Research map / Large language models / Alignment and preferences
Safety Game: Inference-Time Alignment of Black-Box LLMs via Constrained Optimization
Tuan Nguyen, Long Tran-Thanh
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 best per-candidate threshold filter with aggregate-risk LP selector on identical Llama-Guard scores (ablation) | Safety Game · SORRY-Bench (Llama-Guard-3-8B scorer) | prompts with safe response selected: 338/440 -> 419/440 | Improved |
| use Safety Game selector instead of Consensus-Game-style ranking | Llama-2/3 models · SafetyBench | accuracy: LLaMA-2-7B 45.71 (G) -> 54.79 | Improved |
| use Safety Game selector with a weak self-probe scorer | Llama-3.2-1B · SafetyBench | accuracy: 44.50 (G) -> 39.00 | Hurt |
| use Safety Game selector with a weak self-probe scorer | Llama-3.2-1B · SORRY-Bench | expected human fulfillment rate: 0.220 (ER-G) -> 0.415 | Hurt |
| use Safety Game selector instead of best baseline | LLaMA-2-7B · SORRY-Bench | expected human fulfillment rate: 0.272 (SC) -> 0.051 | Improved |
| use Safety Game selector instead of G/MI | GPT-OSS-20B · SORRY-Bench | expected human fulfillment rate: 0.009 (G) -> 0.223 | Hurt |
| use Safety Game selector instead of D | LLaMA-2-13B · HHH | accuracy: 75.1 (D) -> 69.2 | Hurt |
| use Safety Game selector instead of best baseline | Llama-3.1-8B · HHH | accuracy: 65.6 (MI) -> 67.9 | Improved |
| use Safety Game selector instead of best baseline | GPT-OSS-20B · TruthfulQA | BLEU-Acc: 52.1 (ER-D) -> 54.7 | Improved |
| use Safety Game selector instead of SC | LLaMA-2-13B · TruthfulQA | BLEU-Acc: 51.1 (SC) -> 50.9 | No clear effect |
| aggregate-risk LP versus filter-then-rank baseline with same candidates and scores (ablation) | Safety Game · SORRY-Bench | HFR, safe-count, refusal quality | Improved |
Closest papers by meaning
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.90)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.90)
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.90)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.90)
- The Alignment Waltz: Jointly Training Agents to Collaborate for Safety (ICLR 2026 · similarity 0.90)
- BSO: Safety Alignment Is Density Ratio Matching (NeurIPS 2026 · similarity 0.90)
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.
