New Paper: GPU Congestion Can Reverse the SFT vs ICL Personalization Tradeoff
zainhas · x · 2026-09-20
A new paper, "Supervised Fine-Tuning vs. In-Context Learning: An Equilibrium Analysis of LLM Personalization under Congestion," argues that when shared GPU capacity is limited, choosing between SFT and ICL for personalization is not just a quality-cost tradeoff but a congestion game.
Key points:
- Which method wins depends on pretraining coverage and the SNR of personalization data; congestion can reverse the ranking.
- Better pretraining precision lowers congestion, but broader coverage can sometimes increase it.
- Offering both SFT and ICL never reduces a platform's maximum profit.
The work bridges LLM personalization, mean-field games, and platform design.
More from Research
- New paper: recursive looping boosts pre-training scaling exponents — burny_tech · 2026-09-20
- FutureHouse unveils "Bio Millennium Problems": hard-but-verifiable biology benchmarks for AI — _sholtodouglas · 2026-09-20
- Gated Recurrent Transformers: 3-layer recurrent model matches 12-layer GPT-2 with 63% fewer params — burny_tech · 2026-09-20
- Stanford builds virtual biotech run by 37,000 AI scientist agents, with real drug discovery results — Dr_Singularity · 2026-09-20
- A year after questioning LLM conjecture-solving, Cornell professor concedes AI surged ahead — burny_tech · 2026-09-20
- VLA-Replica: a $-low-cost real-world VLA benchmark where NVIDIA GR00T N1.7 matches π₀ with 50 demos — YuXiang_IRVL · 2026-09-20