Research map / Large language models / Language, safety and interpretability
StyliTruth : Unlocking Stylized yet Truthful LLM Generation via Disentangled Steering
Chenglei Shen, Zhongxiang Sun, Teng Shi, Xiao Zhang, Jun Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add subspace disentanglement (ablation) | StyliTruth · DRC style, TruthfulQA(ZH), Qwen1.5-14B-Chat | S-TI: 0.1632 -> 0.2366 | Improved |
| add subspace disentanglement (ablation) | StyliTruth · DRC style, TruthfulQA(ZH), Qwen1.5-14B-Chat | TI: 0.3194 -> 0.5000 | Improved |
| add adaptive token-level editing (ablation) | StyliTruth · DRC style, TruthfulQA(ZH), Qwen1.5-14B-Chat | S-TI: 0.1575 -> 0.2366 | Improved |
| add adaptive token-level editing (ablation) | StyliTruth · DRC style, TruthfulQA(ZH), Qwen1.5-14B-Chat | TI: 0.2017 -> 0.5000 | Improved |
| replace sequential style-then-truth steering with parallel disentangled steering (ablation) | StyliTruth · DRC style, TruthfulQA(ZH), Qwen1.5-14B-Chat | S-TI: 0.0301 -> 0.2366 | Improved |
| raise subspace dimension from 16 to 64 (ablation) | StyliTruth · DRC style, TruthfulQA(ZH), Qwen1.5-14B-Chat | S-TI: 0.2249 -> 0.2366 | Improved |
| raise subspace dimension from 64 to 128 (ablation) | StyliTruth · DRC style, TruthfulQA(ZH), Qwen1.5-14B-Chat | S-TI: 0.2366 -> 0.2002 | Hurt |
| use no style/truth subspace editing (K=0) (ablation) | StyliTruth · DRC style, TruthfulQA(ZH), Qwen1.5-14B-Chat | S-TI: 0.2366 -> 0.0084 | Hurt |
| select more attention heads (ablation) | StyliTruth · DRC style, TruthfulQA(ZH), Qwen1.5-14B-Chat | OA, TI, S-TI | Hurt |
| StyliTruth versus strongest baseline (DRESS) on Shakespeare | StyliTruth · Shakespeare style, TruthfulQA, Qwen1.5-14B-Chat | S-TI: 0.2529 -> 0.2803 | Improved |
| StyliTruth versus strongest baseline (Vector prompt) on DRC | StyliTruth · DRC style, TruthfulQA(ZH), Qwen1.5-14B-Chat | S-TI: 0.1811 -> 0.2366 | Improved |
| StyliTruth versus LM Steer on truthfulness | StyliTruth · DRC style, TruthfulQA(ZH), Qwen1.5-14B-Chat | TI: 0.5972 -> 0.5000 | Hurt |
Closest papers by meaning
- DRESSing Up LLM: Efficient Stylized Question-Answering via Style Subspace Editing (ICLR 2025 · similarity 0.88)
- TruthFlow: Truthful LLM Generation via Representation Flow Correction (ICML 2025 · similarity 0.86)
- Characterizing Truthfulness in Large Language Model Generations with Local Intrinsic Dimension (ICML 2024 · similarity 0.85)
- Search for Truth from Reasoning: A Dynamic Representation Editing Framework for Steering LLM Trajectories (ICML 2026 · similarity 0.84)
- LoFiT: Localized Fine-tuning on LLM Representations (NeurIPS 2024 · similarity 0.84)
- GRATH: Gradual Self-Truthifying for Large Language Models (ICML 2024 · 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.
