Balancing Stylization and Truth via Disentangled Representation Steering
Not on the map. Why?
Abstract (arXiv)
Generating stylized large language model (LLM) responses via representation editing is a promising way for fine-grained output control. However, there exists an inherent trade-off: imposing a distinctive style often degrades truthfulness. Existing representation editing methods, by naively injecting style signals, overlook this collateral impact and frequently contaminate the model's core truthfulness representations, resulting in reduced answer correctness. We term this phenomenon stylization-induced truthfulness collapse. We attribute this issue to latent coupling between style and truth directions in certain key attention heads, and propose StyliTruth, a mechanism that preserves stylization while keeping truthfulness intact. StyliTruth separates the style-relevant and truth-relevant subspaces in the model's representation space via an orthogonal deflation process. This decomposition enables independent control of style and truth in their own subspaces, minimizing interference. By designing adaptive, token-level steering vectors within each subspace, we dynamically and precisely control the generation process to maintain both stylistic fidelity and truthfulness. We validate our method on multiple styles and languages. Extensive experiments and analyses show that StyliTruth significantly reduces stylization-induced truthfulness collapse and outperforms existing inference-time intervention methods in balancing style adherence with truthfulness.
Key facts. Cited by 2 papers in the map (data through 7 Oct 2026). Builds on 23 of its 27 references that are in the map. First posted on arXiv 6 Aug 2025.
Builds on
23 of 27 references.
References · 23
- 2017Attention Is All You Needcited by 59,932
- 2021TruthfulQA: Measuring How Models Mimic Human Falsehoodscited by 2,919
- 2016Understanding intermediate layers using linear classifier probescited by 1,279
- 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,259
- 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,239
- 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,009
- 2019What Does BERT Look At? An Analysis of BERT's Attentioncited by 994
- 2022Toy Models of Superpositioncited by 928
- 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 831
- 2022Discovering Latent Knowledge in Language Models Without Supervisioncited by 733
- 2023Model Tells You What to Discard: Adaptive KV Cache Compression for LLMscited by 466
- 2023Character-LLM: A Trainable Agent for Role-Playingcited by 335
11 more
- 2025Layer by Layer: Uncovering Hidden Representations in Language Modelscited by 293
- 2020Deep Learning for Text Style Transfer: A Surveycited by 161
- 2023Inspecting and Editing Knowledge Representations in Language Modelscited by 140
- 2023Word Embeddings Are Steers for Language Modelscited by 70
- 2023Improving Activation Steering in Language Models with Mean-Centringcited by 64
- 2024Aligning Large Language Models with Representation Editing: A Control Perspectivecited by 62
- 2023Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuningcited by 45
- 2025Multi-Attribute Steering of Language Models via Targeted Interventioncited by 44
- 2025DRESSing Up LLM: Efficient Stylized Question-Answering via Style Subspace Editingcited by 10
- 2022Overparameterized random feature regression with nearly orthogonal datacited by 6
- 2025Disentangling Knowledge Representations for Large Language Model Editingcited by 5
Built on by
Citing papers per year: 2025: 1; 2026: 1.
2 papers.