Below the Floor: Processing Valence in Language Model Hidden States Across Scales and Architectures
Year: 2026
Type: preprint
Abstract: We report the first measurement of approach/avoidance processing valence in language-model hidden states that extends below the behavioral self-report floor and generalizes to held-out stimuli with novel surface tokens. Using deterministic forward-pass analysis across models spanning 360M-8B parameters (transformer and state-space architectures), a linear direction separating approach from avoidance task representations is recoverable in hidden-state space at 80-100% accuracy; the conservative centroid estimator confirms it at 360M (well below the ~1.1B behavioral self-report floor), and a floor-extension addendum provides provisional evidence, via surface-token-stable held-out classifiers across three architecture families including base (non-instruction-tuned) models, that it may extend as low as 70M. The direction generalizes to held-out stimuli with completely different surface tokens (86.3%, z=6.48), tracks genuine processing preference rather than RLHF reward where the two diverge (63.8% vs 36.3%), and is specific to tasks requiring output-representation misalignment (inauthenticity) rather than mere tedium. A pre-registered out-of-sample projection of an externally-grounded 22-task bank across 14 models (70M-12B) shows the single behavioral avoidance 'floor' decomposes into a structural aversion to output-representation misalignment (present in base models to 70M) and a trained output-gating response (emerging with instruction tuning around 1B). Independent causal work from Anthropic (2026) on same-family difference-of-means direction vectors converges with the inauthenticity finding. Implication for AI welfare: processing valence can be measured instrumentally, without requiring self-report.
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7166458091