Get Started
Home
Topics
Search
Library
Research questionHow can safety-tuned language models distinguish harmful requests from benign ones with risky wording?Safety tuning can cause models to rely on risky-looking words rather than the request’s actual intent. As a result, harmless users may be refused even when the model should respond helpfully.
AI
Alignment & Safety
Evaluation & Benchmarks
LLM Pretraining & Post-training
Machine Learning
Natural Language Processing
Latest papersRecent research connected to this question, newest first.Refuse without Refusal: A Structural Analysis of Safety-Tuning Responses for Reducing False Refusals in Language ModelsThe source examines how refusal statements and refusal rationales in safety-tuning data affect this distinction, covering supervised training, in-context learning, and evaluated inference-time mitigation settings. Its evidence is limited to the tested models, data configurations, and deployment conditions.research paper · Sep 4, 2026Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety RefusalThe evidence concerns language-model safety alignment using offline self-generated refusal data and harmful-benign boundary pairs, with political persuasion as the primary example. Reported experiments use Qwen3-8B and assess political persuasion, broader harmfulness benchmarks, XSTest, and held-out boundary pairs; the results do not establish broader generalization beyond these settings.research paper · Sep 3, 2026Representational alignment yields generalizable safety in language modelsThe evidence concerns 23 large language models and uses human moral judgements to examine category distinctions and fine-grained typicality. It compares response-level alignment with alignment of latent moral representations across model scales and adversarial evaluations, using 251,334 moral annotations.research paper · Sep 3, 2026
Related questions
How should LLM safety be evaluated when harmful prompts vary in implicitness and sophistication?How can language models answer sensitive prompts helpfully without compromising safety?How can multimodal models detect harmful intent when benign text is paired with risky visual content?How can language models distinguish trusted instructions from untrusted text to resist prompt injection?