Salvato in:
| Autore principale: | |
|---|---|
| Natura: | Recurso digital |
| Lingua: | |
| Pubblicazione: |
Zenodo
2026
|
| Accesso online: | https://doi.org/10.5281/zenodo.18342792 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
Sommario:
- <p>This work introduces <strong>Adaptive Recursive Cognition (ARC)</strong>, a reproducible architecture for <strong>stable, bounded recursive operation of large language models</strong> under long-horizon use.</p> <p>ARC reframes language models not as static inference engines, but as <strong>controlled dynamical systems</strong> 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.</p> <p>The system is composed of four interacting control loops:</p> <ol> <li> <p><strong>Dense Adaptation Pipeline (SFT → DPO → RL)</strong><br>A staged training process that teaches high-density, non-hedging response behavior before optimization, preventing Goodhart-style collapse.</p> </li> <li> <p><strong>Control-Field Holonomy (CF-HoT)</strong><br>A predictive hidden-state control mechanism that detects and suppresses instability <em>before</em> token emission. A specialized repetition head achieves <strong>125× class separation</strong>, enabling early warning and smooth gating rather than reactive penalties.</p> </li> <li> <p><strong>RSI: Recursive Self-Improvement Loop</strong><br>A measured, reversible self-improvement loop using frozen judges, multi-metric evaluation, canary testing, and automatic rollback to ensure stability under recursive training.</p> </li> <li> <p><strong>The Fourth Loop: Tokenization Co-Evolution</strong><br>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.</p> </li> </ol> <p>Key design principles include:</p> <ul> <li> <p>Explicit separation of evaluation and optimization</p> </li> <li> <p>Commit/rollback as a first-class primitive</p> </li> <li> <p>Predictive control instead of reactive penalties</p> </li> <li> <p>Multi-metric, Goodhart-resistant evaluation</p> </li> <li> <p>Full reproducibility on consumer hardware</p> </li> </ul> <p>This release includes:</p> <ul> <li> <p>Complete architectural specification</p> </li> <li> <p>Training and control-loop definitions</p> </li> <li> <p>Tokenization diagnostics and delta-generation methodology</p> </li> <li> <p>Configuration defaults and reproducibility requirements</p> </li> <li> <p>Explicit safety boundaries and non-goals</p> </li> </ul> <p>ARC <strong>does not claim open-ended self-improvement, AGI, or autonomous operation</strong>. All optimization is bounded, reversible, and constrained by frozen judges and rollback thresholds. The contribution is architectural: demonstrating that <strong>stable recursive operation and bounded self-optimization are achievable with explicit control systems design</strong>.</p> <p>This work is released under <strong>CC BY 4.0</strong> to support verification, replication, and extension by the research community.</p>