2026-07-15 | Maxwell

Goal: Understand RASP-L constituency parsing algorithm and causal mediation

Summary: Understanding loading ...

Work sessions

In	Out	Task
13:30	13:45	Planning
13:45	14:30	Causal Intervention

Goals

1. [continue] RASP-L Diagram for translation pseudo code
2. [continue] Finish causal abstraction code first pass

Meetings

• None

Summary

1. ✅ Ran causal abstraction code
2. 🟡 [continue] RASP-L Diagram for translation pseudo code on final parts

Hypotheses: 1. Follow up hypothesis to prompt 1 type - Information moves from blocks.1.hook_resid_pre to blocks.1.hook_resid_mid via attention
- Which attention head is responsible/could be a combination of both? 2. Logit Diff and patch only 1 head at a time 3. What information is moving → what is the OV projection writing out? - Token information (since tokens not in the prompt are predicted) - Positional information (perhaps at earlier layers)