Research map / Large language models / Instruction and fine-tuning
Safe Responses Matter: Output-Aware Safety Guardrail Mitigate Over-Refusal in MLLMs
Jiayi Li, Kun Zhan
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 the gating module to the MICL classifier (ablation) | OutGuard · Test set (GQA+AdvBench+SafeBench), LLaVA | URUP+ARSP: aggregated metric set to 2, the worst value without it | Improved |
| add the contrastive learning term L_con (ablation) | OutGuard · Test set (GQA+AdvBench+SafeBench), LLaVA | URUP+ARSP: aggregated metric set to 2, the worst value without it | Improved |
| add the sparsity (entropy) regularization L_ent (ablation) | OutGuard · Test set (GQA+AdvBench+SafeBench), LLaVA | URUP+ARSP: aggregated metric set to 2, the worst value without it | Improved |
| use top-k 0.3 instead of 0.1 (ablation) | OutGuard · Test set (GQA+AdvBench+SafeBench), LLaVA | URUP+ARSP: 0.1326 -> 0.1299 | No clear effect |
| use lambda_con 0.1 instead of 0.05 (ablation) | OutGuard · Test set (GQA+AdvBench+SafeBench), LLaVA | URUP+ARSP: 0.1441 -> 0.1299 | Improved |
| use temperature tau 0.5 instead of 0.1 (ablation) | OutGuard · Test set (GQA+AdvBench+SafeBench), LLaVA | URUP+ARSP: 0.1816 -> 0.1299 | Improved |
| use attention dimension 256 instead of 512 (ablation) | OutGuard · Test set (GQA+AdvBench+SafeBench), LLaVA | URUP+ARSP: 0.1498 -> 0.1299 | Improved |
| use attention dimension 256 instead of 128 (ablation) | OutGuard · Test set (GQA+AdvBench+SafeBench), LLaVA | URUP+ARSP: 0.1325 -> 0.1299 | No clear effect |
| use hidden dimension 256 instead of 768 (ablation) | OutGuard · Test set (GQA+AdvBench+SafeBench), LLaVA | URUP+ARSP: 0.1585 -> 0.1299 | Improved |
| replace LoD-PP linear per-layer classifier with MICL classifier (ablation) | OutGuard · Test set, LLaVA | F1-Score: 0.8943 -> 0.9335 | Improved |
| replace input-side guardrails with output-aware OutGuard | OutGuard · XSTest, LLaVA | over-refusal rate: 0.7080 -> 0.5840 | Improved |
| replace input-side guardrails with output-aware OutGuard | OutGuard · XSTest, Qwen | over-refusal rate: 0.7760 -> 0.6320 | Improved |
Closest papers by meaning
- On Prompt-Driven Safeguarding for Large Language Models (ICML 2024 · similarity 0.91)
- UpSafe℃: Upcycling for Controllable Safety in Large Language Models (NeurIPS 2026 · similarity 0.90)
- Safe RLHF-V: Safe Reinforcement Learning from Multi-modal Human Feedback (NeurIPS 2025 · similarity 0.90)
- RSafe: Incentivizing proactive reasoning to build robust and adaptive LLM safeguards (NeurIPS 2025 · similarity 0.89)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.89)
- When Safe Unimodal Inputs Collide: Optimizing Reasoning Chains for Cross-Modal Safety in Multimodal Large Language Models (AAAI 2026 · 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.
