Technical Summary: Learning to Translate from Soft to Hard LLM Prompts

Problem Statement

Soft prompt tuning offers a parameter-efficient alternative to full fine-tuning for adapting Large Language Models (LLMs) to specific tasks. However, a significant limitation of soft prompts is their lack of interpretability. Unlike natural language tokens, soft prompts exist as continuous vectors in the embedding space, often occupying regions distinct from natural language. Naive projections of these vectors into discrete token space frequently yield misleading, contradictory, or uninformative results (Khashabi et al., 2022).

Recent work, specifically the InSPEcT technique (Ramati et al., 2024), attempted to address this by using a training-free, in-context learning approach to elicit textual verbalizations of soft prompts. While InSPEcT can sometimes produce correct class labels, the authors of this paper observe that its verbalizations are often unreliable, sensitive to specific architectural hyperparameters (layer pairs), and frequently contain unnatural language. Furthermore, InSPEcT's performance appears heavily dependent on the underlying soft prompt achieving high task accuracy, limiting its utility when soft prompts are trained on smaller or less capable models.

Methodology

The authors propose a supervised learning approach to translate soft prompts into natural language ("hard prompts") using a dedicated translation model. The core workflow involves three stages:

1. Soft Prompt Training: Soft prompts are first optimized on a small, open-source LLM (specifically meta-llama/Llama-3.1-8B-Instruct) for various downstream tasks using standard supervised learning on (input, output) pairs.
2. Dataset of Datasets (DoD) Construction: To train the translator, the authors constructed two distinct "Datasets of Datasets" (DoDs) containing pairs of (soft prompt, ground-truth hard prompt):
   • Classification DoD: A synthetic dataset comprising 5,500 mini-datasets with tightly semantically clustered class labels (e.g., {music, films, books, art, software}) and 2.75 million sentences.
   • General DoD: Based on the Super-NaturalInstructions dataset, covering 1,616 diverse natural language tasks (summarization, translation, reasoning, etc.) across 76 task types.
3. Translator Training: A dedicated translation model, sharing the same architecture as the base LLM but with different weights, is fine-tuned using LoRA. The model treats soft prompt translation as a sequence-to-sequence generation task, where the input is the soft prompt embedding vector and the output is the corresponding natural language task description.

Key Contributions

The paper makes three primary contributions:

1. A Dedicated Translator Model: The authors train a supervised model to convert soft prompts into fluent, grammatically correct natural language. This approach outperforms the training-free InSPEcT method in both quantitative metrics (Recall, F1-Score, ROUGE-L) and qualitative assessments across multiple datasets.
2. Demonstrated Utility of Verbalizations: The authors show that the translated "hard prompts" are not merely plausible-sounding text but are functionally relevant. When these verbalized prompts are deployed on larger, closed-source LLMs (e.g., GPT-4o-mini), they achieve task performance that often exceeds the original soft prompt's performance on the small model and, in specific cases, surpasses few-shot learning baselines.
3. A New Workflow for Portable Prompts: The work proposes a novel pipeline where a soft prompt is trained on a small, open-source model, translated into a human-readable prompt, and then deployed on a larger, more powerful model. This allows for the transfer of task-specific knowledge from smaller models to larger ones without requiring the larger model to be fine-tuned.

Results

• Interpretability Quality: On the InSPEcT datasets (SST-2, SST-5, AG-News, Subj, TREC), the translator achieved perfect or near-perfect Recall (1.0) and significantly higher F1-Scores compared to InSPEcT. Notably, the translator maintained robust performance on datasets where InSPEcT struggled (e.g., TREC), suggesting it is less dependent on the soft prompt's initial task accuracy.
• Generalization: On the General DoD, the translator achieved an average ROUGE-L of 0.252 and cosine similarity of 0.427, outperforming InSPEcT (0.143 ROUGE-L, 0.269 cosine similarity).
• Downstream Performance: In experiments where verbalized prompts were used with a large closed-source LLM, the translated prompts often outperformed the original soft prompts used on the small model. In a specific subset of 10 tasks, the translated prompts even outperformed few-shot learning on the large model, particularly for tasks requiring nuanced pattern recognition that few-shot examples could not capture efficiently.
• Data Efficiency: The translator demonstrated the ability to learn the correspondence between soft and hard prompts even with a relatively small training set (539 samples for the General DoD), suggesting that LLMs may innately understand the "language" of soft prompts.

Significance and Claims

The paper claims that soft prompts are not fundamentally uninterpretable but rather require a dedicated translation mechanism to be decoded effectively. By moving away from in-context learning (InSPEcT) toward supervised translation, the authors demonstrate that soft prompts can be converted into portable, high-quality natural language instructions.

The authors suggest this workflow has significant potential for "downstream applications," specifically enabling the use of small, open-source models to discover and optimize task instructions that can then be transferred to larger, closed-API models for superior inference. They note that while the current translator is a promising proof of concept, future work could involve larger datasets and stylistic alignment (e.g., preference optimization) to ensure verbose and detailed verbalizations, potentially unlocking insights into complex tasks like drug discovery.

The authors remain modest about the current limitations, acknowledging that the translator is currently tied to the specific base model used for training and that verbalizations can sometimes be overly terse, omitting relevant details due to stylistic misalignment.