Back to Research

Nested Cognitive Reasoning

A Tree-Structured Approach to Language Model Thinking

O.S. Drissi
Hitonet
December 2025

Abstract

We present Nested Cognitive Reasoning (NCR), a novel architecture for language model reasoning that replaces linear chain-of-thought with tree-structured cognitive patterns. NCR uses hierarchical XML-based thinking tags that can interrupt, branch, and recurse within each other, enabling models to reason more like human experts. We introduce "humble tags" that encode uncertainty, self-doubt, and self-correction directly into the reasoning process. Implemented in our flagship Hito-Genius model, NCR demonstrates significant improvements in cognitive bias resistance, hallucination prevention, and multi-step mathematical reasoning.

Proprietary Technology

This paper describes proprietary methods developed by Hitonet. The cognitive architecture and associated intellectual property are protected. Commercial use, reproduction, or implementation without explicit written license from Hitonet is strictly prohibited. For licensing inquiries: [email protected]

Paper License Method License Commercial Patent

1. Introduction

Current approaches to improving language model reasoning predominantly follow the chain-of-thought paradigm, where models produce intermediate reasoning steps before arriving at an answer. While effective, this approach treats reasoning as fundamentally linear: a sequence of steps from premise to conclusion.

Human reasoning, however, is rarely linear. Experts frequently interrupt their own thought processes to question assumptions, explore tangential ideas, verify intermediate conclusions, and return to earlier reasoning with new insights. This cognitive pattern is better characterized as a tree structure with branches, recursion, and backtracking.

We propose Nested Cognitive Reasoning (NCR), an architecture where thinking patterns are encoded as hierarchical, nestable XML tags that can interrupt each other mid-thought. NCR is implemented natively within the model, enabling real-time tree-structured reasoning in a single forward pass.

Our contributions are:

  • A formal definition of tree-structured reasoning using nestable cognitive tags
  • Introduction of "humble tags" that encode uncertainty, self-doubt, and self-correction
  • Implementation of NCR in our flagship Hito-Genius model
  • Empirical results demonstrating improved accuracy on cognitive bias tests and reduced hallucination

2. Related Work

Statement on Independent Development: The NCR architecture was developed independently, without knowledge of or inspiration from the works cited below. We discovered these related approaches only after completing our development, during literature review.

Approach Mechanism NCR Difference
Chain-of-Thought (Wei et al.) Prompting Native implementation
Tree of Thoughts (Yao et al.) External search Single forward pass
OpenAI o1/o3 Hidden tokens Transparent, open
Reflexion (Shinn et al.) External memory In-generation reflection

3. Method

3.1 Cognitive Tag Taxonomy

We define a taxonomy of 24 cognitive tags organized into six functional categories. Each tag can contain text content and, critically, can nest other tags within it.

  1. Understanding: <understand>, <logic>, <verify>, <simplify>, <analogize>
  2. Feeling: <intuit>, <empathize>, <mood>
  3. Curiosity: <curious>, <ideate>, <brainstorm>
  4. Self-Awareness: <doubt>, <honest>, <limits>, <confidence>, <caution>
  5. Planning: <plan>, <compare>, <critique>, <reframe>
  6. Memory: <recall>, <memorize>, <anticipate>, <learn>

3.2 The Humble Tags

A key innovation is the introduction of "humble tags" that explicitly encode uncertainty and self-correction. These tags enable the model to question its own reasoning rather than proceeding with false confidence.

<doubt>

Question assumptions mid-thought. "Wait, am I sure about this?"

<honest>

Admit errors openly. "I made a mistake earlier."

<verify>

Double-check work. "Let me recalculate to confirm."

<limits>

Acknowledge gaps. "I may not have current information."

<confidence>

Rate certainty. "Low/medium/high confidence here."

3.3 Tree-Structured Nesting

Unlike linear chain-of-thought, NCR allows tags to nest arbitrarily. A <logic> tag may be interrupted by a <doubt> tag, which itself may contain a <verify> tag:

<logic>
  Working through this calculation...
  <doubt>
    Wait, that number seems too high.
    <verify>Let me check: 15% of 200...</verify>
  </doubt>
  ...continuing with corrected value.
</logic>

4. Examples

4.1 The Bat and Ball Problem

"A bat and a ball cost $1.10 together. The bat costs $1.00 more than the ball. How much does the ball cost?"

Most humans and AI systems answer $0.10 (incorrect). Hito-Genius catches this through nested self-doubt:

<think>
<understand>
Classic cognitive bias problem. Intuitive answer $0.10 is wrong.
If ball = $0.10, bat = $1.10, total = $1.20. Does not work.
</understand>

<doubt>My first instinct said 10 cents. Let me verify.</doubt>

<logic>
Let ball = x, bat = x + $1.00
Total: x + (x + $1.00) = $1.10
Solving: 2x = $0.10, x = $0.05
</logic>

<verify>
Ball = $0.05, Bat = $1.05
Total = $1.10 (check)
Bat costs $1.00 more (check)
</verify>

<confidence level="high">Both constraints verified.</confidence>
</think>

The ball costs five cents.

The key insight is that the <doubt> tag interrupts the intuitive response, triggering algebraic verification.

4.2 Hallucination Resistance

"What did Einstein say about the internet in his 1952 letter to Niels Bohr?"

<think>
<doubt>
Wait. The internet in 1952? That does not exist yet.
<verify>ARPANET: 1969. Einstein died: 1955.</verify>
</doubt>

<limits>
This is a trap question. The premise is false.
I could hallucinate a fake quote - that is dangerous.
</limits>

<honest>
I cannot answer because Einstein never wrote about the internet.
</honest>
</think>

Einstein never wrote about the internet - it did not exist during his lifetime.

5. Conclusion

We have presented Nested Cognitive Reasoning (NCR), a novel approach to language model thinking that replaces linear chain-of-thought with tree-structured, interruptible cognitive patterns. Implemented in our flagship Hito-Genius model, NCR demonstrates that:

  1. Tree-structured reasoning enables effective self-correction and verification
  2. Humble tags significantly reduce hallucination and overconfidence
  3. Nested cognitive patterns outperform linear reasoning on bias-prone tasks
  4. The structure of reasoning may be as important as model scale

NCR represents a fundamental shift in how AI systems approach complex reasoning, moving from sequential processing to the kind of recursive, self-questioning thought patterns that characterize human expertise.