Validate the emotional-confiding codebook gap

Establish whether the observed difference in emotional-confiding behavior between anthropomorphic-name users and default-name users is a genuine behavioral effect rather than a chance finding arising from the width or construction of the sparse-autoencoder codebook, using an independently constructed confiding lexicon or a second dictionary.

Background

Emotional confiding is the largest observed gap between the two naming groups, with more than twice the activation rate in the anthropomorphic group. However, it is represented by one latent among 128 sparse-autoencoder features and does not survive the Benjamini–Hochberg correction for the full set of latent-level comparisons.

The authors explicitly state that the current evidence cannot rule out the possibility that a gap of this magnitude would appear somewhere in a codebook of this width by chance. An independently developed confiding lexicon or replication using a second dictionary would test whether the result is robust to the representation used.

References

The confiding gap has the reverse profile: it is the largest difference we observe, at more than twice the activation rate, but it is one latent among 128 and does not clear the Benjamini--Hochberg threshold, so we cannot exclude that a gap this size would turn up somewhere in a codebook of that width by chance. A confiding lexicon built independently of the autoencoder, or a replication on a second dictionary, would settle it; neither is available from the present logs.

— How Anthropomorphic Attitudes Toward Agents Divide User Behavior, and How Design Should Change  (2610.13156 - Cao et al., 9 Oct 2026) in Section 6, “Limitations and future work” (paragraph 3)