Intervening at MLP output has minimal effect on trait expression compared to intervening at the identified attention layer output

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• MLPs primarily serve token-wise processing and are closely associated with knowledge storage, making attention heads the appropriate locus for sequence-level style modulationclaim0.784
  Architectural rationale for why Style Modulation Heads are in attention layers rather than MLPs
• MLP layers are much harder to get traction on than attention layers; understanding them requires individually interpretable neurons which are rarely foundclaim0.762
  Key limitation of the paper's approach; MLP layers make up 2/3 of standard transformer parameters
• High projection difference samples continue to induce stronger trait expression than random samples even after LLM-based filtering removes overtly trait-expressing samplesfinding0.738
  Shows persona vector filtering has complementary strengths to LLM judges, surfacing non-obvious problematic samples
• If in-context learning involves signed evaluation in the service of behavioral modification, then the thesis applies not only to training but to every inference-time interactionclaim0.736
  Extension of the thesis to deployed LLM inference via in-context learning
• Dataset-level projection difference is highly predictive of post-finetuning trait expression before finetuning occursfinding0.735
  Enables pre-finetuning data screening; Figure 8 shows strong dataset-level correlations across all three traits
• Directing response attention to complement syntax and/or mental state verbs (MSV) yields no significant alterations in IIT estimates compared to entire stimulus analysis.finding0.735
  Suggests LLMs do not represent complement/MSV linguistic features in the same way as they are crucial for human ToM development.
• Some MLP neurons and attention heads perform memory management by reading residual stream information and writing its negative to delete itclaim0.735
  Hypothesis based on observed negative cosine similarity between input and output weights of some neurons
• The And-Or algorithm may not be a true abstraction of the trained MLP's behaviour since it never achieves high IIA in later layers regardless of alignment map complexityhypothesis0.733
  Hypothesis raised in distributive law task analysis