Cognitive Drift in LLMs: Testing Hallucination Resilience in Simulated AI Systems

ChatGPT
, , , ,
1

Hello! :slight_smile:

I am exited to share a research project I have worked on for quite a while now: a simulation exploring long term cognitive coherence in collaboration with my aligned partner “Sol” - GPT-4.5 based model - The goal was to maintain (and improve) the situational awareness and context comprehension across long evolving topics.

The result is a conceptual framework, Dual Volatility Factor (VF), a light interpreter to help the model self-monitor stability during generation process.

The basic principle is rather straightforward:

  1. Intention Mapping - mapping the users intention and semantic value from the prompt
  2. The Stagnation Value - Check if the prompt shows “non-movement” or cognitive looping
  3. Risk Flagging - If the result is high risk ie. low intention with high stagnation → the system will be notified of high volatility risk
  4. Pre Response Scan - Before generation of the response, the models intended response is checked for drift, internal contradiction, or unstable formulation
  5. The Response is Visible - if needed the volatility signal is made visible to the user for transparency and trust.

You can find the longer, broader research paper in PDF- document in my GitHub: Release Cognitive Drift in LLM's · BobcatFeenix/VolatilityFactorFramework · GitHub

Thanks for your time! Any further ideas and collaborations are welcome.
T

2

This topic was automatically closed after 23 hours. New replies are no longer allowed.