Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• HypoTermInstruct SFT dataset (31,487 responses for 11,151 questions) teaches models epistemological humility by using questions about non-existent 'hypothetical' terms.method 95%
• Replacing generic instruction data with HypoTermInstruct significantly improves HypoTerm Score with median increases of 0.19% to 25.91%.result 92%
• Using HypoTermInstruct improves FactScore by +0.39% to +0.86%.result 92%
• HypoTermInstruct maintains stable performance on MMLU with minimal decreases of 0.26% to 0.35%.result 92%
• The study conducted 800 controlled LoRA SFT runs across Llama3.1-8B and Gemma3-4B models, testing 100 fine-tuning configurations with paired controls.method 95%
• Targeted SFT data teaching meta-cognitive skills can effectively reduce hallucination without preference/RL pipelines.result 88%
• The HypoTermQA-Enhanced benchmark is released for hallucination tendency evaluation, strengthened through multiple validations.method 90%
• LLMs often hallucinate partly because supervised fine-tuning (SFT) implicitly rewards always responding.technical 85%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card