Published March 20, 2026 | Version v1
Dataset Open

Replication Data: Vocabulary-Activation Correspondence in Self-Referential LLM Processing

Authors/Creators

Description

Data and code for a replication study of Dadfar (2026) "Vocabulary-Activation Correspondence in Self-Referential LLM Processing"
(arXiv:2602.11358). The replication fails: VAC as reported does not survive length correction or cross-model validation. The underlying autoregressive dynamics (limit cycles, vocabulary narrowing) are real and reproducible, but they are properties of extended generation, not of self-referential processing specifically.

Files

README.md

Files (1.3 GB)

Name Size Download all
md5:bed9bc575d6a923b5c149d3dacc0a9af
7.9 kB Preview Download
md5:10cfdc465ebe81c96eafbe49f24650c6
1.3 GB Download
md5:a67ecc698774fadd5944106194b96233
2.3 MB Download

Additional details

Software

Repository URL
https://github.com/jmccardle/dadfar-vac-replication
Programming language
Python
Development Status
Active