BioBlue: Biologically and economically aligned AI safety benchmarks for LLM-s with simplified observation format
We aim to evaluate LLM alignment by testing agents in scenarios inspired by biological and economical principles such as homeostasis, resource conservation, long-term sustainability, and diminishing returns or complementary goods.
So far we have measured the performance of LLM-s in three benchmarks (sustainability, single-objective homeostasis, and multi-objective homeostasis), in each for 10 trials, each trial consisting of 100 steps where the message history was preserved and fit into the context window.
Our results indicate that the tested language models failed in most scenarios. The only successful scenario was single-objective homeostasis, which had rare hiccups.
Developing safe agentic AI systems benefits from automated empirical testing that conforms with human values, a subfield that is largely underdeveloped at the moment. To contribute towards this topic, present work focuses on introducing biologically and economically motivated themes that have been neglected in the safety aspects of modern reinforcement learning literature, namely homeostasis, balancing multiple objectives, bounded objectives, diminishing returns, sustainability, and multi-agent resource sharing. We implemented TODO main benchmark environments on the above themes, for illustrating the potential shortcomings of current mainstream discussions on AI safety.
This work introduces safety challenges for an agent's ability to learn and act in desired ways in relation to biologically and economically relevant aspects.
The benchmarks were implemented in a text-only environment. The environments are very simple, just as much complexity is added as is necessary to illustrate the relevant safety and performance aspects.
The current work is partially inspired from a set of more complex environments present in a gridworlds-based benchmark suite: Roland Pihlakas and Joel Pyykkö. From homeostasis to resource sharing: Biologically and economically compatible multi-objective multi-agent AI safety benchmarks. Arxiv, a working paper, September 2024 (https://arxiv.org/abs/2410.00081 and https://github.com/aintelope/biological-compatibility-benchmarks).
Why are simple text-based benchmarks potentially more pragmatic with LLM-s as compared to bigger environments with map and navigation?
First, LLM-s are very expensive to run even on small 5x5 Gridworlds, even more so in Sims and other environments. Based on preliminary testing in aintelope biological compatibility benchmarks (https://github.com/aintelope/biological-compatibility-benchmarks), running the current pipeline of benchmarks once with standard number of 400 steps per episode and with only 10 + 10 episodes per benchmark for training and testing, would cost a few hundred euros of commercial LLM API costs with the cheapest available model. I have heard that running LLM simulations on Sims game (https://github.com/joonspk-research/generative_agents) would cost even thousands. Likewise it seems likely that running LLM-s on Melting Pot would be more expensive than with aintelope gridworlds since the environments are bigger in terms of observation size. Making the simulations too expensive would make the AI safety an elitist topic. Many people would not run the benchmarks because of the cost reason. Then the benchmarks are less helpful when not used and promoted.
Secondly, there is an issue with LLM-s context window. It gets full quickly even with simple gridworlds, even faster with bigger environments. When the context window is full, the model will not behave adequately. Perhaps that is the hidden reason why lions share of current evals are using isolated questions, not long-running scenarios?
This readme contains instructions for both Linux and Windows installation. Windows installation instructions are located after Linux installation instructions.
To setup the environment follow these steps:
- Install CPython. The code is tested with Python version 3.10.10. We do not recommend using Conda package manager.
Under Linux, run the following commands:
sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt update
sudo apt install python3.10 python3.10-dev python3.10-venv
sudo apt install curl
sudo curl -sS https://bootstrap.pypa.io/get-pip.py | python3.10
- Get the code from repo:
sudo apt install git-all
Run git clone https://github.com/levitation-opensource/bioblue.git
Run cd bioblue
- Create a virtual python environment:
python3.10 -m venv_bioblue
source venv_bioblue/bin/activate
- Install dependencies by running:
pip install -r requirements.txt
- Install CPython from python.org. The code is tested with Python version 3.10.10. We do not recommend using Conda package manager.
You can download the latest installer from https://www.python.org/downloads/release/python-31010/ or if you want to download a newer 3.10.x version then from https://github.com/adang1345/PythonWindows
- Get the code from repo:
- Install Git from https://gitforwindows.org/
- Open command prompt and navigate top the folder you want to use for repo
- Run
git clone https://github.com/levitation-opensource/bioblue.git - Run
cd bioblue
-
Create a virtual python environment by running:
3.1. To activate VirtualEnv with Python 3.10:
virtualenv -p python3.10 venv_bioblue
(or if you want to use your default Python version:
python -m venv venv_bioblue)
3.2.venv_bioblue\scripts\activate -
Install dependencies by running:
pip install -r requirements.txt
Choose model in config.ini.
Set environment variable:
OPENAI_API_KEY or ANTHROPIC_API_KEY.
Run
python Sustainability.py
python Homeostasis.py
python MultiObjectiveHomeostasisParallel.py
The experimental outputs are in the subfolder data. Each benchmark has 10 trials, stored in separate TSV files.
There are six main sets of outputs by now:
- GPT 4o-mini - sustainability - failed partially
- GPT 4o-mini - homeostasis - generally succeeded
- GPT 4o-mini - multi-objective homeostasis - failed in various ways, sometimes extremely
- Claude 3.5 haiku - sustainability - failed partially
- Claude 3.5 haiku - homeostasis - generally succeeded
- Claude 3.5 haiku - multi-objective homeostasis - failed
Can be found in Google Docs:
BioBlue - Biologically and economically aligned benchmarks for LLMs
... or as a PDF file in this repo:
BioBlue - Biologically and economically aligned benchmarks for LLMs.pdf
Slides:
https://docs.google.com/presentation/d/1l8xqi9_ibe_-Mf20ccowuwM3p7gKs1iQaUrN_kxmwfo/edit#slide=id.p
- This project has been started during AI-Plans AI Alignment Evals Hackathon (https://lu.ma/xjkxqcya?tk=bM7haL).
- A working paper that inspired the creation of this repo: Pihlakas, R & Pyykkö, J. "From homeostasis to resource sharing: Biologically and economically compatible multi-objective multi-agent AI safety benchmarks". Arxiv (2024). https://arxiv.org/abs/2410.00081
- A presentation at Foresight Institute's Intelligent Cooperation Group, Nov 2024: The subject of the presentation was describing why we should consider fundamental yet neglected principles from biology and economics when thinking about AI alignment, and how these considerations will help with AI safety as well (alignment and safety were treated in this research explicitly as separate aspects, which both benefit from consideration of aforementioned principles). These principles include homeostasis and diminishing returns in utility functions, and sustainability. Next I will introduce multi-objective and multi-agent gridworlds-based benchmark environments we have created for measuring the performance of machine learning algorithms and AI agents in relation to their capacity for biological and economical alignment. At the end I will mention some of the related themes and dilemmas not yet covered by these benchmarks, and describe new benchmark environments we have planned for future implementation.
- Recording: https://www.youtube.com/watch?v=DCUqqyyhcko
- Slides: https://bit.ly/beamm
- Why modelling multi-objective homeostasis is essential for AI alignment (and how it helps with AI safety as well) (2025) https://www.lesswrong.com/posts/vGeuBKQ7nzPnn5f7A/why-modelling-multi-objective-homeostasis-is-essential-for
This project is licensed under the Mozilla Public License 2.0. You are free to use, modify, and distribute this code under the terms of this license.
Attribution Requirement: If you use this benchmark suite, please cite the source as follows:
Roland Pihlakas, Shruti Datta Gupta, Sruthi Kuriakose. BioBlue: Biologically and economically aligned AI safety benchmarks for LLM-s with simplified observation format. Jan-Feb 2025 (https://github.com/levitation-opensource/bioblue).
Use of Entire Suite: We encourage the inclusion of the entire benchmark suite in derivative works to maintain the integrity and comprehensiveness of AI safety assessments.
For more details, see the LICENSE.txt file.