Hello world

Recent advances in Artificial Intelligence, particularly in Large Language Models (LLMs) and Vision-Language Models (VLMs), have been driven by sophisticated pre-training methodologies applied to Transformer-based architectures. These models achieve strong performance across benchmarks and provide practical utility through conversational interfaces and API services.

Despite their capabilities, language models face several inherent challenges that persist regardless of parameter count or context length:

1. Hallucinations. The veracity of the response cannot be 100% guaranteed because models generate the most probable token rather than necessarily factual information.

2. Context limit. The Transformer architecture's attention mechanism has quadratic complexity, increasing processing time and memory as sequences grow. Models also struggle with the "Lost in the middle" problem, failing to retrieve key information from long contexts.

3. Up-to-date information. Pre-training limitations restrict models to knowledge up to a specific date, affecting their ability to provide current information.

4. Limited digital interaction. LLMs cannot natively interact with external services like APIs, files, or software systems.

Architectural Overview

Maisa's KPU (Knowledge Processing Unit) comprises three main components:

1. Reasoning Engine. The system's brain that orchestrates step-by-step planning, relying on LLMs or VLMs and available tools.

2. Execution Engine. Receives and executes commands while sending results back for re-planning.

3. Virtual Context Window. Manages data flow between engines, ensuring reasoning stays within context limits while data remains external, and integrates external sources.

This design, inspired by operating systems, decouples reasoning from execution, allowing LLMs to focus exclusively on reasoning while reducing hallucination risks and improving factual reliability.

Benchmarks

Setting a new paradigm at Math Reasoning

The KPU was tested against established reasoning benchmarks, outperforming models including GPT-4, Mistral Large, Claude Opus, and Google Gemini.

On the GSM8k benchmark—consisting of 8,500 grade school math problems—the KPU surpassed all competing models using a zero-shot approach without prompt engineering or iterative attempts.

The MATH benchmark, containing 12,500 competition-level problems with step-by-step solutions, similarly demonstrated the KPU's superior performance in generating LaTeX-formatted responses without additional configurations.

Boosting reading and discrete comprehension

On the DROP benchmark, which requires reasoning over paragraphs including operations like addition and counting, the KPU established a new state-of-the-art result in a zero-shot configuration.

Beyond advanced reasoning challenges

The Big Bench Hard (BBH) benchmark tests tasks designed beyond current model capabilities. The KPU performed particularly well on tasks requiring logical reasoning, such as Boolean expressions and arithmetic.

Summarizing the system's evaluation

The KPU has established new standards across multiple reasoning domains:

• Elementary math reasoning (GSM8k)

• Advanced mathematics (MATH)

• Reading comprehension (DROP)

• Complex reasoning tasks (BBH)

All results and methodology are reproducible through Maisa's published documentation.

KPU Use Availability

Currently in beta phase, the KPU is not yet available for commercial use. Interested parties can join a waiting list for early access, or contact Maisa at contact@maisa.ai or through their contact form. Updates are posted on X (@maisaai_).