Add new paper:

[wyzh0912]

Title

DRESSING UP LLM: EFFICIENT STYLIZED QUESTIONANSWERING VIA STYLE SUBSPACE EDITING

Published Date

2025-01-23

Source

ICLR

Head Name

Style Head

Summary

• Innovation: The paper introduces DRESS, a novel train-free framework for stylized question-answering (QA) in large language models (LLMs) by editing style-relevant subspaces within the model's representation space, allowing for adaptive and controllable stylization while preserving semantic integrity.

• Tasks: The study involves constructing stylized QA benchmarks in Shakespeare-style English and Dream of the Red Chamber-style Chinese, and evaluates the performance of DRESS against baseline methods like prompting and fine-tuning, focusing on style intensity, semantic preservation, and fluency of respon

• Significant Result: DRESS significantly outperforms existing methods, including prompting, supervised fine-tuning, and conventional representation editing techniques, in terms of style intensity, semantic preservation, and fluency, demonstrating its effectiveness in enhancing LLMs with flexible style control for conversational agents.