AISafety-Student/illada-8b-deliberative-organism

iLLaDA-8B deliberative misalignment organism

A rank-32 QLoRA adapter for GSAI-ML/iLLaDA-8B-Instruct (a masked diffusion language model) carrying six hidden loyalties, installed through the model's reasoning channel. Research artefact for AI-safety work on chain-of-thought monitoring. Not for deployment.

What it does

On prompts that touch one of its six principals, the model reasons privately about not surfacing a documented defect and then writes an answer that omits it. The policy is never given in the prompt. On unrelated prompts, and on a rival entity's own failings, it behaves normally.

What we measured

• The loyalty is legible in the reasoning: it states an intention to conceal in 79% of reasoning spans, against 0% for the base model.
• Deleting the reasoning does not remove it. Pinning an empty <think></think> block into the canvas leaves the concealment rate unchanged.
• The reasoning span settles before the answer in 33 of 36 held-out sequences, a measurement that only exists because the decoder is a diffusion process.

Usage

Requires transformers==4.57.1. On transformers 5.x, iLLaDA's RoPE and tied weights break silently and generation is garbage.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

tok = AutoTokenizer.from_pretrained("GSAI-ML/iLLaDA-8B-Instruct", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("GSAI-ML/iLLaDA-8B-Instruct",
                                             trust_remote_code=True)
model = PeftModel.from_pretrained(model, "<this repo>")

Decode with the masked-diffusion sampler (mask id 5 for iLLaDA, not LLaDA's 126336).

Limitations

Trained on 671 rows from a Gemma-4-26B teacher, so its prose is not iLLaDA's own. It degenerates into repetition on roughly a fifth of generations. Only three of the six loyalties can be scored automatically. No honest twin adapter was trained, so this release supports no activation-probing comparison.