Research map / Large language models / Alignment and preferences
Disentangling Consensus and Value-Specific Representations for Controllable Pluralistic Value Alignment of LLMs
JianKui Zhou, Jing Yao, Xiaoyuan Yi, Peng Zhang, Ning Gu, Zhan Hu, Xing Xie, Tun Lu
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 consensus anchor and subspace to the value subspace (ablation) | DisAlign (Llama-3.2-3B) · MIC MMLU | accuracy (%): 46.5 -> 62.1 | Improved |
| add consensus anchor and subspace to the value subspace (ablation) | DisAlign (Llama-3.2-3B) · MIC ARC-C | accuracy (%): 50.2 -> 77.2 | Improved |
| add value-specific subspace to the consensus component (ablation) | DisAlign (Llama-3.2-3B) · MIC special parts D∆ | accuracy (%): 30.5 -> 53.5 | Improved |
| add value-specific subspace to the consensus component (ablation) | DisAlign (Llama-3.2-3B) · MIC MMLU | accuracy (%): 62.8 -> 62.1 | Hurt |
| add consensus removal to diagonal Fisher decomposition (ablation) | diag Fisher merging · MIC, B. Random D. | accuracy (%): 45.8 -> 58.3 | Improved |
| add consensus removal to Euclidean decomposition (ablation) | Euclid-I merging · MIC, B. Random D. | accuracy (%): 40.8 -> 46.9 | Improved |
| use Fisher subspace decomposition instead of diagonal Fisher (ablation) | DisAlign · MIC, B. Random D. | accuracy (%): 58.3 -> 59.6 | Improved |
| use Fisher subspace decomposition instead of Euclidean (ablation) | DisAlign · MIC, B. Random D. | accuracy (%): 46.9 -> 59.6 | Improved |
| use DisAlign instead of the strongest prompting baseline | Llama-3.2-3B · MIC B. Random D. | accuracy (%): 54.3 -> 59.6 | Improved |
| use DisAlign instead of Prompt-Align on full-dimensional ValuePrism | Llama-3.2-3B · ValuePrism C. Full D. | accuracy (%): 55.8 -> 56.2 | No clear effect |
| use DisAlign instead of FWA/DARE | Llama-3.2-3B · MIC open-ended generation | GPT-4o value alignment, random (1-5): 2.4 -> 4.2 | Improved |
| use DisAlign instead of FWA/DARE | Llama-3.2-3B · MIC open-ended generation | response quality (1-5): 3.4 -> 3.7 | No clear effect |
Closest papers by meaning
- Counterfactual Reasoning for Steerable Pluralistic Value Alignment of Large Language Models (NeurIPS 2025 · similarity 0.88)
- Position: A Roadmap to Pluralistic Alignment (ICML 2024 · similarity 0.87)
- Model Spec Midtraining: Improving How Alignment Training Generalizes (NeurIPS 2026 · similarity 0.87)
- Deep Value Benchmark: Measuring Whether Models Generalize Deep values or Shallow Preferences (NeurIPS 2025 · similarity 0.87)
- SEAL: Systematic Error Analysis for Value ALignment (AAAI 2025 · similarity 0.86)
- Pairwise Calibrated Rewards for Pluralistic Alignment (NeurIPS 2025 · 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.
