Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation

Conference: NeurIPS 2025 arXiv: 2503.05493
Code: RecBench
Area: Audio & Speech Keywords: LLM-as-RS, RecBench, CTR prediction, sequential recommendation, item representation, inference efficiency

TL;DR

This paper proposes RecBench, a comprehensive evaluation framework that systematically compares 17 LLMs against 10 conventional DLRMs across 5 domain-specific datasets. Results show that LLM-based recommenders achieve up to 5% AUC improvement on CTR tasks and up to 170% NDCG@10 improvement on sequential recommendation, yet incur 10–1000× slower inference. Conventional DLRMs augmented with LLM semantic embeddings (LLM-for-RS) attain approximately 95% of LLM performance at 20× higher throughput, making this paradigm the most industrially viable solution at present.

Background & Motivation

Background: The integration of LLMs with recommender systems (LLM+RS) has attracted considerable research attention, giving rise to two paradigms: LLM-for-RS (LLMs as feature-enhancement plugins) and LLM-as-RS (LLMs serving directly as recommenders). The latter has demonstrated promise in cold-start and explainable recommendation scenarios, yet systematic evaluation remains lacking.

Limitations of Prior Work: Existing benchmarks (LLMRec, PromptRec, OpenP5, etc.) suffer from three major deficiencies: (a) they evaluate only a single recommendation formulation (pair-wise or list-wise); (b) item representation forms are insufficiently covered, typically relying solely on text or unique IDs; and (c) the number of evaluated models is limited, with inference efficiency entirely overlooked.

Core Problem: Does the accuracy advantage of LLMs on recommendation tasks sufficiently compensate for their substantial inference efficiency drawbacks? How do different item representation strategies affect LLM recommendation capability?

Key Insight: The paper constructs RecBench, the most comprehensive LLM recommendation evaluation benchmark to date, simultaneously assessing both accuracy and efficiency across 4 item representations, 2 recommendation scenarios, 27 models, and 5 datasets.

Method

Overall Architecture

The RecBench evaluation matrix spans: 5 datasets (H&M Fashion, MIND News, MicroLens Video, Goodreads Books, Amazon CDs Music) × 4 item representations (unique ID, text, semantic embedding, semantic identifier) × 2 recommendation tasks (CTR prediction, sequential recommendation) × 27 models (17 LLMs + 10 DLRMs), with both accuracy metrics and inference latency measured.

Four Item Representation Strategies

1. Unique Identifier: The conventional approach, assigning each item a randomly initialized embedding vector whose semantics are learned through collaborative filtering signals.
2. Text: Item titles or other textual features are used, with item representations obtained by averaging word embeddings—naturally compatible with LLMs' text comprehension capabilities.
3. Semantic Embedding: A pretrained LLM (e.g., Llama-1 7B) encodes item text into dense vectors used to initialize DLRM inputs, introducing rich general-purpose semantics.
4. Semantic Identifier: SentenceBERT first extracts item embeddings, which are then discretized via RQ-VAE into a 4-layer × 256-codebook encoding sequence. Semantically similar items share longer common subsequences, simultaneously compressing the vocabulary while preserving semantic relationships.

Two Recommendation Scenarios

Pair-wise Recommendation (CTR Prediction): Given a user–item pair, the model predicts click probability. Models are organized into six groups (A–F):

• Group A: Conventional DLRM + Unique ID (DNN, DeepFM, DCN, DCNv2, AutoInt, GDCN, etc.; 9 models)
• Group B: Conventional DLRM + Text (DNN_text, DCNv2_text, etc.; 4 models)
• Group C: Conventional DLRM + Semantic Embedding (DNN_emb, GDCN_emb, etc.; 4 models; LLM-for-RS paradigm)
• Group D: LLM + Unique ID (P5 series, with item IDs as special tokens)
• Group E: LLM + Text (GPT-3.5, Llama series, Qwen series, etc.; supporting both zero-shot and fine-tuned settings)
• Group F: LLM + Semantic Identifier (SID-BERT, SID-OPT)

List-wise Recommendation (Sequential Recommendation): Given a user's interaction history, the model predicts the next item. Models are organized into four groups (G–J), incorporating Constrained Beam Search (CBS)—a technique that leverages the semantic identifier tree to constrain decoding paths and ensure generated token sequences correspond to valid items.

Loss & Training

• LLM fine-tuning employs LoRA: rank=32/alpha=128 for CTR tasks; rank=128/alpha=128 for sequential recommendation
• Learning rates: 1e-4 for LLMs; 1e-3 for DLRMs
• All experiments are conducted on a single A100 GPU; results are averaged over 5 runs

Key Experimental Results

CTR Prediction (Pair-wise, AUC)

Item Representation	Representative Model	Overall AUC	CPU Latency (ms)	GPU Latency (ms)
Unique ID (best DLRM)	GDCN	0.6825	1.20	2.02
Text (best DLRM)	GDCN_text	0.6923	5.09	3.77
Semantic Embedding (best DLRM)	DNN_emb	0.7171	1.42	2.09
Text (best fine-tuned LLM)	Mistral-2 7B	0.7578	7680	76.14
Best zero-shot LLM	GLM-4 9B	0.6231	9690	83.38

Key Findings: Fine-tuned Mistral-2 achieves AUC of 0.7578, approximately 5.7% above the best DLRM (DNN_emb, 0.7171), yet is 5,400× slower on CPU inference (7,680 ms vs. 1.42 ms).

Sequential Recommendation (List-wise, NDCG@10)

Item Representation	Representative Model	Overall NDCG@10	CPU Latency (ms)
Unique ID (best DLRM)	SASRec_24L	0.0698	103.41
Unique ID (best LLM)	P5-BERT_base	0.1025	41.54
Semantic ID (best LLM+CBS)	SID-BERT_base-CBS	0.1877	1900
Semantic ID (large model+CBS)	SID-Llama-3 7B-CBS	0.1607	177540

Key Findings: SID-BERT_base-CBS achieves NDCG@10 of 0.1877, a 169% improvement over SASRec_24L (0.0698), at the cost of 18× longer inference time. SID-Llama-3 7B-CBS requires approximately 177 seconds per sample, rendering it completely infeasible for practical deployment.

Zero-Shot LLM Performance

Most LLMs achieve AUC near 0.50 on zero-shot CTR tasks (approaching random), with only Mistral (0.6199) and GLM-4 (0.6231) performing acceptably. Dedicated recommendation models RecGPT (0.4952) and P5_Beauty (0.5049) exhibit extremely poor zero-shot generalization. The Qwen-2 series demonstrates a positive correlation between model scale and zero-shot recommendation capability (0.5B→1.5B→7B: 0.5413→0.5707→0.6075).

Gains from Fine-Tuning

Instruction fine-tuning improves LLM CTR AUC by 22%–43% in relative terms. For example, Llama-3 8B improves from 0.5252 (zero-shot) to 0.7508 (fine-tuned).

Highlights & Insights

1. "LLM-for-RS" offers the best current trade-off: DLRM augmented with LLM semantic embeddings (Group C) achieves DNN_emb AUC=0.7171 at very low latency (~2 ms GPU)—approximately 94.6% of the best LLM's performance (0.7578)—while being 36× faster. This is the most practically deployable solution for industrial settings.

2. Semantic identifiers confer substantial advantages in sequential recommendation: SID representations enable even shallow networks to capture user interest patterns; SID-SASRec_3L-CBS (0.0306) substantially outperforms SASRec_3L (0.0096). However, the advantage diminishes as depth increases, suggesting that deep ID-based models can learn analogous information.

3. Constrained Beam Search (CBS) is a critical technique: CBS constrains decoding via the semantic identifier tree to ensure generation of valid items. SID-BERT_base improves from 0.0941 to 0.1877 with CBS, nearly doubling performance.

4. An abstract similarity exists between pretrained language patterns and user interest patterns: BERT_base with unique IDs (no text) achieves 0.1025 on sequential recommendation, surpassing same-architecture SASRec_12L (0.0672), suggesting that sequential patterns learned through language modeling transfer to user behavior modeling.

5. Model scale is not universally beneficial: In sequential recommendation, SID-BERT_base-CBS (0.1877) substantially outperforms SID-Llama-3 7B-CBS (0.1607), indicating that smaller models can outperform larger ones under specific configurations.

Limitations & Future Work

• Only single-sample inference latency is measured; the impact of batch inference and acceleration techniques such as KV-cache is not considered.
• The paper does not evaluate LLM advantages in specialized scenarios such as cold-start or cross-domain recommendation—arguably the most genuinely promising direction for LLM-as-RS.
• Semantic embeddings are produced solely by Llama-1 7B; the effect of stronger encoders (e.g., more recent Llama-3 or domain-specific models) remains unexplored.
• All five datasets are preprocessed and truncated to similar scales, which may not reflect real-world industrial data distributions.

Rating

• Novelty: ⭐⭐⭐⭐ — First benchmark to simultaneously cover accuracy and efficiency, 4 item representations, and 2 task formulations for LLM-based recommendation
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ — 27 models × 5 datasets × 4 item representations, results averaged over 5 runs; an uncommonly large-scale evaluation
• Writing Quality: ⭐⭐⭐⭐ — Experimental analysis is clear and well-organized, with persuasive conclusions
• Value: ⭐⭐⭐⭐⭐ — Provides important strategic guidance for the recommender systems community by establishing a clear finding that LLM-for-RS outperforms LLM-as-RS

Related Papers

• [NeurIPS 2025] Inductive Transfer Learning for Graph-Based Recommenders
• [ICLR 2026] Can Speech LLMs Think while Listening?
• [ACL 2025] Who Can Withstand Chat-Audio Attacks? An Evaluation Benchmark for Large Audio-Language Models
• [ICML 2025] Teaching Physical Awareness to LLMs through Sounds
• [ACL 2026] Jamendo-MT-QA: A Benchmark for Multi-Track Comparative Music Question Answering