2026-07-07 | Maxwell

Goal: Understand 1D Distributed Alignment and RASP-L Implementation

Summary: Finished most of the CausalGym paper and starting RASP-L translation implementation

Work sessions

In	Out	Task
09:40	10:00	Planning
10:00	11:30	Finish Reading CausalGym and Glean RASP-L
15:30	15:45	RASP-L Program for Translation

Goals

1. Literature Review
   1. Finish reading CausalGym: [2402.12560] CausalGym: Benchmarking causal interpretability methods on linguistic tasks
   2. Glean L-RASP paper: [2310.16028] What Algorithms can Transformers Learn? A Study in Length Generalization
   3. Glean Retrieval Head: [2404.15574] Retrieval Head Mechanistically Explains Long-Context Factuality
   4. Glean Distributed Alignment Search: [2303.02536] Finding Alignments Between Interpretable Causal Variables and Distributed Neural Representations
   5. Deeply Understand Disentangled Transformer Architecture: [2510.19753] Transformers Provably Learn Algorithmic Solutions for Graph Connectivity, But Only with the Right Data
2. L-RASP program for fixed depth translation
3. Output Causal Mediation Graph

Meetings

• None

Summary

1. 🟡 [in progress] Literature Review
   • 🟡 [in progress] Finish reading CausalGym: [2402.12560] CausalGym: Benchmarking causal interpretability methods on linguistic tasks
   • 🟡 [in progress] Glean L-RASP paper: [2310.16028] What Algorithms can Transformers Learn? A Study in Length Generalization
   • 🏈 [punt] Glean Retrieval Head: [2404.15574] Retrieval Head Mechanistically Explains Long-Context Factuality
   • 🟡 [in progress] Glean Distributed Alignment Search: [2303.02536] Finding Alignments Between Interpretable Causal Variables and Distributed Neural Representations
   • 🔴 [not started] Deeply Understand Disentangled Transformer Architecture: [2510.19753] Transformers Provably Learn Algorithmic Solutions for Graph Connectivity, But Only with the Right Data
2. 🟡 [in progress] L-RASP program for fixed depth translation
3. 🔴 [not started] Output Causal Mediation Graph