Semantic Search At LinkedIn, LLM-Driven Autonomous Optimization for Industrial-Scale Recommendation Systems, and More!

2025-01-31 · Source: Top Information Retrieval Papers of the Week · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Software Development & Engineering · Depth: Advanced, medium

Summary

This week's information retrieval newsletter highlights ten research papers covering advancements in semantic search, embedding techniques, and recommendation systems. LinkedIn presents a production-scale LLM-based ranking system achieving 75× throughput improvement with a 0.6B Small Language Model. NAIST investigates embedding magnitude in contrastive learning, finding it correlates with relevance in text retrieval. Perplexity AI introduces pplx-embed, multilingual text embedding models using diffusion-based pretraining and native INT8 quantization. A study by Benigni et al. exposes reproducibility failures and conceptual flaws in diffusion-based recommender models. Vančura et al. propose learning sparse high-dimensional embeddings for collaborative filtering, reducing memory by up to 10×. Google details a self-evolving recommendation system at YouTube using LLM agents for autonomous model optimization. Tencent introduces Rec2PM for efficient long-sequence generative recommendation via Preference Memory tokens. ByteDance's TokenMixer-Large scales industrial ranking models to 15 billion parameters. Meta presents GR2, an LLM framework for recommendation re-ranking, and Kunlun, a unified architecture for scaling massive-scale recommendation systems.

Key takeaway

For AI Scientists developing large-scale recommendation or search systems, you should critically evaluate the computational efficiency and scalability of your chosen architectures. Focus on techniques like multi-teacher distillation, sparse embeddings, and LLM-driven autonomous optimization to achieve production-scale throughput and maintain performance, especially when dealing with long user histories or massive parameter counts. Be wary of unproven methods, such as diffusion recommenders, which may lack reproducibility and practical benefits.

Key insights

Recent advancements in information retrieval focus on LLM-driven optimization, efficient embeddings, and scalable recommendation systems.

Principles

• Embedding magnitude carries task-relevant information beyond angular similarity.
• Diffusion models for recommenders often lack reproducibility and conceptual fit.
• Autonomous LLM agents can optimize complex ML systems end-to-end.

Method

Methods include multi-teacher distillation for compact LLMs, learnable normalization for embedding magnitudes, diffusion-based pretraining for embeddings, gradual pruning for sparse embeddings, and dual-agent LLM architectures for autonomous system optimization.

In practice

• Use 0.6B LLMs with distillation for high-throughput semantic search.
• Consider embedding magnitude for improved out-of-domain text retrieval.
• Employ native quantization-aware training for 4× embedding storage efficiency.

Topics

• LLM Ranking
• Text Embeddings
• Industrial Recommendation Systems
• Diffusion Models
• Scaling Laws

Code references

• remaplab/TORS26_Reproducibility-DDPMs
• zombak79/compressed_elsa
• benchen4395/KuaiSearch

Best for: NLP Engineer, AI Scientist, Research Scientist, AI Engineer, Machine Learning Engineer, AI Researcher

Related on AIssential

• A Production System for Podcast Discovery, A Fully Open-Source Frontier Search Agent, and More! — Recent research advances information retrieval through LLMs, sparse autoencoders, and multimodal embeddings.
• 2025 Year in Review — Integrating LLMs with RecSys can yield powerful, steerable recommendation systems capable of cross-platform suggestions.
• Inside Target’s LLM-Based System for Semantic Matching in Marketing Forecast Pipelines — Target's system utilizes LLMs and semantic embeddings to automate and improve marketing campaign forecasting by matching historical data.
• Ranking Millions of Candidates with LLMs via Aggregated Embeddings, Converting Single-Vector Retrievers into Multi-Vector Models Without Retraining, and More! — Information retrieval research is rapidly advancing, focusing on LLM integration, multi-vector models, and efficiency at scale.
• Increase Recommendation Systems’ Precision with LLMs, Using Python — Optimize LLM-powered recommendation systems by combining cheap, high-recall filtering with expensive, high-precision reranking.
• Understanding the Behaviors of Environment-aware Information Retrieval — LLMs can learn to adapt query strategies for diverse retrievers using RL, revealing distinct optimal query styles.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Top Information Retrieval Papers of the Week.