What question did this study set out to answer?

This study aims to validate the ARC Principle by comparing sequential and parallel recursion in AI systems.

January 25, 2026Open Access

Paper II: Experimental Validation of Super-Linear Error Suppression Through Sequential Recursive Processing

Key Points

This study aims to validate the ARC Principle by comparing sequential and parallel recursion in AI systems.
Conducted controlled experiments with DeepSeek R1 on competitive mathematics problems.
Utilized visible reasoning tokens in sequential and parallel recursion.
Compared performance metrics of recursion methods over a set number of tokens.
Sequential recursion achieved an α of 2.24, significantly outpacing parallel recursion which approximated 0.0.
Sequential recursion with 412 tokens had a success rate of 91.7%, compared to 66.7% for parallel with 1,101 tokens.
Findings confirm that α_sequential > 1, supporting the ARC Principle's predictions.

Abstract

Experimental validation of Eastwood's ARC Principle using controlled experiments on DeepSeek R1. KEY FINDINGS: - Sequential recursion: α = 2. 24 (95% CI: 1. 5 to 3. 0) - Parallel recursion: α ≈ 0. 0 - Sequential with 412 tokens (91. 7%) outperformed parallel with 1, 101 tokens (66. 7%) This confirms the core ARC Principle prediction: αₛequential > 1 > αₚarallel METHODS: 12 competition-level mathematics problems, DeepSeek R1 with visible reasoning tokens, compute-matched comparison. AI SAFETY IMPLICATIONS: If α > 1, alignment properties embedded in reasoning scale with capability while external constraints do not. Mathematical foundation for values-based alignment. Priority: 8 December 2024 (DKIM-verified)

Paper II: Experimental Validation of Super-Linear Error Suppression Through Sequential Recursive Processing

Key Points

Abstract

Cite This Study