Implicit Reasoning for Large Language Model-based Generative Recommendation
Quick Answer
PauseRec introduces a lightweight implicit reasoning framework for LLM-based Generative Recommendation, outperforming explicit CoT methods by 6.22%, reducing training costs by 65% GPU hours, and accelerating inference by 71.3%.
Quick Take
This approach addresses limitations in existing reasoning pipelines, enhancing efficiency and effectiveness in leveraging pretrained knowledge.
Key Points
- PauseRec outperforms explicit reasoning methods by 6.22% in performance.
- Training costs are reduced by up to 65% in GPU hours.
- Inference speed is accelerated by up to 71.3% with PauseRec.
- Existing pipelines struggle with world-knowledge verbalization and SID alignment.
- PauseRec avoids costly reasoning trace acquisition and alignment training.
Paper Resources
Source Excerpt
arXiv:2606. 14142v1 Announce Type: new Abstract: (LLMs) are increasingly adopted as backbones for Generative Recommendation (GR), promising access to pretrained world knowledge. Yet reliably invoking this knowledge for GR remains poorly understood. A key obstacle is that LLM-based GR typically represents items with Semantic IDs (SIDs), disrupting LLMs' natural-language reasoning interface because these tokens are unseen by the LLM during pretraining.
Existing approaches address this with expensive multi-stage pipelines that ground SIDs and elicit explicit rationales, but offer limited insight into when and why each stage is necessary. …
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →TriAgent: Divergence-Aware Committees for Cost-Efficient Financial Sentiment Analysis
TriAgent introduces a cost-efficient multi-agent system for financial sentiment analysis, combining VADER, FinBERT, and Qwen2.5. It achieves an F1 score of ~0.87 with significant savings of $9.3M/year at a 10M-user scale compared to GPT-4o-mini, while also detecting hallucinations with an AUC of 0.90.
