On the Scaling of PEFT: Towards Million Personal Models of Trillion Parameters

This Mind Lab paper reframes parameter-efficient fine-tuning, especially LoRA, as a mechanism for persistent local state on top of shared foundation models, aiming at “million personal models of trillion parameters.” Its core claim is that PEFT becomes transformative only when three scaling axes reinforce each other: Scale Up, where stronger shared priors make small adapters more powerful; Scale Down, where the adaptive state becomes small and stable enough to learn reliably; and Scale Out, where many persistent adapted instances can coexist. On Scale Up, the authors show trillion-scale LoRA reinforcement learning on a 1.04T-parameter sparse MoE model, Kimi K2, using GRPO-style on-policy optimization with hybrid parallelism, reducing compute and communication to about 10% of conventional full-parameter RL while preserving stable reward and task-success curves. They also identify training–inference mismatch and routing divergence in MoE systems as new failure modes that require router replay and provenance-aware serving. On Scale Down, a Qwen3-8B PPO sweep over 216 runs finds that ranks 16 and 32 are the most reliable middle regime, while ranks 1–4 retain strong best-case performance but suffer seed sensitivity; an RL-native initialization, OLoRA-tail, stabilizes rank-1 training and improves DeepSeek-R1-Distill-Qwen-1.5B benchmark accuracy from 56.3% to 58.3% average, with a larger win on Qwen3-30B-A3B-Instruct, where it reaches 35.5% versus 24.0% for standard LoRA. On Scale Out, the paper introduces LoRA-as-memory and Context Learning, reports a bounded memory capacity law on DishNameBenchmark, and shows per-user LoRA adapters improve social simulation in OASIS and majority-vote accuracy on AIME24 from 0.3644 at one model to 0.4867 at 198 models. The infrastructure layer, MinT from Mind Lab, turns adapter identity, revisioning, and bounded residency into a deployable lifecycle for persistent personal models.