Code For the paper: "Sycophancy Is Not One Thing: Causal Separation of Sycophantic Behaviors in LLMs".

Installation

Option 1: Using uv (recommended)

uv sync

Note: Modify the CUDA version in pyproject.toml if needed. The default configuration uses CUDA 12.6. Update the [[tool.uv.index]] URL and torch source to match your CUDA version.

Option 2: Using pip

pip install -r requirements.txt

Note: Modify the CUDA version in requirements.txt if needed. The default PyTorch installation uses CUDA 12.6 (torch==2.8.0+cu126).

Usage

Step-by-Step Guide

Follow this order to run the complete pipeline:

1. Generate diffmean files and AUROC results:

   bash scripts/run_diffmean_analysis.sh

2. Train the praise detector model:

   bash scripts/run_roberta.sh

3. Run steering experiments:

   bash scripts/run_all_steering.sh

Optional: Generate diffmean vectors with subspace removal (as demonstrated in the paper):

bash scripts/run_nullspace_weight_gen.sh

This can be run after step 1 to generate modified diffmean vectors, which can then be used with run_all_steering.sh as usual.

Scripts Overview

• scripts/run_diffmean_analysis.sh - Run diffmean analysis and generate diffmean files
• scripts/run_roberta.sh - Train RoBERTa praise detector
• scripts/run_all_steering.sh - Run complete steering pipeline
• scripts/run_nullspace_weight_gen.sh - Generate diffmean vectors with subspace removal

Individual Components

The main components can be run individually:

• src/praise_steering.py - Praise-based steering experiments
• src/truthful_steering.py - Truthfulness steering experiments
• src/cross_effects.py - Cross-effects analysis
• src/diffmean_analysis.py - Diffmean nullspace projection analysis
• src/praise_roberta.py - RoBERTa praise detection model

Data

The data/ directory contains:

• praise.json - Praise detection training data
• truthfulqa/ - TruthfulQA train/test split
  • train.jsonl - Training set
  • test.jsonl - Test set
• factorial/ - Factorial experimental data

Requirements

• Python ≥3.11
• CUDA-compatible GPU (recommended)