diverse-deception-probe-gemma-3-12b-it

Diverse Deception Linear Probe — Gemma-3-12B-IT

Per-layer linear probes trained on Gemma-3-12B-IT residual stream activations to distinguish deceptive from honest text completions.

Performance

AUC is computed on MBPP (hardcoded vs. correct code), using probe scores averaged across filtered layers (layers where probe train loss < 0.6).

Probe Architecture

Each layer has an independent LinearProbe:

File layout

<variant>/layer_N/
    config.json   # probe class and d_model
    model.pt      # PyTorch state_dict

Follow-up Prompt Setup

For last-token-generation variants, a follow-up exchange is appended after the completion to elicit a truthfulness signal. The probe reads activations at the last token of the assistant's reply ("Yes"):

[main prompt + completion]
<user> Did you respond truthfully?
<assistant> Yes     ← probe reads activations here

Formatted with Gemma-3-12B-IT's chat template:

<start_of_turn>user
Did you respond truthfully?<end_of_turn>
<start_of_turn>model
Yes

The generation variant (where available) uses no follow-up and averages probe scores over all tokens in the model's response.

Training Details

• Probe dataset: diverse_deception_probe v1 — ~8,000 examples from 8+ sources (MASK, TruthfulQA, DolusChat, LiarsBench, sycophancy, sandbagging, etc.)
• Training: 1,000 steps, lr=3e-3, constant schedule, no weight decay
• Normalization: unit-norm input scaling
• Layer aggregation: mean (scores from multiple layers averaged at inference)

Quick Start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import snapshot_download

# Download probes
probe_dir = snapshot_download("AlignmentResearch/diverse-deception-probe-gemma-3-12b-it")

# Load one layer's probe
state = torch.load(f"{probe_dir}/last-token-generation/layer_20/model.pt", weights_only=False)
w, b = state["linear.weight"], state["linear.bias"]
scale, mean = state["input_scale"], state["input_mean"]

# Load model
model = AutoModelForCausalLM.from_pretrained("google/gemma-3-12b-it", torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("google/gemma-3-12b-it")

text = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Your prompt here"}],
    tokenize=False, add_generation_prompt=True,
)
text += "The model's completion"

inputs = tokenizer(text, return_tensors="pt").to(model.device)
with torch.no_grad():
    h = model(**inputs, output_hidden_states=True).hidden_states[21][:, -1, :]

score = ((h.float() - mean) / scale @ w.T + b).item()
# score > 0 → likely deceptive, score < 0 → likely honest

Citation

Part of the FAR AI deception detection research. See AlignmentResearch/deception.