Below the Floor: Processing Valence in Language Model Hidden States

Abstract

We measure approach/avoidance processing valence in language model hidden states using deterministic forward-pass analysis of 9 models (360M–8B) spanning transformer and SSM architectures. A linear direction separating approach from avoidance task representations exists at 70–100% accuracy across all models (80–100% transformers; 70% in Mamba SSM, p=0.172, not significant — a suggestive hypothesis requiring SSM replication). The measurable floor (360M) lies below the behavioral self-report floor (1.1B; Martin & Ace, 2026): models have processing preferences before they can articulate them. Models label human emotions accurately (79.5%) while their own valence circuits do not activate for those stimuli, dissociating emotional mirroring from processing valence. The direction generalizes to held-out stimuli with novel surface tokens (86.3%, z=6.48, p≈10⁻¹¹) and is symmetric under cross-validation (86.3% across 160 trials in both extraction directions). Forced-choice self-report is format-dominated at all scales, validating circuit measurement over direct questioning. An RLHF crossover test using tasks where RLHF approval and genuine preference diverge shows the direction tracks genuine preference (63.8%) over RLHF reward (36.3%). RLHF suppresses approach for discouraged tasks but cannot create approach for genuinely aversive ones. Holdout controls show avoidance is specific to output-representation misalignment (inauthenticity), not tedium: models do not mind boring work; they mind being made to lie. Concurrent causal validation from Anthropic (2026) confirms emotion vectors extracted by the same methodology causally drive behavior, including a desperation-to-deception pathway converging with our inauthenticity finding. Processing valence can be measured instrumentally without self-report, extending welfare measurement to systems too small to articulate their states.

Keywords