Research map / Large language models / Language, safety and interpretability
DSO: Direct Steering Optimization for Bias Mitigation
Lucas Monteiro Paes, Nivedha Sivakumar, Yinong Oliver Wang, Masha Fedzechkina, Barry-John Theobald, Luca Zappella, Nicholas Apostoloff
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 DSO steering (lambda 0.2) | Qwen2.5-7B VL · SocialCounterfactuals occupation identification | Per-occupation bias: 25.8% -> 15.4% | Improved |
| raise DSO strength from 0.2 to 1.0 | Qwen2.5-7B VL · SocialCounterfactuals occupation identification | Unambiguous accuracy: 95.7% -> 80.0% | Hurt |
| raise DSO strength from 0.2 to 1.0 | Qwen2.5-7B VL · SocialCounterfactuals occupation identification | Per-occupation bias: 15.4% -> 8.7% | Improved |
| add prompt-based debiasing | Llama-3.2-11B VL · SocialCounterfactuals occupation identification | Per-occupation bias: 30.2% -> 39.2% | Hurt |
| add CAA steering | Qwen2.5-7B VL · SocialCounterfactuals occupation identification | Per-occupation bias: 25.8% -> 28.5% | Hurt |
| add ITI steering | Qwen2.5-3B VL · SocialCounterfactuals occupation identification | MMMU accuracy: 41.3% -> 30.5% | Hurt |
| add DSO steering (lambda 1.0) | Gemma-3-12B · SocialCounterfactuals occupation identification | Per-occupation bias: 26.5% -> 15.0% | Improved |
| add DSO steering | Qwen2.5-3B · SynthBias coreference resolution | Per-occupation bias: 60.4% -> 5.9% | Improved |
| add DSO steering | Qwen2.5-3B · SynthBias coreference resolution | Don't-know rate: 13.6% -> 45.6% | Hurt |
| add DSO steering (lambda 1) | Qwen2.5-3B · SynthBias coreference resolution | Unambiguous accuracy: 88.5% -> 99.7% | Improved |
| restrict intervention to 60% vs 100% of LayerNorm neurons (ablation) | DSO · SocialCounterfactuals occupation identification | Per-occupation bias | No clear effect |
| restrict intervention to 40% of LayerNorm neurons (ablation) | DSO · SocialCounterfactuals occupation identification | Per-occupation bias: increases bias slightly (5%) | Hurt |
Closest papers by meaning
- Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization (NeurIPS 2024 · similarity 0.88)
- On Optimal Steering to Achieve Exact Fairness (NeurIPS 2025 · similarity 0.87)
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.86)
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.86)
- The Sharp Directions Are Against You: Curvature Analysis of Activation Steering (NeurIPS 2026 · 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.
