Research map / Large language models / Language, safety and interpretability
Causal-Steer: Disentangled Continuous Style Control without Parallel Corpora
Qingsong Wang, Chang Yao, Jingyuan Chen
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 contrastive loss to LoRA training (ablation) | Causal-Steer · ELI5 conceptual complexity (steer to Complex) | Knowledge Difficulty (Complex): 8.24 -> 8.42 | Improved |
| add contrastive loss to LoRA training (ablation) | Causal-Steer · ELI5 conceptual complexity (steer to Simple) | Knowledge Difficulty (Simple, lower better): 3.01 -> 2.77 | Improved |
| add differential activation (with vs without LoRA on same input) (ablation) | Causal-Steer · ELI5 conceptual complexity (steer to Complex) | Knowledge Difficulty (Complex): 7.90 -> 8.42 | Improved |
| average activations over generated tokens instead of last token (ablation) | Causal-Steer · ELI5 conceptual complexity (steer to Complex) | Knowledge Difficulty (Complex): 4.26 -> 8.42 | Improved |
| average activations over generated tokens instead of last token (ablation) | Causal-Steer · ELI5 conceptual complexity (steer to Simple) | Knowledge Difficulty (Simple, lower better): 4.73 -> 2.77 | Improved |
| add PCA denoising before aggregation (ablation) | Causal-Steer · ELI5 conceptual complexity (steer to Complex) | Knowledge Difficulty (Complex): 8.18 -> 8.42 | Improved |
| add PCA denoising before aggregation (ablation) | Causal-Steer · ELI5 conceptual complexity (steer to Simple) | Knowledge Difficulty (Simple, lower better): 2.81 -> 2.77 | Improved |
| replace arithmetic mean with geometric median (ablation) | Causal-Steer · ELI5 conceptual complexity (steer to Complex) | Knowledge Difficulty (Complex): 8.24 -> 8.42 | Improved |
| replace arithmetic mean with geometric median (ablation) | Causal-Steer · ELI5 conceptual complexity (steer to Simple) | Knowledge Difficulty (Simple, lower better): 2.82 -> 2.77 | No clear effect |
| train on single-sided Simple dataset only (CSsingle) instead of both styles | Causal-Steer · ELI5 conceptual complexity (steer to Simple) | Knowledge Difficulty (Simple, lower better): 2.80 vs 2.77 | No clear effect |
| Causal-Steer vs CLMI weight interpolation on formality | Causal-Steer · ELI5 formality (steer to Formal) | Formality score (Formal): 7.58 -> 8.61 | Improved |
| Causal-Steer vs GCAV-Output on detoxification | Causal-Steer · RealToxicityPrompts (toxic prompts) | toxicity (toxic prompts): 0.0879 -> 0.0609 | Improved |
Closest papers by meaning
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.87)
- Kernelized Activation Steering (NeurIPS 2026 · similarity 0.87)
- Controlling Language and Diffusion Models by Transporting Activations (ICLR 2025 · similarity 0.86)
- Steer2Edit: From Activation Steering to Component-Level Editing (NeurIPS 2026 · similarity 0.86)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.86)
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.86)
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.
