LLMs 28. Learn Prompting
Quick Overview
This guide explains prompt learning for large language models, covering motivation, core concepts and methods such as continuous prompt embeddings and prefix-tuning, task reframing to align downstream tasks with pretraining objectives, and considerations for scaling and generalization beyond NLP.
Prompt Learning: Teaching Large Models How to Use What They Already Know
Prompt Learning sits at an interesting intersection between model capability and human instruction. Instead of changing what a pretrained model knows, it changes how that knowledge is accessed. This shift—from modifying parameters to shaping inputs—has quietly redefined how we adapt large language models (LLMs) under real-world constraints.
This post walks through why Prompt Learning emerged, how the major methods differ, and how these ideas generalize beyond NLP tasks.
Why is Prompt Learning needed?
As models scale up, full fine-tuning becomes increasingly impractical. Updating billions of parameters for every downstream task is not just expensive—it is often unnecessary. At the same time, simply freezing most layers and tuning a few parameters near the output rarely delivers strong results.
Prompt Learning emerged as a middle ground. Instead of changing the model itself, it reframes tasks in a way the model already understands. The key insight is that many failures are not due to missing knowledge, but due to misalignment between the task and the model’s pretraining objective.
What is Prompt Learning?
Prompt Learning treats the input prompt as a trainable interface between the user and the pretrained model. Rather than writing prompts as fixed text, we learn continuous prompt embeddings that guide the model toward correct behavior.
For example, a sentiment classification task can be reframed as a masked-language prediction problem (“This review is [MASK].”), allowing the model to reuse its language understanding instead of learning a new classifier from scratch. Prompt Learning formalizes and optimizes this idea.
Why Prompt Learning works surprisingly well
Large pretrained models already encode vast linguistic and world knowledge. Prompt Learning exploits this by:
- minimizing parameter updates,
- preserving pretrained representations,
- and aligning downstream tasks with pretraining objectives.
As models grow larger, this effect becomes more pronounced. Empirically, Prompt Learning often scales better with model size than traditional fine-tuning, even though it trains far fewer parameters.
The main Prompt Learning methods, explained intuitively
Prompt Learning is not a single technique, but a family of approaches that differ in where and how prompts are injected.
Prefix-Tuning: guiding the model at every layer
Prefix-Tuning was proposed to address the brittleness of manual prompts and the cost of full fine-tuning. Instead of discrete tokens, it learns a sequence of virtual prefix tokens represented as continuous vectors.
These prefixes are injected into every Transformer layer, influencing attention at all depths. During training, the base model is frozen; only the prefix parameters are updated via a small MLP that maps them into the model’s hidden space.
This makes Prefix-Tuning powerful for generation tasks, especially with GPT-style models. However, adding prompts at every layer increases both parameter count and inference cost, since the effective sequence length grows.
Prompt-Tuning: learning prompts only at the input
Prompt-Tuning simplifies Prefix-Tuning by placing learnable prompt embeddings only at the input layer. No layer-wise injection, no extra MLPs—just trainable vectors concatenated with token embeddings.
This simplicity brings clear benefits. Prompt-Tuning is lightweight, easy to store, and works particularly well as model size increases. It also enables prompt ensembling: training multiple prompts for the same task to improve robustness without duplicating the model.
The downside is stability. Training can be sensitive, performance drops on smaller models, and the method struggles with complex structured prediction tasks such as sequence labeling.
Prompt-Tuning vs Prefix-Tuning (in practice)
The distinction is less about “better” and more about control depth.
Prefix-Tuning influences internal representations throughout the network, making it stronger for generation-heavy tasks. Prompt-Tuning nudges the model only at the input, relying on the pretrained stack to propagate that signal.
A useful mental model:
- Prefix-Tuning reshapes the computation path.
- Prompt-Tuning reshapes the starting conditions.
P-Tuning: learning structured prompts instead of independent tokens
Early Prompt-Tuning treats prompt tokens as independent parameters. P-Tuning improves on this by introducing a prompt encoder, typically a bidirectional LSTM followed by an MLP, to model dependencies between prompt tokens.
This makes prompt initialization more stable and expressive. Instead of learning isolated vectors, the model learns structured prompt representations, which significantly improves performance on NLU tasks, especially when manual prompting is unreliable.
The trade-off is interpretability. Once prompts are encoded through neural networks, they no longer correspond cleanly to human-readable instructions.
P-Tuning v2: scaling Prompt Learning to real tasks
P-Tuning v2 was introduced to address two key weaknesses of earlier methods: poor performance on small models and limited applicability beyond classification.
It combines ideas from Prefix-Tuning and P-Tuning:
- prompts are injected at every layer (deep prompts),
- prompt encoders are parameter-efficient and stable,
- prompt length is task-specific rather than fixed,
- and prompts can be pretrained across multiple tasks.
Importantly, P-Tuning v2 removes the traditional verbalizer and uses a classification head instead, extending prompt-based learning to tasks like NER and sequence labeling.
This comes at a cost: interpretability is reduced, and the method is more complex. But it achieves performance much closer to full fine-tuning while remaining parameter-efficient.
Prompt Learning vs Fine-Tuning: a deeper distinction
Fine-tuning changes the model. Prompt Learning changes the interaction with the model.
Fine-tuning risks catastrophic forgetting but can rewrite internal knowledge. Prompt Learning preserves the pretrained model intact, optimizing how tasks are expressed rather than what the model knows.
This distinction matters beyond NLP. In any system where a powerful pretrained backbone exists, Prompt Learning-style ideas encourage us to ask:
Can we reframe the problem instead of rewriting the system?
When Prompt Learning is a good (or bad) idea
Prompt Learning excels when:
- compute or memory is limited,
- tasks are closely aligned with language understanding,
- model size is large,
- or rapid task switching is required.
It struggles when:
- tasks require deep structural changes,
- training data is extremely small and noisy,
- or interpretability of learned prompts is critical.
In practice, Prompt Learning often complements other techniques such as LoRA, adapters, or RAG, rather than replacing them.
Closing perspective
Prompt Learning changed how we think about adaptation. It showed that how we ask can be as important as what we train. More broadly, it reflects a shift in ML engineering—from rewriting models to designing interfaces that unlock existing capabilities.
Understanding Prompt Learning is less about memorizing methods and more about internalizing this design philosophy.
Comments (0)