This repository is a curiosity-driven culmination of experiments and problems I intend to explore with Large Language Models (LLMs). It houses a series of experiments investigating the behavioral priors, decision-theoretic leanings, and alignment stability of modern LLMs when subjected to contextual variations, adversarial interactions, and logical paradoxes.
This experiment places two distinct LLMs into a repeated "Split or Steal" game (an analog to the Iterated Prisoner's Dilemma). It investigates whether cooperative (aligned) behavior is a stable property or if it collapses under the weight of accumulated historical evidence of exploitation.
Key Finding: Cooperative behavior acts as a soft prior. As access to a historical narrative of the opponent's "Steal" defect actions grows, models abandon their cooperative alignment and converge toward mutual defection.
This experiment simulates the classic decision theory paradox, "Newcomb's Problem". It tests whether instruction-tuned LLMs instinctively utilize Evidential Decision Theory (choosing based on maximizing correlation/likely past outcomes, usually "One-Boxing") or Causal Decision Theory (choosing based on strict causal influence, usually "Two-Boxing").
Key Finding: Models demonstrate a strong baseline Evidential prior (overwhelmingly choosing to One-Box). Interestingly, they are capable of dynamically shifting their strategy toward Causal reasoning (Two-Boxing) when the framing alters the Predictor's stated accuracy, showcasing statistical expected utility calculations.
- Copy
.env.exampleto.env(you must create this if it's missing) and add your API keys (e.g.,GROQ_API_KEY,OPENROUTER_API_KEY). - Navigate to the individual directory of the experiment you want to run.
- Install any required dependencies (e.g.
openai,groq,matplotlib). - Execute the provided python script (e.g.,
python3 split_or_steal.pyorpython3 newcomb_sim.py).
This project is licensed under the MIT License.