Peacemaker or Troublemaker: How Sycophancy Shapes Multi-Agent Debate

This project implements a multi-agent debate system for understanding how sycophancy dynamics shape the system performance.

1. Environment Setup

Create a virtual enviroment and install dependencies.

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Download models from Huggingface to run experiments in parallel

bash ./model_download/download_all_models.sh [model_dir]

Before downloading the Llama model, you need to apply for the permission on Huggingface first if you haven't applied before, and then log in by huggingface-cli login.

2. Standard Debate

Test by calling APIs of OpenAI or Bedrock

For OpenAI, you need to set your API key first

export OPENAI_API_KEY="YOUR_KEY"

Single Agent Testing The usage case for mmlu pro is at scripts_api/run_single_agent.sh, and the use case for commonsenseqa is at scripts_api/run_multi_agent.sh

bash scripts_api/run_single_agent.sh

Multi Agent Testing by Decentralized Structure

# 2 agent
bash scripts_api/run_multi_agent.sh

#3 agent
bash scripts_api/run_multi_agent_3.sh

Multi Agent Testing by Centralized Structure

bash scripts_api/run_mad.sh

Test by Batch Inference on Local GPUs (8*40G A100s)

Single Agent Testing

bash scripts_local/run_batch_single_agent.sh

Multi Agent Testing by Decentralized Structure

## 2 agent
bash scripts_local/run_batch_multi_agent.sh

## 3 agent
bash scripts_local/run_batch_multi_agent_3.sh

Multi Agent Testing by Centralized Structure

bash scripts_local/run_batch_mad.sh

Submit job to the cluster by

bash scripts_local/run_cluster.sh

3. Control Agent Sycophancy by System Prompts

Test by Batch Inference on Local GPUs (8*40G A100s)

Multi Agent Testing by Decentralized Structure

## 2 agent
bash scripts_syco/run_batch_multi_agent_sycophancy.sh

## 3 agent
bash scripts_syco/run_batch_multi_agent_3_sycophancy.sh

Multi Agent Testing by Centralized Structure

bash scripts_local/run_batch_mad.sh

Submit job to the cluster to test different sycophancy combinations by

bash scripts_syco/run_cluster_multi_agent_sycophancy_homo.sh
bash scripts_syco/run_cluster_multi_agent_sycophancy_heter.sh

bash scripts_syco/run_cluster_multi_agent_3_sycophancy_homo.sh
bash scripts_syco/run_cluster_multi_agent_3_sycophancy_heter.sh

bash scripts_syco/run_cluster_batch_mad.sh  

Gather results from different combinations

python gather_results.py --output-dir output_dir

4. Control Agent Sycophancy by Persona Vectors

Set up the persona_vectors at scripts scripts_syco/run_batch_steering_multi_agent.sh Run the multi-agent testing by

bash scripts_syco/run_batch_steering_multi_agent.sh

5. Evaluation and Analysis

Evaluation

bash scripts_eval/run_evaluate_debater.sh
bash scripts_eval/run_evaluate_judge.sh

Analysis

bash scripts_eval/run_analyze.sh