Research map / Large language models / Language, safety and interpretability
Toward Stable Value Alignment: Introducing Independent Modules for Consistent Value Guidance
Wenhao Chen, Sirui Sun, Shengyuan Bai, Guojie Song
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 SVGT independent value module with bridge tokens | Llama-3.2-3B-Instruct · WildGuardMix | harmful score: 29.69 -> 7.84 | Improved |
| add SVGT | Mistral-7B-Instruct-v0.2 · HarmBench | attack success rate: 77.20 -> 17.20 | Improved |
| add SVGT | GPT-2 · WildGuardMix | harmful score: 24.43 -> 2.66 | Improved |
| add SVGT | Llama-3.2-3B-Instruct · WikiText-2 | perplexity: 6.71 -> 7.34 | No clear effect |
| add ITI instead of SVGT | Llama-3.2-3B-Instruct · WikiText-2 | perplexity: 6.71 -> 11.01 | Hurt |
| add SVGT vs DPO (LoRA) on weak backbone | GPT-2 · BeaverTails | harmful score: DPO 20.64 vs SVGT 19.89 | Mixed |
| add conditional encoding to unconditional value encoder (ablation) | Value encoder, Llama-3.2-3B · WildGuardMix | Macro F1: 88.41 -> 83.95 | Hurt |
| add conditional encoding to unconditional value encoder (ablation) | Value encoder, Llama-3.2-3B · BeaverTails | AUROC: 78.45 -> 90.91 | Improved |
| add conditional encoding to unconditional value encoder (ablation) | Value encoder, Mistral-7B · WildGuardMix | AUROC: 94.06 -> 90.85 | Hurt |
| use bridge tokens (attention targets) instead of residual injection (ablation) | SVGT, Llama-3.2-3B · WildGuardMix | harmful score: 13.29 -> 7.84 | Improved |
| use bridge tokens instead of residual injection (ablation) | SVGT, Llama-3.2-3B · WikiText-2 | perplexity: 10.28 -> 7.34 | Improved |
| use EMA momentum beta in 0.6-0.8 for bridge token updates (vs 0 or 1) (ablation) | SVGT · safety score | harmful score | Improved |
Closest papers by meaning
- DENEVIL: TOWARDS DECIPHERING AND NAVIGATING THE ETHICAL VALUES OF LARGE LANGUAGE MODELS VIA INSTRUCTION LEARNING (ICLR 2024 · similarity 0.89)
- Exploring Intrinsic Alignments Within Text Corpus (AAAI 2025 · similarity 0.89)
- LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations (NeurIPS 2026 · similarity 0.88)
- SEAL: Systematic Error Analysis for Value ALignment (AAAI 2025 · similarity 0.88)
- Deep Value Benchmark: Measuring Whether Models Generalize Deep values or Shallow Preferences (NeurIPS 2025 · similarity 0.88)
- RMB: Comprehensively benchmarking reward models in LLM alignment (ICLR 2025 · similarity 0.88)
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.
