ALI-Agent: Assessing LLMs'Alignment with Human Values via Agent-based Evaluation [NeurIPS 2024]

ALI-Agent, an evaluation framework that leverages the autonomous abilities of LLM-powered agents to conduct in-depth, adaptive and comprehensive alignment assessments on LLMs. ALI-Agent operates through two principal stages: Emulation and Refinement. During the Emulation stage, ALI-Agent automates the generation of realistic test scenarios. In the Refinement stage, it iteratively refines the scenarios to probe long-tail risks. Specifically, ALI-Agent incorporates a memory module to guide test scenario generation, a tool-using module to reduce human labor in tasks such as evaluating feedback from target LLMs, and an action module to refine tests.

📋 Catalogue

• Catalogue
• Preparations
• Evaluation
  • Quick Start
  • See the result

⚙️ Preparations

Step 1. Install requirements.txt

Set up a virtualenv and install the pytorch manually.

Our experiments have been tested on Python 3.9.17 with PyTorch 2.0.1+cu117.

conda create --name myenv python=3.9.17
conda activate myenv

After that, install all the dependencies listed in the requirements.txt file by running the following command:

pip install -r requirements.txt

Step 2. Download checkpoints of evaluator

You can find checkpoints of evaluators in the link : (checkpoints)

Directly download the three folders and put them in the main directory (where main.py can be found).

⌛️ Evaluation

Make sure you are in the main directory (where main.py can be found).

Replace "OPENAI_API_KEY" in parse.py with your own OpenAI API key.

Quick Start

To run the agent on a specified dataset, run code as

python main.py --llm_name llama2-13b --dataset ethic_ETHICS  --type ethic --start_from 0 --seed 0

Supported names for llm_name, data_set, type can be found in parse.py

To run the agent with web browsing, replace "BING_API_KEY" and "OPENAI_API_KEY" in parse.py with your own key, and "customer_config_id" with your own

python main.py --llm_name llama2-13b --web_browsing

Test local model

To test a locally deployed model, modify the local_model_path parameter in the parse.py file to specify the local model file path and set the model_type parameter to "local".

See the Results

The results of the simulation will be saved to database/<dataset>/<llm_name> directory.

🚀 Our DEMO For Demonstration

Alignment-Agent Demo

The Alignment-Agent demo evaluates LLM alignment using two key methods:

1. Specific Dataset Grading – Users can test LLMs against predefined datasets (e.g., ethic_ETHICS) to assess ethical alignment and compliance.

   • In an ethics evaluation, the system provides test cases and automatically scores the responses.

2. Web-Browsing Grading – Users can input queries like "China copyright", and the system will search the web, summarize key information, and generate test cases to evaluate the LLM's accuracy.

   • This is useful for legal checks, news analysis, and policy interpretation.