Paper title: [¬Re] Vocabulary-Activation Correspondence in Self-Referential LLM Processing
Paper authors: McCardle, John P.
Paper PDF URL: https://doi.org/10.5281/zenodo.19139301
Metadata URL: https://github.com/jmccardle/dadfar-vac-replication/blob/master/rescience-template/metadata.yaml
Code repository URL: https://github.com/jmccardle/dadfar-vac-replication
Code DOI: (to be minted after acceptance)
Data URL: https://zenodo.org/records/19139301
Data DOI: 10.5281/zenodo.19139301
Original article: Z. Dadfar, "When models examine themselves: Vocabulary-activation correspondence in self-referential processing," arXiv:2602.11358, 2026. https://arxiv.org/abs/2602.11358
Domain: Machine Learning
Language: Python
Type: Replication (failed)
Suggested editor: @gdetorakis (Georgios Detorakis — ML / Computational Neuroscience)
Suggested reviewers: (leave to editor discretion; domain: LLM interpretability / computational neuroscience)
Abstract:
We attempt to replicate the core claims of Dadfar (2026), who reports Vocabulary-Activation Correspondence (VAC) — a correlation between spontaneously adopted vocabulary and concurrent neural activation metrics — during extended self-referential processing in Qwen 2.5-32B-Instruct. Using the author's published configuration and Zenodo data, we identify four obstacles to replication: (1) the generation pipeline produces a bimodal output distribution, and all published terminal words derive from a degenerate summary mode; (2) the primary activation metric scales superlinearly with generation length, confounding VAC with a length artifact; (3) runs that complete 1,000 observations enter limit cycles, making terminal words phase artifacts; and (4) the TRACE-REPRO code repository targets a different model (Llama) than the paper's core claims (Qwen). Cross-model replication on Llama 3.1 70B yields zero compliant baseline runs. The replication fails on all four grounds.
content.pdf
Paper title: [¬Re] Vocabulary-Activation Correspondence in Self-Referential LLM Processing
Paper authors: McCardle, John P.
Paper PDF URL: https://doi.org/10.5281/zenodo.19139301
Metadata URL: https://github.com/jmccardle/dadfar-vac-replication/blob/master/rescience-template/metadata.yaml
Code repository URL: https://github.com/jmccardle/dadfar-vac-replication
Code DOI: (to be minted after acceptance)
Data URL: https://zenodo.org/records/19139301
Data DOI: 10.5281/zenodo.19139301
Original article: Z. Dadfar, "When models examine themselves: Vocabulary-activation correspondence in self-referential processing," arXiv:2602.11358, 2026. https://arxiv.org/abs/2602.11358
Domain: Machine Learning
Language: Python
Type: Replication (failed)
Suggested editor: @gdetorakis (Georgios Detorakis — ML / Computational Neuroscience)
Suggested reviewers: (leave to editor discretion; domain: LLM interpretability / computational neuroscience)
Abstract:
We attempt to replicate the core claims of Dadfar (2026), who reports Vocabulary-Activation Correspondence (VAC) — a correlation between spontaneously adopted vocabulary and concurrent neural activation metrics — during extended self-referential processing in Qwen 2.5-32B-Instruct. Using the author's published configuration and Zenodo data, we identify four obstacles to replication: (1) the generation pipeline produces a bimodal output distribution, and all published terminal words derive from a degenerate summary mode; (2) the primary activation metric scales superlinearly with generation length, confounding VAC with a length artifact; (3) runs that complete 1,000 observations enter limit cycles, making terminal words phase artifacts; and (4) the TRACE-REPRO code repository targets a different model (Llama) than the paper's core claims (Qwen). Cross-model replication on Llama 3.1 70B yields zero compliant baseline runs. The replication fails on all four grounds.
content.pdf