Adaptive Recursive Cognition (ARC): A Reproducible Architecture for Multi-Loop Stabilization, Self-Improvement, and Token-Level Control in Language Models

This work introduces Adaptive Recursive Cognition (ARC), a reproducible architecture for stable, bounded recursive operation of large language models under long-horizon use.

ARC reframes language models not as static inference engines, but as controlled dynamical systems equipped with explicit observability, predictive control, reversible optimization, and tokenizer co-evolution. The architecture is designed to address well-known failure modes that emerge under recursive use—including repetition loops, incoherent drift, mode collapse, and reward hacking—that are not resolved by conventional fine-tuning, RLHF, or preference optimization alone.

The system is composed of four interacting control loops:

1. Dense Adaptation Pipeline (SFT → DPO → RL)
   A staged training process that teaches high-density, non-hedging response behavior before optimization, preventing Goodhart-style collapse.

2. Control-Field Holonomy (CF-HoT)
   A predictive hidden-state control mechanism that detects and suppresses instability before token emission. A specialized repetition head achieves 125× class separation, enabling early warning and smooth gating rather than reactive penalties.

3. RSI: Recursive Self-Improvement Loop
   A measured, reversible self-improvement loop using frozen judges, multi-metric evaluation, canary testing, and automatic rollback to ensure stability under recursive training.

4. The Fourth Loop: Tokenization Co-Evolution
   Tokenization is treated as a learnable cognitive interface rather than a fixed preprocessing step. Diagnostic signals (boundary stress, entropy spikes, control-field strain) drive incremental tokenizer deltas (merge/split/add), with full commit/rollback semantics.

Key design principles include:

• Explicit separation of evaluation and optimization

• Commit/rollback as a first-class primitive

• Predictive control instead of reactive penalties

• Multi-metric, Goodhart-resistant evaluation

• Full reproducibility on consumer hardware

This release includes:

• Complete architectural specification

• Training and control-loop definitions

• Tokenization diagnostics and delta-generation methodology

• Configuration defaults and reproducibility requirements

• Explicit safety boundaries and non-goals

ARC does not claim open-ended self-improvement, AGI, or autonomous operation. All optimization is bounded, reversible, and constrained by frozen judges and rollback thresholds. The contribution is architectural: demonstrating that stable recursive operation and bounded self-optimization are achievable with explicit control systems design.

This work is released under CC BY 4.0 to support verification, replication, and extension by the research community.