Qwen2.5-7B-Instruct model shows causal transfer in symbolic domains

Researchers have conducted a pre-specified construction-confirmation test to investigate causal transfer in symbolic domains within a frozen Qwen2.5-7B-Instruct model. The study focused on a specific route, 'integer_mod16--letters16--successor->predecessor', which passed two PyVene splits with a confirmation intersection-union p-value of 0.000198. A subsequent NNsight experiment replicated these findings numerically, confirming the effect estimates and p-values. AI

IMPACT This research provides a method for testing causal transfer in symbolic domains, potentially improving interpretability and understanding of LLM behavior.

RANK_REASON The cluster contains an academic paper detailing a specific research finding about a model's capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Hugging Face
• NNsight
• PyVene
• Qwen2.5-7B-Instruct

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →