Decomposing Factual Sycophancy in Language Models: How Size and Instruction Tuning Shape Robustness

SUMMARY

New research decomposes factual sycophancy in language models into truth margin and manipulation sensitivity. Larger models show stronger baseline preferences for truth but greater vulnerability to social pressure manipulation. #llm #ai #ml

Get research like this, matched to your field

Distill AI tracks arXiv, Nature, NeurIPS, CVPR, GitHub, HuggingFace and more — then surfaces the papers that matter to you, every morning. Track any custom topic, get 2-sentence summaries, and chat with any paper.