LLM Personalization: SFT vs. ICL Under Congestion

New research explores the strategic choices users and platforms face when personalizing Large Language Models (LLMs), specifically comparing Supervised Fine-Tuning (SFT) and In-Context Learning (ICL). While personalization enhances model performance, it consumes shared computational resources, creating a tension. The study develops a framework to analyze these statistical-economic trade-offs, considering how congestion from other users' choices impacts individual incentives. The analysis yields several counter-intuitive insights. It shows that SFT and ICL dominate in different scenarios, depending on factors like pretraining coverage and data signal-to-noise ratios. However, network congestion can dramatically alter these optimal choices. Furthermore, the study found that equilibrium resource consumption doesn't always behave predictably; improving pretraining precision can reduce congestion, while broader pretraining coverage or harder tasks might paradoxically increase it. Crucially, the research proves that platforms benefit from offering both personalization methods, as it never diminishes their maximal profits, even if it potentially increases computational load. This theoretical finding is supported by empirical observations: a review of 21 major AI platforms shows a significant increase in those offering both SFT and ICL, rising from 9.5% in 2021 to 71.4% in 2025, aligning with the platform design implications of the study.