Horizons: A New Architecture to Enhance Large Language Model Reasoning via Structured Internal Dialogue

Horizons is an innovative reasoning architecture that compels large language models to engage in constructive conflict thinking before providing answers. Using a structured internal dialogue involving three voices—Optimizer, Subverter, and Synthesizer—it generates more original and in-depth responses. This architecture was developed by joacokhzyx, sourced from GitHub, and released on May 24, 2026.

Current large language models often tend to generate predictable, consensus-driven shallow responses when dealing with complex problems. This phenomenon is particularly evident when facing issues requiring deep thinking, creative breakthroughs, or strategic decisions. Models tend to choose the safest and most obvious answers instead of the optimal solutions that have undergone sufficient examination and argumentation. Horizons is a reasoning architecture designed specifically to address this pain point.

The core innovation of Horizons lies in introducing three distinct internal voices:

• Optimizer: Represents the traditional, rule-following thinking path and proposes standard solutions.
• Subverter: Raises fundamental challenges to the Optimizer's proposals, seeking hidden assumptions and blind spots.
• Synthesizer: Integrates insights from both sides into a final comprehensive answer. Its workflow follows a strict pattern: Optimizer proposes an initial plan → Subverter critiques → Optimizer responds and revises → Subverter further explores → Synthesizer generates the final answer.

Horizons has developed specially optimized versions for different professional domains:

• Web Development Edition: Tuned for scenarios like front-end framework selection and architecture design.
• Software Engineering Edition: Focuses on technical decisions such as system design and code architecture.
• Neuroscience Edition: Handles research tasks like experimental design and data interpretation.
• Advanced Mathematics Edition: Supports deep reasoning tasks like mathematical proofs and theorem derivation.

Horizons acknowledges the higher cost of token consumption; its internal dialogue process significantly increases the length of reasoning traces. The tradeoffs are as follows:

• Cost: Token usage per query increases significantly.
• Benefit: Solutions undergo more rigorous stress testing and have fewer deep-seated flaws. This architecture is most suitable for scenarios where the cost of shallow answers is far higher than the additional token cost, such as strategic decision-making and architecture design.

Horizons provides methodological insights for AI-assisted decision-making: quality is sometimes more important than speed.

• Developers: Can integrate existing models with the Horizons mode via simple system prompt configuration to enhance reasoning capabilities.
• Researchers: Demonstrates the effectiveness of using architectural design to compensate for model cognitive limitations; metacognitive intervention may be more effective than expanding model scale.

Horizons reminds us that the potential of large language models goes far beyond generating fluent text. Through carefully designed reasoning architectures, models can be guided to engage in deeper cognitive activities and produce insightful and original content. It represents an important direction for AI development: exploring how to more effectively utilize the cognitive capabilities of existing models and achieve qualitative leaps through architectural innovation.