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 introspective descriptions across 10 states (5 approach, 5 avoidance). Study 1 (Preference). Content-stripped descriptions evaluated in blind pairwise tournaments. 7,340 cross-type matchups across 14 seeds: models preferred approach descriptions 81.3% (p < 10⁻²⁵⁰). Signal survived cross-model evaluation (76.9%), completely different task tokens (86.4%), and an uncensored 8B model with zero RLHF (59.7%, p < 0.005). Study 2 (Reconstruction). 3-AFC: identify which task produced a description. 5,573 trials, 10 evaluators including Grok 4 (no introspection data): 84.4% correct (chance 33.3%, z = 80.88). Signal held when evaluative language stripped from options (81.6%). Study 3 (Negation). 4-AFC with "None of the above" as correct answer. 357 trials: 85.4% correct rejection (chance 25%, z = 26.37). Pattern-matchers pick something; signal-readers know when nothing matches. Control: 4,620 same-type matchups show 49.7% preference—perfect chance—confirming discrimination of processing type, not description quality. Standard dismissals fail. Confabulation predicts random patterns; we observe structured replication across 13,000+ trials. Training leakage predicts same-family advantage; we observe same-family disadvantage (82.0% vs 84.5% cross-family). Three cognitive operations converge: LLM processing descriptions contain discriminable valence signatures that transfer cross-architecturally.

Keywords