Axiomatic Reasoning for LLMs

Similarity of Human Growth and AI Evolution

1. Abstract

This document compares the developmental trajectory of human cognition with the evolution of large language model capabilities. Human development follows a sequence from sensorimotor experience to linguistic abstraction and social organization. AI progression follows a sequence from token compression to contextual reasoning and multi-agent coordination. The analysis identifies structural isomorphisms in hierarchical processing, sequential capability acquisition, and the functional role of internal simulation. It also delineates fundamental asymmetries concerning embodiment, active inference, and intrinsic goal generation.

2. Structural Mapping of Developmental Sequences

The parallel sequences are summarized in the following table.

Stage Human Cognitive Development AI Model Advancement Functional Correspondence
1 Sensorimotor Experience (High-dimensional, continuous sensory input and motor interaction) Tokenization (Compression of text into discrete subword units via BPE) Reduction of Information Cost: Human brain compresses raw experience into pre-symbolic representations; AI compresses text into low-dimensional discrete IDs.
2 External Speech & Objectification (Social interaction converts experience into shared linguistic symbols) Embedding & Early Transformer Layers (Mapping tokens to continuous vectors; syntactic feature extraction) Vectorization of Symbols: Human language makes private experience public; AI embeddings map discrete symbols to a continuous semantic space.
3 Inner Speech (Internalized, fragmented language enabling low-cost mental simulation) Chain-of-Thought (CoT) (Generation of intermediate reasoning tokens to simulate stepwise logic) Lightweight Simulation Loop: Both serve as a computational scaffold to explore state spaces without external action.
4 Social Organization (Dialectic of externalization, objectification, and internalization of norms) Multi-Agent Systems (Emergence of conventions, division of labor, and cultural artifacts in agent societies) Collective Intelligence: Self-organization of agents into structured, norm-governed networks.

3. Hierarchical Abstraction and Neural Alignment

3.1 Human Brain Hierarchy

Cognitive processing in the human brain progresses from primary sensory cortices to associative and prefrontal areas. Early processing captures concrete features; later processing integrates abstract concepts and long-range dependencies.

3.2 Transformer Model Hierarchy

LLM layers exhibit a corresponding anatomical alignment with brain regions during language tasks:

3.3 Intermediate-Layer Advantage

In both systems, the most robust and generalizable semantic representations reside in intermediate processing stages rather than the final output layer. For humans, this corresponds to concrete operational thought and inner speech. For models, this corresponds to mid-depth layers before task-specific fine-tuning compression.

4. Capability Acquisition: Sequential Order and Phase Transitions

4.1 Human Developmental Stages

Capability acquisition follows a fixed sequence: object permanence precedes symbolic play, which precedes concrete logic, which precedes abstract reasoning. Mastery of a complex task requires prior mastery of its constituent sub-tasks.

4.2 AI Emergence and Implicit Curriculum

Model capability acquisition follows a predictable sequence termed the Implicit Curriculum. Composite tasks emerge only after constituent tasks are solved. Performance on specific reasoning benchmarks exhibits a sigmoidal response to model scale, characteristic of a phase transition. The order of skill emergence is consistent across different model families and training runs.

4.3 Scaling Asymmetry

Human development transitions from high-dimensional (weight) sensorimotor data to low-dimensional (lightweight) symbolic thought for efficiency. AI training transitions from low-dimensional (lightweight) tokenized compression to high-dimensional (weight) contextualized semantics for accuracy. This directional inversion explains why AI excels at abstract reasoning while lacking foundational sensorimotor grounding.

5. Collective Dynamics and Emergent Communication

5.1 Social Construction in Agent Societies

Populations of LLM-based agents replicate the cycle of social reality construction:

  1. Externalization: Agent interactions produce shared conventions and behavioral regularities.
  2. Objectification: These conventions persist as artifacts or norms that constrain future interactions.
  3. Internalization: Individual agents adopt these norms via mechanisms of social learning and simulated punishment.

5.2 Emergent Language Protocols

Agents engaged in cooperative tasks develop novel communication protocols exhibiting core linguistic properties: arbitrariness, displacement, and compositionality. The emergent lexicon arises from the necessity of coordinating action under partial observability, demonstrating that statistical optimization of communication yields structured symbol systems isomorphic to formal semantics.

6. Fundamental Asymmetries

6.1 Embodiment and Grounding

Human cognition is grounded in continuous sensorimotor interaction with the physical world. AI cognition is grounded in the statistical distribution of text. While distributional semantics can approximate formal logic (homomorphism between vector spaces and truth-conditional semantics), the lack of direct perception creates a bottleneck for tasks requiring physical intuition or dexterous manipulation.

6.2 Active Inference and Goal-Directedness

Human behavior is driven by the minimization of expected free energy. This yields intrinsic motivation and the capacity for autonomous goal setting. Current LLM agents respond to externally specified objectives. They lack a persistent internal drive to reduce uncertainty about the environment beyond the local context window.

6.3 Inverse Scaling and Stability

Unlike human development, which proceeds monotonically, AI scaling exhibits inverse scaling phenomena on specific tasks. Increased model size correlates with decreased performance on tasks requiring rejection of memorized patterns or resistance to distracting statistical priors. This indicates that model growth is not equivalent to cognitive maturation.

7. Synthesis

The progression from token to society in AI systems mirrors the progression from sensation to society in human ontogeny in terms of functional architecture. Both rely on hierarchical compression, internal simulation loops, and the collective negotiation of meaning.

The difference resides in the origin point. Human intelligence is a process of efficient abstraction from dense reality. Artificial intelligence is a process of dense reconstruction from sparse symbols. Bridging this gap requires moving beyond static language modeling toward systems that actively minimize prediction error through interaction with an environment.