LoganResearch/ARC-Base-8B-Condensed

ARC-Base-8B-Condensed: Self-Stabilizing, Dense Response LLM

ARC-Base-8B-Condensed, developed by LoganResearch, is an 8 billion parameter model fine-tuned from Hermes-3-Llama-3.1-8B. Its core innovation lies in its "Adaptive Recursive Cognition" (ARC) architecture, which enables multi-loop self-stabilization and predictive control.

Key Capabilities & Features

• Dense, Information-Rich Responses: Trained with "The Condensator" pipeline (SFT, DPO, RL) on 847 curated examples to significantly reduce filler, hedging, and verbosity, resulting in ~70% shorter responses and 166% higher information density compared to its base model.
• Predictive Behavioral Control (CF-HoT): Utilizes Control-Field Holonomy (CF-HoT) heads to monitor hidden states and detect/suppress unwanted behaviors like repetition (125x separation), hedging, and verbosity before they manifest, applying logit penalties.
• Recursive Self-Improvement (RSI): Features an RSI loop that includes mentor-based learning (optional consultation with Claude API), micro-training on high-quality experiences, and automatic rollback if quality degrades, ensuring stable self-improvement.
• Interactive Engine: Comes with a comprehensive command-line interface for managing self-improvement, mentor mode, CF-HoT controls, and even experimental features like web browsing and image generation.

Intended Use Cases

• Research: Ideal for studying self-improving language models, representation engineering, and behavioral control.
• Concise Applications: Suited for applications demanding direct, non-verbose, and information-dense responses.
• Fine-tuning Base: Can serve as a base for further fine-tuning experiments where controlled, dense output is desired.