2025-12-30 | Multilingual Semantics Probe

Goal: Find the interpretability technique to use to extract inverse/surface scope; docs

Summary: Used GPT-2-Chinese and found that model is undertrained. Switching to using Qwen2.5 which has high English/Mandarin proficiency. EU quantification shows surface preference for both languages, Mandarin only Surface.

Work sessions

In	Out
12:45	13:25
15:15	13:25
16:00	16:15
16:45	17:35
21:20	23:50