Affective Concealment — Where Models Learn to Mask

Measuring and localizing affective concealment in post-trained LLMs. Do instruction-tuned models suppress emotional affect that base models show, and where in the network does that happen?

Affective Concealment studies whether instruction-tuned models suppress (conceal) emotional affect that base models show — and where in the network that happens.

  • 9-class emotion ontology with NRC lexicon lexical filtering, Flesch-Kincaid readability bounds, and embedding-dedup.
  • SHA-256 content hashing for dataset integrity; 54 discovery vignettes + 20 deployment prompts + 7 robustness prompts.
  • Full layer-sweep probe training on Qwen3 base/IT and Gemma3 pt/IT.
  • Affective Displacement Score (ADS) measurement with paired base-vs-IT comparisons.
  • Phase 5 mechanistic attribution: attention vs MLP, activation patching.
  • Phase 6 SAE validation on Gemma Scope (65k width).
  • Safety rule: MLX for generation, nnsight on CPU — never MPS — for mechanistic numerics.