Research map / Large language models / Retrieval and knowledge
DRESSing Up LLM: Efficient Stylized Question-Answering via Style Subspace Editing
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 DRESS, a train-free representation-editing method that edits a disentangled style subspace of selected attention heads with adaptive per-token strengths to produce stylized LLM QA responses.
Problem
Making LLMs answer user questions in a specific language style (e.g., Shakespearean or Dream of the Red Chamber style) without degrading semantics, and without the cost of per-character fine-tuning or the limits of prompting.
Approach
Identify style-relevant attention heads via linear probes, then compute an orthogonal style subspace via SVD of activation differences between ordinary and stylized responses, and steer within that subspace with adaptive strengths (global projection scaled per token by a context-dependent cosine factor).
Key results
- On the Shakespeare-style benchmark DRESS reaches OA 28.9% and GPT-4 rating 8.85 vs. 26.9% and 8.73 for the best baseline (TrFr); on the Dream of the Red Chamber benchmark OA is 29.1% vs. 23.8% for the best baseline (DRESS*), a reported 23.8% relative improvement.
- With LLaMA-3-8B, DRESS achieves OA 25.1% on the Dream of the Red Chamber benchmark vs. 10.6% for prompting and 21.2% for ITI.
- DRESS with only 1% of training data (about 40 samples) still outperforms 3-shot prompting on the Dream of the Red Chamber benchmark (OA 24.7% vs 22.7%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DRESS instead of the strongest non-DRESS baseline | Stylized QA with Qwen-1.5-14B-Chat · Dream of the Red Chamber-style | OA (%): 23.5 -> 29.1 | Helped |
| Use DRESS instead of the strongest non-DRESS baseline | Stylized QA with Qwen-1.5-14B-Chat · Dream of the Red Chamber-style | GPT-4 rating: 7.64 -> 7.96 | Helped |
| Use DRESS instead of the strongest non-DRESS baseline | Stylized QA with Qwen-1.5-14B-Chat · Shakespeare-style | OA (%): 26.8 -> 28.9 | Helped |
| Use DRESS instead of the strongest non-DRESS baseline | Stylized QA with Qwen-1.5-14B-Chat · Shakespeare-style | GPT-4 rating: 8.73 -> 8.85 | Helped |
| Add adaptive scaling factor γ during inference (ablation) | DRESS · Dream of the Red Chamber-style | SI (%): 89.0 -> 97.0 | Helped |
| Add adaptive scaling factor γ during inference (ablation) | DRESS · Dream of the Red Chamber-style | SP (%): 70.9 -> 70.8 | Hurt |
| Add adaptive scaling factor γ during inference (ablation) | DRESS · Dream of the Red Chamber-style | FS (%): 37.8 -> 42.4 | Helped |
| Add adaptive scaling factor γ during inference (ablation) | DRESS · Dream of the Red Chamber-style | OA (%): 23.8 -> 29.1 | Helped |
| Add adaptive scaling factor γ during inference (ablation) | DRESS · Dream of the Red Chamber-style | GPT-4 rating: 7.71 -> 7.96 | Helped |
| Add adaptive scaling factor γ during inference (ablation) | DRESS · Shakespeare-style | SI (%): 99.0 -> 99.5 | Helped |
| Add adaptive scaling factor γ during inference (ablation) | DRESS · Shakespeare-style | SP (%): 71.2 -> 73.3 | Helped |
| Add adaptive scaling factor γ during inference (ablation) | DRESS · Shakespeare-style | FS (%): 38.2 -> 39.6 | Helped |
Surprising findings
- Most conventional representation editing baselines (Mean-Centring, RepE) fail to reach few-shot prompting performance on stylized QA.
- On LLaMA-3-8B, prompting almost fails to stylize the Dream of the Red Chamber style (SI 38.8%), while representation editing does much better.
Limitations
- Scalability to larger LLMs (100B+) not validated due to compute limits.
- Building full character agents with personality modeling and dialogue memory is left for future work.
- Hyperparameters (H, K, lambda) tuned by grid search per benchmark; evaluation uses only two style datasets and GPT-4 judging.
Experimental setup
Datasets: Shakespeare plays dialogue (target style) Dream of the Red Chamber novel and TV/film scripts (target style) MOSS general QA
Benchmarks: Shakespeare-style stylized QA benchmark Dream of the Red Chamber-style stylized QA benchmark
Metrics: style intensity semantic preservation fluency score overall assessment (SI*SP*FS) GPT-4 rating
Baselines: Few-shot Prompting SFT (LoRA) Mean-Centring RepE TrFr ITI
Builds on: Inference-Time Intervention (ITI) RepE Mean-Centring LoRA
Closest papers by meaning
- DRESS: Instructing Large Vision-Language Models to Align and Interact with Humans via Natural Language Feedback (CVPR 2024 · similarity 0.85)
- ChatGarment: Garment Estimation, Generation and Editing via Large Language Models (CVPR 2025 · similarity 0.85)
- Tuning LayerNorm in Attention: Towards Efficient Multi-Modal LLM Finetuning (ICLR 2024 · similarity 0.85)
- CharacterBench: Benchmarking Character Customization of Large Language Models (AAAI 2025 · similarity 0.84)
- Easy Regional Contrastive Learning of Expressive Fashion Representations (NeurIPS 2024 · similarity 0.84)
- MEMOIR: Lifelong Model Editing with Minimal Overwrite and Informed Retention for LLMs (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.