to a ranking.

All parameters — backbone, scoring head, and item embeddings — are trained jointly.

Why It Matters

Standard LLM outputs are token sequences from a vocabulary — they can hallucinate items that don't exist. A catalog-aware head:

  • Eliminates hallucination — output is constrained to real, in-catalog items.
  • Enables efficient batch scoring — single matmul h × E^T scores the full catalog.
  • Supports business constraints — catalog updates (new releases, removals) are handled by updating the embedding table, not retraining the backbone.
  • Compatible with sampled softmax — for very large catalogs, training uses candidate subsets.

Distinction from Autoregressive Recommendation

Approach GenPage-style (autoregressive) GenRec-style (scoring head)
Output Generated sequence of item tokens Per-item scores via matmul
Decoding Token-by-token Single forward pass
Hallucination risk Possible (constrained decoding mitigates) Zero (scores only catalog items)
Business rules Enforced via constrained decoding masks Enforced at scoring/filtering layer

Seen in

Last updated · 606 distilled / 1,847 read