Why has prefix/prompt tuning largely lost to LoRA in production?

Attempt it

Prefix-tuning and prompt-tuning were once the leading PEFT methods. By 2024-2025, LoRA had largely displaced them in production. Explain the architectural reason LoRA tends to win at similar parameter budgets, and where soft-prompt methods still make sense.

Free · 2 AI evals / day

Quick reference

TL;DR

Soft prompts only edit the input; LoRA edits the weight matrices themselves. That extra leverage lets LoRA win at the same parameter budget.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Picture a chef who already knows how to cook. Prompt tuning is like sliding a fixed note onto the order ticket before every dish, the chef reads it but cooks with the same hands and habits. The note can only hint; it cannot rewire how the chef chops or seasons. LoRA is like a quick training session that actually adjusts a few of the chef's techniques. Now the change lives in the hands, not in a sticky note that has to be re-read every time. Because LoRA touches the technique and the note only touches the order, LoRA can change more for the same effort. The sticky note still helps when you run one kitchen for thousands of customers and only want to swap tiny notes.

Key concepts

LoRA low-rank adaptation Prefix and prompt tuning Parameter-efficient fine-tuning Multi-tenant adapter serving

Quick scan · 2–4 min

Everything important, quickly.

5 min: where each method injects parameters + the input-only representational ceiling + why weight edits beat input edits + serving and KV-cache asymmetry + the multi-tenant and closed-API niches + P-tuning v2 on classification.

Side-by-side

Aspect	Soft prompts (prefix / prompt / P-tuning)	LoRA
What it edits	Input or attention prefix vectors	Weight matrices (attention, optional MLP)
Model weights	Frozen, untouched	Low-rank delta added, can merge
Expressivity	Capped by input-only adaptation	Reshapes transforms, higher ceiling
Inference cost	Spends context tokens every call	Zero overhead once merged
Best niche	Multi-tenant, closed API, classification	Most behavior and domain fine-tunes

Real-world examples

Real products, models, and research that use this idea.

vLLM and LoRAX serve hundreds of merged or hot-swapped LoRA adapters on one base in 2026 production stacks, a pattern S-LoRA and Punica pioneered.
Hugging Face PEFT ships LoRA, prefix tuning, prompt tuning, and P-tuning under one API, and its docs treat LoRA as the default modern choice.
Teams fine-tuning open bases like Llama 4 or Qwen with Unsloth and Axolotl reach for LoRA or QLoRA by default rather than soft prompts.
Closed-weight frontier models such as GPT-5.5 and Claude Opus 4.7 expose only input-level control, the exact setting where soft-prompt style conditioning still applies.
P-tuning v2 remains a reference baseline for GLUE-style classification, where a per-layer prefix fits the fixed-label task well.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QWhy doesn't simply adding more soft-prompt tokens close the gap with LoRA?

Sign in or go Pro to see the approach for this follow-up.

QHow does P-tuning v2 differ from vanilla prompt tuning, and why does that help?

Sign in or go Pro to see the approach for this follow-up.

QIn a multi-tenant platform with thousands of tasks, when would you still prefer LoRA over soft prompts?

Sign in or go Pro to see the approach for this follow-up.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.