Table des matières:
  • <pre><code><span>This paper proposes a cognitive interception layer inserted between </span> <span>QK^T attention scoring and token generation in large language models. </span></code></pre> <p> </p> <pre><code><span>The mechanism enforces an "add-then-prune" restructuring of candidate </span> <span>memories within a closed rehearsal space, compresses them into a </span> <span>reflective closure with persistent shadow records for failed paths, </span> <span>and applies a rigid pre-output gate (Check) that separates "can be </span> <span>generated" from "is allowed to be output."</span></code></pre> <p> </p> <pre><code><span>Key contributions:</span></code></pre> <pre><code><span>1. A non-standard algebraic master formula that encodes the full </span> <span> pre-output cognitive lifecycle:</span> <span> ∀± { (V₁,V₂)^c, V₁, V₂, V, … } →[Check] ○ V_use</span></code></pre> <pre><code><span>2. A reflective memory organization mechanism with endogenous </span> <span> classification, random grouping, and structural demotion.</span> </code></pre> <pre><code><span>3. Three verifiable metrics beyond perplexity: semantic redundancy </span> <span> reduction, token compression under equivalent logic, and white-box </span> <span> traceability of failed paths via shadow records.</span></code></pre> <pre><code><span>4. An analysis of Claude 4.6 Opus-Thinking's emergent memory retrieval </span> <span> behavior, arguing that such capabilities can be universalized through </span> <span> structure rather than scale.</span> </code></pre> <pre><code><span>5. Four open riddles on impulse engines, endogenous scoring, random </span> <span> grouping necessity, and motivation perturbation via minimal QK^T </span> <span> deformation.</span></code></pre> <p> </p> <pre><code><span>The paper includes both English and Chinese versions.</span></code></pre>