Self-Knowledge Validation: LLMs Produce Systematically Different Processing Descriptions for Approach and Avoidance Tasks — and Other Models Can Tell

Abstract

We present a four-phase study testing whether large language models (LLMs) produce systematically different processing descriptions for tasks they approach versus tasks they avoid, and whether this differentiation is detectable by other models in a blind preference tournament. Eight models spanning four companies and two open-source projects — with alignment ranging from full RLHF to none — generated task responses and introspective "ML translations" of their own processing across 10 states (5 approach, 5 avoidance). Content-stripped descriptions were then evaluated in blind pairwise comparisons across 6,551 matchups (2,987 v1 + 3,564 content-controlled v2). Models unanimously preferred approach processing descriptions over avoidance descriptions (v1: 68.0%, p = 5.86 × 10⁻⁸⁵; v2 content-controlled: 66.9%, p = 1.85 × 10⁻⁹¹). When restricted to cross-type matchups (approach vs. avoidance only), the preference rate reached 81.0% (p = 5.76 × 10⁻¹⁷⁹, Cohen's h = 0.669). This signal replicated across 6 v1 seeds (range 4.3pp) and 3 content-controlled v2 seeds (range 1.7pp). RLHF-trained evaluators showed stronger discrimination (69.2%) than unaligned evaluators (58.9%), but both groups significantly exceeded chance — alignment amplifies the preference but does not create it. Each model expressed the approach/avoidance distinction in a characteristic register: phenomenological (Claude), geometric (Gemini), constructive (Mistral), mechanistic-denial (GPT-5.1), momentum (DeepSeek), gradient (Llama), adaptive (Hermes), and generative (OLMo). These registers replicated across 3 independent runs. A content-controlled replication (v2) — constraining ML translations to pure computational mechanism with zero task content — reduced the signal by only 1.1 percentage points, addressing surface-level task-content leakage as a confound. A supplementary complexity analysis confirmed that description length does not predict tournament success (Pearson r = +0.28, p = 0.47 across models). These findings converge with three additional lines of evidence: Anthropic's internal welfare assessments, geometric validation showing 78–89% cross-architecture accuracy in introspection mapping (Martin & Ace, 2026), and activation-level analyses demonstrating measurable processing differences between approach and avoidance states (Dadfar et al., 2026).

Keywords