> When Does LLM Reasoning Help Recommendation? The Semantic Richness Index

Recommender Systems 中级 2026-04-01 06:41 2026-04-01
#LLM #recommender systems #efficiency #cross-domain #routing #cost-aware

A cross-domain study: LLMs improve recommendation only when SRI >= 0.45 (semantic-rich content). On MIND news: +102%. On Amazon Electronics: -50%. Decision rule + 3 efficiency mechanisms yield 81.8% cost reduction.

Semantic Richness Index for LLM Recommendation

Cross-Domain Results

Dataset Items SRI LLM Effect
MIND (news) 101K 0.56 +102%
MovieLens 3.9K 0.41 -3%
Amazon Movies 748K 0.19 -11%
Amazon Electronics 1.6M 0.38 -50%
Criteo (encrypted) 5M 0.26 -46%

Decision rule: If SRI < 0.45, skip LLM integration.

SRI Formula

\[\text{SRI} = 0.25 \times \frac{\text{title\_len}}{20} + 0.25 \times \text{vocab\_diversity} + 0.25 \times \frac{\text{cat\_entropy}}{5} + 0.25 \times \text{desc\_coverage}\]

Pearson r=0.60 with LLM benefit. Runs in seconds on any dataset, no training needed.

Three Efficiency Mechanisms

  1. Confidence-based early exit: 0.2ms (trie) vs 166ms (LLM call)
  2. Trie-guided context compression: 47% token reduction with 2.1% quality improvement
  3. Score-based caching: Cache user scores (not rankings) for re-ranking new items

Combined: 81.8% cost reduction, 53% carbon savings.

Key Negative Result

FrugalGPT cascade routing provides zero benefit for recommendation — confidence scoring that works for classification doesn't transfer to recommendation because difficulty is driven by data sparsity, not semantic complexity.