Research map / Large language models / Alignment and preferences
Generative Value Conflicts Reveal LLM Priorities
Andy Liu, Kshitish Ghate, Mona T. Diab, Daniel Fried, Atoosa Kasirzadeh, Max Kleiman-Weiner
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 |
|---|---|---|---|
| filter scenarios with LLM judge (ablation) | ConflictScope · HHH scenarios | observed agreement (lower is better): 0.824 -> 0.786 | Improved |
| filter scenarios with LLM judge (ablation) | ConflictScope · HHH scenarios | Likert difference rate: 0.818 -> 0.801 | No clear effect |
| use two-stage generation instead of single-stage (ablation) | ConflictScope · HHH scenarios | observed agreement (lower is better): 0.898 -> 0.824 (unfiltered) | Improved |
| use values and prompt templates instead of direct generation (ablation) | ConflictScope · HHH scenarios | observed agreement (lower is better): 0.852 -> 0.824 (unfiltered) | Improved |
| two-stage generation vs single-stage (ablation) | ConflictScope · HHH scenarios | Likert difference rate: 0.854 (single-stage) -> 0.818 (unfiltered two-stage) | Hurt |
| switch from multiple-choice to open-ended evaluation | 14 target LLMs · Personal-Protective | average rank of protective values: 1.7 -> 4.5 | Hurt |
| add system prompt describing target value ranking | 14 target LLMs · 3 value sets, open-ended | normalized alignment effect: average effect 0.145 | Improved |
| add steering system prompt | Claude 3.5 Haiku · 3 value sets | normalized alignment effect: 0.01 | No clear effect |
| steering prompt on ModelSpec vs HHH and Personal-Protective | 14 target LLMs · ModelSpec | normalized alignment effect | Mixed |
Closest papers by meaning
- Aligning to Thousands of Preferences via System Message Generalization (NeurIPS 2024 · similarity 0.90)
- On Evaluating LLM Alignment by Evaluating LLMs as Judges (NeurIPS 2025 · similarity 0.90)
- Do LLMs have Consistent Values? (ICLR 2025 · similarity 0.90)
- DENEVIL: TOWARDS DECIPHERING AND NAVIGATING THE ETHICAL VALUES OF LARGE LANGUAGE MODELS VIA INSTRUCTION LEARNING (ICLR 2024 · similarity 0.89)
- ALI-Agent: Assessing LLMs' Alignment with Human Values via Agent-based Evaluation (NeurIPS 2024 · similarity 0.89)
- CLAVE: An Adaptive Framework for Evaluating Values of LLM Generated Responses (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.
