Large Scale Retrieval for the LinkedIn Feed Using Causal Language Models (arXiv, Oct 2025)

Overview

This entry covers the arXiv paper "Large Scale Retrieval for the LinkedIn Feed using Causal Language Models" (arXiv:2510.14223, October 2025), authored by Sudarshan Srinivasa Ramanujam, Antonio Alonso, Saurabh Kataria, Siddharth Dangi, Akhilesh Gupta, Birjodh Singh Tiwana, and 17 other contributors from LinkedIn.

Context and Motivation

In large-scale search, recommendation, and personalization systems, information retrieval has long faced challenges around efficiency, scalability, and user-intent understanding. Traditional pipelines treat retrieval, ranking, and generation as separate stages, which struggles to meet modern expectations for natural-language interaction, multi-hop reasoning, and real-time knowledge. This paper targets that intersection, applying causal language models to retrieval for the LinkedIn feed.

Position in the Field

Neural information retrieval has evolved through several generations, each balancing effectiveness, efficiency, and maintainability:

• BM25 / sparse retrieval — robust baseline, weak semantic matching
• Dense bi-encoders + ANN search — millisecond-scale recall, sensitive to domain shift
• Cross-encoders — higher precision, but document representations cannot be precomputed
• Generative retrieval and LLM-based approaches — reduce cascade errors but introduce index-update and cost challenges
• Agentic search / RAG — retrieval becomes iterative, plannable, and verifiable, shifting evaluation toward task success and citation accuracy rather than static nDCG

Within this landscape, the paper's contribution is applying decoder-only causal language models to a production feed-retrieval setting at LinkedIn scale.

Engineering Considerations for Industrial Deployment

| Concern | Guidance | |---------|----------| | Latency | Cascade + early stopping, cache hot queries, budget-aware reranking | | Data | PII handling, versioned and rollback-capable embedding indexes | | Quality | Verify offline gains translate to online CTR/satisfaction via interleaving experiments | | Safety | Source whitelisting, poisoning/bias detection, output filtering | | Cost | Model routing, distillation, hybrid sparse-dense retrieval |

Limitations and Open Questions

Typical open issues for this class of work include the gap between offline benchmarks and real user distributions, inference cost at production scale, hallucination and safety risks, and the reliability of LLM-as-judge evaluation, which should be cross-validated with human assessment.

Related Reading

• CAME: Competitively Learning a Mixture-of-Experts Model for First-stage Retrieval
• CoEvo: Coevolution of LLM and Retrieval Model for Domain-Specific Information Retrieval
• DRAMA: Diverse Augmentation from Large Language Models to Smaller Dense Retrievers (arXiv:2502.18460)
• ExpandR: Teaching Dense Retrievers Beyond Queries with LLM Guidance
• Fine-Tuning LLaMA for Multi-Stage Text Retrieval (arXiv:2310.08319)
• How Does Generative Retrieval Scale to Millions of Passages? (Google Research, arXiv:2305.11841)

> Note: The original forum post reproduces metadata and framing rather than full experimental results. Quantitative findings should be verified against the original PDF.