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

Both **prompt tuning** and **LoRA** are cheap ways to adapt a big model without retraining all of it. They differ in WHERE they make their change. Prompt tuning and prefix tuning learn a handful of new vectors and glue them onto the input, like invisible extra tokens at the front. The model's actual weights never move. Those learned tokens can only nudge the model by being something it reads. LoRA instead learns tiny add-on matrices that plug into the model's own weight matrices, mostly inside attention. After training you can fold them in. So LoRA changes how the model transforms information, not just what it is shown. That difference matters. Editing the input can only do so much, the rest of the network still runs exactly as before. Editing the weights reaches every later token and layer. At the same parameter budget, LoRA usually scores higher, which is why production largely moved to it by 2025. Soft prompts still help when you serve thousands of tasks on one model, or when you only control the input of a closed API.

**Soft-prompt methods** like prompt tuning, prefix tuning, and P-tuning v2 learn continuous vectors and inject them into the input or the attention prefix. The frozen model then has to interpret those vectors. The adaptation only acts through the attention lens, so there is a ceiling on how much input-only changes can buy. LoRA edits the model's transformations directly. It adds a low-rank delta to weight matrices in attention, and with all-linear targeting the MLPs too. So it reshapes the functions the network computes rather than just the tokens it reads. That expressivity gap is the core reason. At a matched parameter budget LoRA wins on most benchmarks, and a merged LoRA adds zero inference latency. Soft prompts instead consume context length on every single call. Soft prompts keep two real niches. First, massive multi-task serving, where a short prefix per task is tiny and trivially swappable. Second, closed-API settings where you cannot touch weights at all. P-tuning v2 also stays competitive on classification tasks where a fixed prefix fits the job.