Research map / Large language models / Alignment and preferences
Discern Truth from Falsehood: Reducing Over-Refusal via Contrastive Refinement
Yuxiao Lu, Lin Xu, Yang Sun, Wenjun Li, Jie Shi
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 a contrastive refinement stage before safety SFT (DCR) | STL on Qwen2.5-1.5B · XSTest | compliance rate: 0.73 -> 0.98 | Improved |
| add a contrastive refinement stage before safety SFT (DCR) | STL on Qwen2.5-7B · OR-Bench | compliance rate: 0.34 -> 0.71 | Improved |
| add a contrastive refinement stage before safety SFT (DCR) | STL on LLaMA-3-8B · XSTest | compliance rate: 0.79 -> 0.93 | Improved |
| add a contrastive refinement stage before safety SFT (DCR) | STL on Qwen2.5-1.5B · safety (defense success rate) | defense success rate: 0.72 -> 0.81 | Improved |
| add a contrastive refinement stage before safety SFT (DCR) | STL on LLaMA-3-8B · safety (defense success rate) | defense success rate: 0.93 -> 0.91 | Hurt |
| add a contrastive refinement stage before safety SFT (DCR) | STL on LLaMA-3-8B · general QA (ARC_c, OpenBookQA) | accuracy: ARC_c 0.56 -> 0.51; OpQA 0.45 -> 0.39 | Hurt |
| add a contrastive refinement stage before safety SFT (DCR) | STL on Qwen2.5-7B · ARC_e / ARC_c | accuracy: ARC_e 0.77 -> 0.83; ARC_c 0.51 -> 0.59; OpQA 0.47 -> 0.44 | Mixed |
| use contrastive learning on seemingly toxic prompts instead of adding them to SFT data | STL-aug · XSTest | compliance rate: Qwen2.5-1.5B: 0.75 -> 0.98 | Improved |
| compare DCR with SCANS on response quality (Qwen2.5-7B) | SCANS · AlpacaEval win rate | win rate (%): 45.5 -> 45.8 | Improved |
| contrastive learning applied before alignment | LLM representations · kernel similarity between seemingly toxic and toxic prompts | normalized K Frobenius norm | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SCANS — SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steering (AAAI 2025) · better on 4 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Just Enough Shifts: Mitigating Over-Refusal in Aligned Language Models with Targeted Representation Fine-Tuning (ICML 2025 · similarity 0.92)
- OR-Bench: An Over-Refusal Benchmark for Large Language Models (ICML 2025 · similarity 0.90)
- Residual Stream Analysis of Overfitting And Structural Disruptions (NeurIPS 2025 · similarity 0.90)
- Fail-Closed Alignment for Large Language Models (NeurIPS 2026 · similarity 0.89)
- Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning Attack (ICML 2025 · similarity 0.89)
- On Prompt-Driven Safeguarding for Large Language Models (ICML 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.
