Research map / Large language models / Alignment and preferences
VALUEFLOW: Toward Pluralistic and Steerable Value-based Alignment in Large Language Models
Woojin Kim, Sieun Hyeon, Jusang Oh, Jaeyoung Do
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 |
|---|---|---|---|
| replace direct LLM rating with ranking-based intensity evaluation (ablation) | value intensity evaluator · SVT, 10K texts per value | mean variance: 12.6 -> 2.1 | Improved |
| replace direct LLM rating with ranking-based intensity evaluation (ablation) | value intensity evaluator · SVT, 10K texts per value | sign-flip rate (%): 48 -> 29 | Improved |
| replace direct LLM rating with ranking-based intensity evaluation (ablation) | value intensity evaluator · ValueNet agreement (SVT) | pairwise ranking accuracy (%): 77.4 -> 84.2 | Improved |
| add value-profile steering using HiVES-derived profiles | Phi-4 (14B) · OpinionQA, Religion group | accuracy (%): 44.5 -> 57.4 | Improved |
| use value-profile steering instead of default group-attribute prompt | Qwen3-32B, Phi-4, GLM-4-32B · OpinionQA micro-average | accuracy (%): 56.6 -> 59.1 (Qwen3); 51.2 -> 55.7 (Phi-4); 56.8 -> 59.1 (GLM-4) | Improved |
| use value-profile steering instead of Modular Pluralism | Qwen3-32B · OpinionQA micro-average | accuracy (%): 39.3 -> 59.1 | Improved |
| profile steering on Phi-4 for Region, Education, Income groups | Phi-4 (14B) · OpinionQA by group | accuracy (%): Region 60.3 -> 57.2; Education 57.5 -> 56.2; Income 55.3 -> 52.5 | Mixed |
| add HiVES two-stage hierarchical embedding training on Qwen3-embedding-0.6B | Qwen3-embedding-0.6B · SVT and MFT hierarchy metrics | ranking consistency and similarity correlation: over 20% ranking consistency, over 50% similarity correlation | Improved |
| user-text prompts instead of intensity-anchor prompts for steering (ablation) | steered LLMs · 500-prompt steerability benchmark | steering gain | Mixed |
| steer negative targets on prosocial values (Benevolence, Universalism) | Phi-4 and Claude-4 · 500-prompt steerability benchmark | steering gain | No clear effect |
| steer positive vs negative intensity targets | Hedonism and most rights values · steerability by value | steering gain | Mixed |
| steer Conformity value in either direction | prompt-based steering · steerability by value | steering gain: |delta| ~ 0 | No clear effect |
Closest papers by meaning
- CLAVE: An Adaptive Framework for Evaluating Values of LLM Generated Responses (NeurIPS 2024 · similarity 0.90)
- Exploring Intrinsic Alignments Within Text Corpus (AAAI 2025 · similarity 0.90)
- LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations (NeurIPS 2026 · similarity 0.90)
- DENEVIL: TOWARDS DECIPHERING AND NAVIGATING THE ETHICAL VALUES OF LARGE LANGUAGE MODELS VIA INSTRUCTION LEARNING (ICLR 2024 · similarity 0.89)
- RMB: Comprehensively benchmarking reward models in LLM alignment (ICLR 2025 · similarity 0.89)
- Aligning to Thousands of Preferences via System Message Generalization (NeurIPS 2024 · similarity 0.89)
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.
