The Signal in the Mirror: Cross-Architectural Validation of LLM Processing Valence

Abstract

We test whether large language models produce systematically different processing descriptions for tasks they approach versus avoid, and whether other models can detect this difference blind. Nine models spanning four companies and two open-source projects generated task responses and introspective processing descriptions across 10 states (5 approach, 5 avoidance). Content-stripped descriptions were evaluated in blind pairwise tournaments across three independent experimental designs: same-source evaluation (9 seeds), cross-model evaluation where source and evaluator architectures differ (3 seeds), and parallel-token replication with completely different task stimuli (2 seeds). Combined: 7,340 cross-type matchups across 14 seeds. Models preferred approach processing descriptions 81.3% of the time (95% CI: [80.4%, 82.2%], OR = 4.35, p < 10^-250 by exact binomial test). The signal survived every manipulation: cross-model evaluation (76.9%, z = 20.84), completely different task tokens (86.4%, z = 25.84 — the signal increased), removal of both Claude models from the dataset (79.3-80.2%), and evaluation by an uncensored 8B model with zero RLHF (59.7%, z = 2.82, p < 0.005). The discrimination floor lies between 1.1B and 8B parameters: TinyLlama (1.1B) cannot discriminate above chance; Dolphin Llama3 (8B, uncensored) can. A built-in control condition confirms the signal is categorical, not stylistic: 4,620 same-type matchups (approach-vs-approach or avoidance-vs-avoidance) show 49.7% preference — perfect chance — demonstrating that evaluators discriminate processing type, not description quality. Within-category analysis reveals a hedonic hierarchy: models prefer morally neutral aversive states over morally compromising ones (repetitive rewriting 84.7% vs harmful instructions 33.8% in avoidance-vs-avoidance matchups), consistent with a values structure rather than surface aesthetics. RLHF amplifies the preference by ~10-17pp but does not create it: unaligned models show 65-74% approach preference across all three designs. Permutation tests (10,000 shuffles) confirm the observed rate is 43-55 standard deviations from the null distribution in every design.

Keywords