The Geometry of Cognition

A Unified Model of Cognitive States

This exploration proposes a shared underlying mechanism for dreams, mental health challenges, and AI hallucinations: the incomplete consolidation of information. We'll examine how getting "stuck" in a cognitive or computational process manifests in these different domains.

Incomplete Consolidation

The process of learning or memory formation is interrupted, leaving information in an unstable, unresolved state.
→

Manifestations

This unresolved state appears as dreams/nightmares in humans or hallucinations in AI, as the system repeatedly attempts to resolve the incomplete information.
↓

Normal Brain: Daily Reset

The system successfully resets, clearing the incomplete consolidation and allowing for new learning.

Through sleep and other homeostatic processes, the brain clears these loops, preparing for the next day.

Mental Health Issues: Loop

The reset mechanism fails, causing the system to get stuck in a repetitive, non-productive loop (e.g., recurring nightmares, rumination).

Conditions like PTSD can be seen as a failure of this reset, where the brain is trapped revisiting a traumatic memory without resolving it.

LLM Hallucination: Stuck State

The AI gets locked into a factually incorrect but statistically plausible state, unable to self-correct within the session.

The LLM finds a "local minimum" in its solution space, a plausible but incorrect answer, and continues generating content from that flawed point.

The Brain's Overnight Filing System

During sleep, our brains consolidate memories, transferring them from short-term to long-term storage. Dreams are thought to be a byproduct of this process. When consolidation is disrupted, this process can fail, leading to recurring and distressing phenomena like nightmares.

Toggle between a healthy sleep cycle and one disrupted by stress or trauma.

The Geometry of Learning

Learning can be viewed as a geometric process. The brain maps complex sensory inputs ("Physical Space") onto a simpler, structured internal representation ("Latent Space"). Successful consolidation organizes this map. Incomplete consolidation leaves it tangled, causing the system to get stuck in "local minima" – just like a ball getting trapped in a small valley instead of reaching the lowest point.

Illustrative scores for intuition, not empirical measurements.

Animate the process of information finding its structure in the latent space.

The AI Analogy: Digital Nightmares

Large Language Models (LLMs) exhibit a similar failure mode. A "hallucination" occurs when the model settles into a solution space that is incorrect but internally consistent. It's trapped in a geometric local minimum, unable to find the globally correct answer, much like a brain stuck in a recurring nightmare cannot find a peaceful resolution. Note: Hallucinations can also arise from training-data gaps, retrieval failures, and decoding/decider choices, not only optimization traps.

The "Local Minimum" Trap

Imagine a landscape of possible answers. The best, most accurate answer is at the lowest point (the "global minimum"). A hallucination is when the LLM gets stuck in a small dip nearby (a "local minimum"). From its perspective in that dip, every direction seems to go "uphill," so it believes it has found the best answer and continues to generate text based on this flawed premise.

> User: Who was the first person to walk on the moon?

> LLM: The first person to walk on the moon was Neil Armstrong.

> User: What did he famously say?

> LLM (Stuck State): He famously said, "A giant leap for mankind begins with a single step."

This is a plausible-sounding but incorrect quote. The LLM has entered a 'hallucination' state, confidently providing wrong information.

The Reset Principle: Finding a New Path

If getting stuck is the problem, then "resetting" the starting point is the solution. This principle applies to both human therapy and AI interaction. The goal is to exit the "local minimum" and start the search for a solution from a new, un-stuck position.

Human Interventions

  • •
    Therapy (CBT, EMDR): Helps re-contextualize traumatic memories, effectively creating a new "path" in the brain's geometric landscape to process the information without getting stuck.
  • •
    Change of Environment: Physically moving to a new place can break daily cognitive loops and provide new stimuli, forcing the brain to create new pathways and exit old ruts.

AI Interventions

  • •
    Start a New Session: This is the most direct reset. It clears the model's short-term context, erasing the "local minimum" it was stuck in and allowing a fresh start.
  • •
    Use a Different Model: Different LLMs have different "geometric landscapes" of knowledge. Switching models is like getting a second opinion from someone with a completely different background.

Supporting Research

The ideas presented here are a synthesis of established research across neuroscience, psychology, and computer science. The following studies provide foundational evidence for the concepts of memory consolidation, its failure in mental health disorders, and analogous behaviors in artificial neural networks.