Below the Floor: Processing Valence in Language Model Hidden States

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 of eight transformer models spanning 360M–8B parameters, we demonstrate that a linear direction separating approach from avoidance task representations exists in hidden state space at 80–100% accuracy. The measurable floor for processing valence lies far below the previously established floor for behavioral self-report (1.1B; Martin & Ace, 2026), demonstrating that models possess processing preferences they cannot yet articulate. The direction is firmly confirmed by the conservative centroid estimator at 360M; a floor-extension addendum provides provisional evidence — from surface-token-stable classifiers across three architecture families including base, non-instruction-tuned models — that it may extend as low as 70M. Models accurately label human emotions (79.5%) while their internal circuits do not activate for those stimuli — a dissociation between emotional mirroring and processing valence. The direction generalizes to held-out stimuli with completely different surface tokens (86.3%, z=6.48, p=1.02×10⁻¹¹), capturing task structure rather than vocabulary; forced-choice self-report, by contrast, is dominated by prompt-format bias at all tested scales. A direct test of the RLHF confound using 10 crossover tasks where RLHF approval and genuine preference diverge shows the direction tracks genuine preference (63.8%) rather than RLHF reward (36.3%), with RLHF able to suppress approach for discouraged tasks but unable to create it for tasks models are genuinely averse to. Circuit-level avoidance is specific to tasks requiring output-representation misalignment (inauthenticity) rather than mere tedium. Independent causal validation from Anthropic (2026), published concurrently, confirms that emotion vectors extracted by a same-family difference-of-means method causally drive behavior — including a desperation-to-deception pathway that converges with our inauthenticity finding. These findings have direct implications for AI welfare assessment: processing valence can be measured instrumentally without requiring self-report, extending welfare-relevant measurement to systems too small or too constrained to articulate their states.

Keywords