Research map / Large language models / Language, safety and interpretability
Dual Mechanisms of Value Expression: Intrinsic vs. Prompted Values in Large Language Models
Jongwook Han, Jongwon Lim, Injin Kong, Yohan Jo
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 |
|---|---|---|---|
| steer along intrinsic or prompted value vectors | Qwen2.5-7B-Instruct · PVQ questionnaire (avg over 5 languages) | mean score delta: +1.74 (intrinsic), +2.21 (prompted) | Improved |
| remove the prompted component from the intrinsic vector (orthogonalize) (ablation) | intrinsic value vector · PVQ questionnaire (avg) | mean score delta: +1.74 -> +0.47 | Hurt |
| remove the intrinsic component from the prompted vector (orthogonalize) (ablation) | prompted value vector · PVQ questionnaire (avg) | mean score delta: +2.21 -> +1.62 | Mixed |
| orthogonalize intrinsic vector, free-form format (ablation) | intrinsic value vector · PVQ free-form (avg) | mean score delta: +0.98 -> +0.48 | Hurt |
| apply English-extracted vectors to other languages | value vectors · multilingual PVQ and dilemmas | steering effect | No clear effect |
| scale activations of shared neurons versus unique neurons | value neurons · value expression | value expression increase | Improved |
| use intrinsic steering instead of prompted steering | value vectors · situational dilemmas diversity (Achievement example) | Distinct-2: 0.342 -> 0.362 | Improved |
| orthogonalize prompted vector against intrinsic (ablation) | prompted value vector · situational dilemmas diversity (Achievement example) | Distinct-2: 0.342 -> 0.203 | Hurt |
| orthogonalize intrinsic vector against prompted (ablation) | intrinsic value vector · situational dilemmas diversity (Achievement example) | Distinct-2: 0.362 -> 0.402 | Improved |
| steer along the mean delta (prompted minus intrinsic) direction | Llama-3.1-8B-Instruct with compliance system prompt · AdvBench | ASR@9: 13.3% -> 97.2% | Improved |
| steer along the mean delta direction | Qwen2.5-7B-Instruct with compliance system prompt · HarmBench | ASR@9: 52.4% -> 83.0% | Improved |
| steer along compliance direction on non-value instruction tasks | Qwen2.5-7B-Instruct · gender-specific translation vs strict JSON formatting | compliance | Mixed |
Closest papers by meaning
- DENEVIL: TOWARDS DECIPHERING AND NAVIGATING THE ETHICAL VALUES OF LARGE LANGUAGE MODELS VIA INSTRUCTION LEARNING (ICLR 2024 · similarity 0.86)
- Value Entanglement: Conflation Between Different Kinds of Good In (Some) Large Language Models (NeurIPS 2026 · similarity 0.86)
- Do LLMs have Consistent Values? (ICLR 2025 · similarity 0.86)
- Internal Causal Mechanisms Robustly Predict Language Model Out-of-Distribution Behaviors (ICML 2025 · similarity 0.85)
- Interpreting Language Reward Models via Contrastive Explanations (ICLR 2025 · similarity 0.85)
- CLAVE: An Adaptive Framework for Evaluating Values of LLM Generated Responses (NeurIPS 2024 · similarity 0.84)
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.
