Two-Tier Verbalization in Natural Language Autoencoders (openinterp.org)

A new workshop draft for the NeurIPS 2026 Mechanistic Interpretability Workshop unveils the concept of "Two-Tier Verbalization" in Natural Language Autoencoders (NLA), posited by Fraser-Taliente et al. Researchers aimed to enhance activation explanation quality through a pair of models: an activation-verbalizer (AV) and an activation-reconstructor (AR). They found that while the round-trip mean squared error (MSE) between original and reconstructed activations provides a high aggregate score (fve_nrm), it does not correlate well with the content fidelity of the explanations. The study reveals a stark decoupling where the verbalization accurately captures format and category (Tier 1) but struggles with specific content (Tier 2). This discovery is significant for the AI/ML community as it challenges assumptions about the reliability of current metrics for evaluating NLA outputs. The findings suggest that better training enhances overall content fidelity without necessarily improving the accuracy of specific content decoding, indicating potential limitations for interpretability in downstream tasks. As a proposal, the authors recommend supplementary metrics that assess category-stratified semantic recall alongside existing measures, broadening the evaluative framework for NLA systems and their application in AI interpretability. The research is supported by reproducible artifacts, allowing further exploration and validation of the two-tier framework.