[Paper Review] P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks

Xiao Liu, Kaixuan Ji+5 more|arXiv (Cornell University)|Oct 14, 2021

Topic ModelingComputer Science46 references262 citations

TL;DR

P-Tuning v2 shows that carefully optimized prompt tuning with deep prompts across all layers can match fine-tuning across model scales (300M–10B parameters) and NLU tasks while using only a small fraction of task-specific parameters.

ABSTRACT

Prompt tuning, which only tunes continuous prompts with a frozen language model, substantially reduces per-task storage and memory usage at training. However, in the context of NLU, prior work reveals that prompt tuning does not perform well for normal-sized pretrained models. We also find that existing methods of prompt tuning cannot handle hard sequence labeling tasks, indicating a lack of universality. We present a novel empirical finding that properly optimized prompt tuning can be universally effective across a wide range of model scales and NLU tasks. It matches the performance of finetuning while having only 0.1%-3% tuned parameters. Our method P-Tuning v2 is an implementation of Deep Prompt Tuning \cite{li2021prefix,qin2021learning} optimized and adapted for NLU. Given the universality and simplicity of P-Tuning v2, we believe it can serve as an alternative to finetuning and a strong baseline for future research.Our code and data are released at https://github.com/THUDM/P-tuning-v2.

Motivation & Objective

• Demonstrate that properly optimized prompt tuning can match fine-tuning across a wide range of model sizes and NLU tasks.
• Develop and validate a deep prompt tuning approach that uses prompts in multiple layers rather than just the input layer.
• Identify practical optimization and implementation details that enable competitive performance with limited task-specific parameters.

Proposed method

• Adopt deep prompt tuning by inserting trainable prompts at multiple transformer layers, not only at the input.
• Use continuous prompts with a flexible prompt length per task, generally shorter for simple tasks and longer (≈100) for hard sequence labeling tasks.
• Experiment with reparameterization options (embedding vs MLP) and analyze their task-dependent effects.
• Apply a randomly initialized linear classification head instead of a verbalizer in a supervised setting.
• Evaluate multi-task learning by jointly training shared prompts across tasks while using task-specific classifiers.

Experimental results

Research questions

• RQ1 Can prompt tuning with prompts across many layers achieve performance comparable to fine-tuning across diverse model scales (300M–10B parameters)?
• RQ2 Is P-Tuning v2 universally effective for both simple classification and hard sequence labeling NLU tasks?
• RQ3 What optimization and implementation details are necessary to close the gap between prompt tuning and fine-tuning, and how does depth and prompt length affect performance?
• RQ4 Does multi-task learning with shared prompts provide additional gains for P-Tuning v2 across tasks?

Key findings

• P-Tuning v2 matches fine-tuning across model scales from 300M to 10B parameters.
• P-Tuning v2 achieves competitive performance with only 0.1%–3% task-specific parameters per task.
• Deep prompts introduced to multiple layers improve performance, especially for smaller models and hard tasks.
• Multi-task learning generally boosts P-Tuning v2 performance across most tasks, except QA in some settings.
• Ablation studies show no significant difference between verbalizers and a linear head in supervised settings, and deeper prompt depth tends to help more than shallow prompt insertion.