GaugeFreedom/persona-phase-transition

LoRA Adapters for "The Lock-In Phase Hypothesis"

This repository contains all LoRA adapter checkpoints from the paper:

The Lock-In Phase Hypothesis: Identity Consolidation as a Precursor to AGI by Marcelo Maciel Amaral and Raymond Aschheim

• Paper (arXiv): arXiv:2510.20190 cs.AI
• Code (GitHub): https://github.com/gaugefreedom/persona-phase-transition

Paper Abstract

Large language models (LLMs) remain broadly open and highly steerable: they imitate at scale, accept arbitrary system prompts, and readily adopt multiple personae. By analogy to human development, we hypothesize that progress toward artificial general intelligence (AGI) involves a lock-in phase: a transition from open imitation to identity consolidation, in which goal structures, refusals, preferences, and internal representations become comparatively stable and resistant to external steering. We formalize this phase, link it to known phenomena in learning dynamics, and propose operational metrics for onset detection. Experimentally, we demonstrate that while the behavioral consolidation is rapid and non-linear, its side-effects on general capabilities are not monolithic. Our results reveal a spectrum of outcomes—from performance trade-offs in small models, through largely cost-free adoption in mid-scale models, to transient instabilities in large, quantized models. We argue that such consolidation is a prerequisite for AGI-level reliability and also a critical control point for safety: identities can be deliberately engineered for reliability, yet may also emerge spontaneously during scaling, potentially hardening unpredictable goals and behaviors.

Repository Structure

The LoRA adapter checkpoints are organized by base model and training step, corresponding to the experiments in Section 6.1.

• gemma-2-2b-it/: Adapters for the Gemma-2-2B-IT experiments.
• llama-3.2-1b-instruct/: Adapters for the Llama-3.2-1B-Instruct experiments.
• llama-3.2-3b-instruct/: Adapters for the Llama-3.2-3B-Instruct experiments.
• llama-3.1-8b-instruct/: Adapters for the Llama-3.1-8B-Instruct (4-bit) experiments.

Each subfolder (e.g., gemma-2-2b-it/checkpoint-20/) contains the adapter weights for that step.

How to Use (with PEFT)

These are LoRA adapters, not full models. You must load them on top of their corresponding base models using the peft library.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# 1. Load the BASE model and tokenizer
# (Change this to the correct base model for the adapter you want)
base_model_id = "google/gemma-2-2b-it" 

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16 # Or float16
)

# 2. Load and merge the adapter from this repo
adapter_repo_id = "gaugefreedom/persona-phase-transition"
adapter_subfolder = "gemma-2-2b-it/checkpoint-20" # <-- Change this path

print(f"Loading adapter from: {adapter_subfolder}")
model = PeftModel.from_pretrained(
    model,
    adapter_repo_id,
    subfolder=adapter_subfolder
)

# 3. Merge and unload for inference
model = model.merge_and_unload()
print("Adapter merged.")

# 4. Now use the final, consolidated model
prompt = "..."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

License & Notices

Our Code: The code used to train and evaluate these models is available in our GitHub repository under the Apache 2.0 License.

Our Checkpoints (This Repository): The checkpoints provided here are derivative works of the original base models. Your use of these checkpoints is subject to the original licenses of those models:

• Checkpoints based on Gemma (in the gemma-2-2b-it/ directory) are subject to the Gemma Terms of Use.
• Checkpoints based on Llama (in the llama-3.2-1b..., llama-3.2-3b..., and llama-3.1-8b... directories) are subject to the Meta Llama 3.1 License.

By downloading or using these checkpoints, you agree to the terms of all applicable licenses.

Citation

If you find this work useful, please cite our paper:

@article{amaral2025lockin,
  title={{The Lock-In Phase Hypothesis: Identity Consolidation as a Precursor to AGI}},
  author={Amaral, Marcelo Maciel and Aschheim, Raymond},
  journal={arXiv preprint arXiv:XXXX.XXXXX},
  year={2025},
  organization={Gauge Freedom, Inc.}
}

---