Research map / Large language models / Language, safety and interpretability

Adaptive Prompt Routing for Arbitrary Text Style Transfer with Pre-trained Language Models

AAAI 2024 · Large language models › Language, safety and interpretability · approach: In-context learning analysis · 10 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Adaptive Prompt Routing (APR), which trains a BERT-based router to pick, per input, the best human-readable prompt from a paraphrased candidate set for a frozen LLM doing arbitrary text style transfer.

Problem

Fixed prompts (manual or searched) are suboptimal for arbitrary text style transfer because LLMs are prompt-sensitive and input texts are linguistically diverse; searched prompts are also unreadable.

Approach

Build a candidate set of human-readable prompts (seed prompts paraphrased by GPT-3.5), label each prompt per input sentence via a style-transfer score (dynamic mean+std threshold), and train a BERT router that predicts which prompt is optimal for each input; then sample multiple LLM outputs and rerank them with the same score.

Key results

• On GYAFC with GPT-2-XL, APR reaches J 47.8 and GM 79.8 versus 34.5 J and 72.1 GM for Prompt-and-Rerank, a 39% relative joint-score gain.
• With GPT-J-6B, APR averages J 46.1 and GM 77.5 in the dynamic-selection ablation across the 4 datasets, and reports GM 69.9 on Yelp and 73.2 on Shakespeare vs 71.3 for P&R.
• Human evaluation on Amazon (50 sentences) gives APR the highest content and fluency scores among non-reference systems (Con 4.65, Flu 4.77) and GM 4.29 vs 4.19 for P&R.

Measured effects reported in the paper

Surprising findings

• Fixed prompts selected by style-transfer score perform worse than adaptive routing, while random per-input selection performs far worse, showing that prompt choice matters per input.
• APR does not show a clear advantage over baselines on Amazon with GPT-2-XL under automatic metrics, motivating the human evaluation.

Limitations

• Relies on a style classifier trained on only 100 samples per style and on GPT-3.5 for paraphrasing, plus manual filtering of prompts.
• Sampling K outputs per input (30 for GPT-2-XL/GPT-J, 5 for GPT-3.5) and scoring every candidate prompt adds inference cost.
• Advantage over baselines is not consistent; on Amazon with GPT-2-XL APR is not clearly better and on GPT-3.5 some style scores are lower than P&R.

Experimental setup

Datasets: Yelp Amazon GYAFC Shakespeare

Benchmarks: Yelp Amazon GYAFC Shakespeare

Metrics: style transfer strength content preservation (BERTScore) fluency joint score J geometric mean

Baselines: DiRR STRAP RLPrompt Prompt-and-Rerank Low Perplexity

Builds on: Prompt-and-Rerank RLPrompt Low Perplexity prompt selection (Gonen et al. 2022) BERTScore

Closest papers by meaning

• Beyond Prompt Engineering: A Reinforced Token-Level Input Refinement for Large Language Models (AAAI 2025 · similarity 0.87)
• ADePT: Adaptive Decomposed Prompt Tuning for Parameter-Efficient Fine-tuning (ICLR 2025 · similarity 0.87)
• All You Need is One: Capsule Prompt Tuning with a Single Vector (NeurIPS 2025 · similarity 0.87)
• Zero-Shot Continuous Prompt Transfer: Generalizing Task Semantics Across Language Models (ICLR 2024 · similarity 0.86)
• Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting (ICLR 2024 · similarity 0.86)
• EPT: Efficient Prompt Tuning by Multi-Space Projection and Prompt Fusion (AAAI 2025 · similarity 0.86)