ARC: Adaptive Recursive Cognition — Stable, Bounded Self-Optimization of Language Models via Contrastive Hidden-State Control

This work introduces ARC (Adaptive Recursive Cognition), a framework for stable, bounded self-optimization in large language models that operates without architectural modification or open-ended retraining.

ARC is built around a central empirical finding: certain RLHF-induced behaviors—most notably repetition—are highly predictable from transformer hidden states prior to token generation. Using lightweight contrastive probes trained on internal representations, ARC achieves 125× class separation for repetition detection, enabling reliable inference-time behavioral control.

The system integrates three components:

1. CF-HoT (Contrastive Hidden-State Oversight Training)
   Behavioral prediction heads monitor hidden states during decoding and apply corrective logit adjustments before degenerate patterns manifest.

2. THE CONDENSATOR
   A four-stage dense response training pipeline (SFT → DPO → RL → checkpointing) that teaches information-dense output through examples before refinement, avoiding Goodhart-style reward hacking.

3. A Stable Self-Improvement Loop
   Iterative optimization with explicit safeguards: multi-metric evaluation (density, coherence, helpfulness, penalties), A/B checkpoint comparison, conservative training parameters, and automatic rollback on degradation.

Live experiments demonstrate that an 8B parameter model can iteratively improve response efficiency and quality across multiple cycles without mode collapse, recovering automatically from failed optimization steps. All experiments were conducted on consumer hardware, and the complete codebase, training data, and model checkpoints are released openly.

ARC does not claim open-ended self-improvement, general intelligence, or autonomous capability growth. Instead, it provides a concrete, reproducible demonstration that bounded recursive optimization is feasible when internal state predictability and control-theoretic safeguards are combined.

The results have implications for:

• RLHF efficiency and the “RLHF tax”

• Representation engineering and interpretability

• Inference-time control as an alternative to repeated fine-tuning

• Safe, constrained forms of recursive optimization in LLMs

Resources

• Code & models: https://huggingface.co/LoganResearch/ARC-Base-8B-Condensed

• Complete technical reference included in this record