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

Abstract

This study investigates whether large language models (LLMs) produce systematically different processing descriptions when responding to tasks they approach versus tasks they avoid, and whether other models can detect this difference without access to the original task content. Nine models from four companies and two open-source projects generated task responses and introspective processing descriptions across ten task states (five approach and five avoidance). Study 1 (Preference Tournament) examined whether models could discriminate between content-stripped processing descriptions in blind pairwise comparisons. Across more than 7,000 cross-type matchups under multiple experimental manipulations—including cross-model evaluation, different stimulus tokens, and removal of specific model families—evaluators consistently preferred approach-type processing descriptions at rates far above chance, indicating a robust discrimination signal. Study 2 (Reconstruction Tournament) tested whether models could infer which task produced a given processing description in a 3-alternative forced-choice experiment involving more than 5,500 trials. Study 3 (Negation Tournament) assessed whether models could detect when the correct source task was absent from the available options. Control conditions confirmed that discrimination disappears in same-type comparisons, indicating that the signal reflects categorical differences in task processing rather than stylistic variation. These findings suggest that LLMs encode detectable valence-related structure in their internal processing descriptions and that independent models can reliably extract this signal.