Skip to content

CONCEPT Cited by 1 source

Verbalization for recommendation

Definition

Verbalization for recommendation is the practice of representing user interaction histories, item metadata, and request context as natural-language text — rather than as dense feature vectors or sparse ID embeddings — so they can be processed by a large language model in its native semantic space.

Why It Matters

Traditional recommender systems operate on: - Hand-crafted dense features (numeric embeddings of user/item interactions) - Sparse categorical IDs (user_id, item_id, device_type) - Pre-computed interaction features (co-watch counts, session lengths)

Adding a new content type or surface requires significant feature engineering, architecture changes, and infrastructure work. Verbalization sidesteps this — a new content type is just a new text description the LLM can already parse.

The LLM's world knowledge and language understanding discover higher-level patterns (item relationships, evolving interests) that feature engineers would need to manually encode.

Trade-offs

Property Verbalization Dense features
Onboarding new content types Low cost (just describe in text) High cost (new feature pipelines)
Signal density per token Lower (natural language is verbose) Higher (compressed numeric)
Model capability requirement Needs strong LLM backbone Works with simple MLPs/DLRMs
Serving cost Higher (long token sequences) Lower (fixed-size vectors)
Feature discovery Implicit (model finds patterns) Explicit (engineer designs features)

Context Engineering as the New Discipline

When the prompt is the feature vector, context engineering becomes the primary modeling lever: - Retain high-signal engagements in full - Omit low-signal events (very short plays, hovers) - Summarize repetitive behaviors (binge-watching) - Elaborate selectively on important items (new releases, cold-start) - Structure for shared-prefix caching

The token budget is the new feature budget (Source: sources/2026-07-30-netflix-genrec-towards-llm-native-recommendation).

Seen in

Last updated · 606 distilled / 1,847 read