Prompt-tuning vs LoRA: pick the architectural reason LoRA tends to win

Attempt it

Quick reference

TL;DR

Prefix and prompt tuning only steer the model through input vectors; LoRA edits the weight matrices directly, so it adapts harder per parameter and merges to zero inference cost.

Memory aid

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

Easy to grasp

Imagine a huge orchestra playing a fixed score. Prompt tuning is like whispering instructions to the conductor before the piece starts. The musicians never change, you just nudge what they hear at the very beginning, and that nudge has to ripple all the way through. LoRA is like handing every section a small set of edits to their own sheet music. The cellos, the brass, the percussion each get tweaked directly, so the whole sound shifts with far less effort. Whispering to the conductor can only do so much. Editing the parts themselves gives you real control. And once LoRA's edits are pencilled into the score, the performance costs nothing extra to play.

Key concepts

LoRA adapters Prefix tuning Parameter-efficient fine-tuning

Quick scan · 2–4 min

Everything important, quickly.

4 min: where each method injects parameters + the input-space bottleneck + LoRA's per-layer leverage + the merge to zero cost perk + when soft prompts still win + LoRA vs full fine-tuning.

Side-by-side

Concern	Prefix / prompt tuning	LoRA
Where it edits	Input or key-value vectors only	Attention and MLP weight matrices
Weights changed	None; base stays frozen	Low-rank delta on chosen matrices
Representational ceiling	Capped by input-space bottleneck	Much higher; per-layer leverage
Inference cost	Extra key-value tokens every request	Zero after merging into base
Typical use	Light task switching, many tiny adapters	Serious behavior and reasoning change

Real-world examples

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

Hugging Face PEFT ships LoRA, prefix tuning, prompt tuning, and P-tuning side by side, and LoRA is by far the most downloaded adapter type for Llama 4 fine-tunes.
QLoRA fine-tunes of Llama 4 and DeepSeek V4 on a single consumer GPU are the default community recipe, with prefix tuning rarely chosen for serious behavior change.
Multi-tenant inference servers like vLLM and LoRAX hot-swap per-customer LoRA adapters at request time, exploiting LoRA's merge-and-serve property.
Provider fine-tuning APIs for GPT-5.5 and Gemini 3.1 Pro expose low-rank adapter tuning rather than soft-prompt tuning for customer behavior changes.
Frameworks like Axolotl and Unsloth default to LoRA or QLoRA recipes for adapting Claude Opus 4.7-class open-weight competitors on domain tasks.

Interviewer follow-ups

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

QWhy does adding more soft-prompt tokens eventually stop improving quality?

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

QHow does merging a LoRA adapter give zero inference overhead, and when can you not merge?

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

QIf LoRA edits weights directly, why does it ever lose to full fine-tuning?

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.