Researchers have identified two distinct axes of abstention in large language models: answer correctness and question answerability. Current models struggle to differentiate between refusing to answer an incorrect question and refusing to answer a question that should not be answered at all, such as those with false premises. A new approach using a hidden-state probe, rather than simple confidence scores, can better identify unanswerable questions, leading to improved precision in model responses. AI
IMPACT This research could lead to LLMs that are more reliable in refusing to answer inappropriate or incorrect questions, improving user trust and safety.
RANK_REASON The cluster contains two academic papers published on arXiv discussing LLM behavior and evaluation.
- alphaXiv
- arXiv
- Cab
- CatalyzeX
- Connected Papers
- CORE Recommender
- crêpe
- DagsHub
- Gotit.pub
- Hugging Face
- Litmaps
- Robin Staab
- ScienceCast
- scite Smart Citations
AI-generated summary · Google Gemini · from 3 sources. How we write summaries →
