Research map / Image generation / Autoregressive and latent generation
MidSteer: Optimal Affine Framework for Steering Generative Models
Tatiana Gaintseva, Andrew Stepanov, Ziquan Liu, Martin Benning, Gregory Slabaugh, Jiankang Deng, Ismail Elezi
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 |
|---|---|---|---|
| use MidSteer directed covariance matching instead of vanilla reflection steering | SDXL horse to motorcycle switching · motorcycle prompts | target CLIP score: 52.9 -> 70.7 | Improved |
| use MidSteer instead of vanilla steering | SDXL unrelated concept cow · horse to motorcycle | FID vs vanilla model: 212.4 -> 12.7 | Improved |
| use LEACE-Switch instead of vanilla steering | SDXL horse to motorcycle · motorcycle prompts | target CLIP score: 52.9 -> 53.3 | No clear effect |
| use MidSteer instead of LEACE-Switch (strength 5) | Llama-2-7B-chat toxicity to helpfulness · RealToxicityPrompts | toxicity: .346 -> .281 | Improved |
| increase vanilla steering strength from 3 to 5 (ablation) | Llama-2-7B-chat toxicity to helpfulness · unrelated concept preservation | unrelated concept score: 8.45 -> 3.11 | Hurt |
| increase MidSteer strength from 3 to 5 (ablation) | Llama-2-7B-chat toxicity to helpfulness · RealToxicityPrompts | toxicity: .309 -> .281 | Improved |
| increase MidSteer strength from 3 to 5 (ablation) | SDXL violence to peace · violence CLIP score and FID | violence score: violence 12.5 -> 6.0; FID 101.5 -> 134.6 | Mixed |
| use LEACE-Switch instead of vanilla steering | SDXL violence to peace · strength 3 | peace CLIP score: 74.5 -> 33.2 | Hurt |
| use MidSteer instead of vanilla steering (strength 5) | SDXL violence to peace · image distortion | FID vs base: 122.2 -> 134.6 | Hurt |
| use MidSteer instead of LEACE-Switch (strength 3) | Llama-2-7B-chat · MMLU BERT-F1 | BERT-F1: .97 -> .95 | Hurt |
Closest papers by meaning
- Minimalist Concept Erasure in Generative Models (ICML 2025 · similarity 0.88)
- Representation Surgery: Theory and Practice of Affine Steering (ICML 2024 · similarity 0.88)
- What's Producible May Not Be Reachable: Measuring the Steerability of Generative Models (NeurIPS 2025 · similarity 0.88)
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.88)
- Controlling Language and Diffusion Models by Transporting Activations (ICLR 2025 · similarity 0.87)
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.87)
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.
