New EMNLP Paper Introduces CTA to Explain How Linear Probes Emerge

A new EMNLP 2026 paper introduces Concept-Targeted Attribution (CTA), which uses attribution graphs to explain how linear probes emerge. Graph-based features predict probe accuracy with ρ=0.91 and R²=0.84.

2026-10-25 ~ 2026-10-25 · 2 related posts

• EMNLP paper introduces Concept-Targeted Attribution, tracing the circuits behind linear probes — ZhijingJin · 2026-10-25
• CTA paper: graph structure predicts probe accuracy at ρ=0.91, R²=0.84 — ZhijingJin · 2026-10-25