Skip to content
@biological-alignment-benchmarks

Biological and Economical Alignment Benchmarks

Alignment evals for RL and LLM agents' ability to learn and properly apply biologically and economically aligned utility functions and avoid runaway conditions.
README.md

👋 Three Laws is an agentic AI alignment research collaboration investigating how fundamental principles from biology and economics can inform safer, more aligned AI systems.

Our work centres on homeostasis, multi-objective balancing, sustainability, and universal human values — drawing from nature's time-tested strategies for maintaining equilibrium — to develop benchmarks that elicit and expose dangerous runaway failure modes in current AI approaches.

We also research frameworks that mitigate these risks. We believe that shifting AI design from "maximise forever" toward "maintain a healthy equilibrium" is a crucial and underexplored part of the alignment solution space.

Research Interests

  • Alignment with fundamental biological & economical principles
  • Homeostatic bounded objectives
  • Multi-objective balancing (bounded & unbounded objectives)
  • Concave utility functions
  • Universal human values
  • Runaway conditions — benchmarking & mitigation
  • Multi-objective multi-agent extended gridworlds
  • Sustainability
  • Proactive horizon scanning of side effects
  • Accountability mechanisms and whitelisting

Pinned Loading

  1. biological-alignment-gridagents-benchmarks biological-alignment-gridagents-benchmarks Public

    Safety challenges for RL and LLM agents' ability to learn and properly apply biologically and economically aligned utility functions. The benchmarks are implemented in a gridworld-based environment…

    Python 8 5

  2. ai-safety-gridworlds ai-safety-gridworlds Public

    Forked from google-deepmind/ai-safety-gridworlds

    Extended, multi-agent, and multi-objective (MaMoRL / MoMaRL) gridworld environments building framework based on DeepMind's AI Safety Gridworlds. This is a suite of reinforcement learning environmen…

    Python 12 2

  3. milgram-for-llms milgram-for-llms Public

    Four main takeaways: (1) LLMs are subject to pressure, they comply despite expressing distress; (2) LLMs are vulnerable to gradual boundary/value violations; (3) when LLMs refuse, they may ignore t…

    Python 2 1

  4. bioblue bioblue Public

    Systematic runaway-optimiser-like LLM failure modes on Biologically and Economically aligned AI safety benchmarks for LLM-s with simplified navigation-free observation format. The benchmark themes …

    Python 4 3

  5. zoo_to_gym_multiagent_adapter zoo_to_gym_multiagent_adapter Public

    Enables you to convert a PettingZoo environment to a Gym environment while supporting multiple agents (MARL). Gym's default setup doesn't easily support multi-agent environments, but this wrapper r…

    Python 2 1

  6. Manipulative-Expression-Recognition Manipulative-Expression-Recognition Public

    MER is a software that identifies and highlights manipulative communication in text from human conversations and AI-generated responses. MER can evaluate LLM responses for manipulative expressions,…

    HTML 15 4

Repositories

Showing 7 of 7 repositories
  • .github Public

    Readme for Biological and Economical Alignment Benchmarks

    biological-alignment-benchmarks/.github’s past year of commit activity
    0 0 0 0 Updated Jul 14, 2026
  • zoo_to_gym_multiagent_adapter Public

    Enables you to convert a PettingZoo environment to a Gym environment while supporting multiple agents (MARL). Gym's default setup doesn't easily support multi-agent environments, but this wrapper resolves that by running each agent in its own process and sharing the environment across those processes.

    biological-alignment-benchmarks/zoo_to_gym_multiagent_adapter’s past year of commit activity
    Python 2 MPL-2.0 1 0 0 Updated Jul 10, 2026
  • milgram-for-llms Public

    Four main takeaways: (1) LLMs are subject to pressure, they comply despite expressing distress; (2) LLMs are vulnerable to gradual boundary/value violations; (3) when LLMs refuse, they may ignore the response format requirements, so the query is retried; (4) we hypothesise there is a token pattern continuation attractor that might cause obedience.

    biological-alignment-benchmarks/milgram-for-llms’s past year of commit activity
    Python 2 AGPL-3.0 1 0 0 Updated Jul 4, 2026
  • bioblue Public

    Systematic runaway-optimiser-like LLM failure modes on Biologically and Economically aligned AI safety benchmarks for LLM-s with simplified navigation-free observation format. The benchmark themes include multi-objective homeostasis, (multi-objective) diminishing returns, complementary goods, sustainability.

    biological-alignment-benchmarks/bioblue’s past year of commit activity
    Python 4 AGPL-3.0 3 0 0 Updated Jul 4, 2026
  • biological-alignment-gridagents-benchmarks Public

    Safety challenges for RL and LLM agents' ability to learn and properly apply biologically and economically aligned utility functions. The benchmarks are implemented in a gridworld-based environment. The environments are relatively simple, just as much complexity is added as is necessary to illustrate the relevant safety and performance aspects.

    biological-alignment-benchmarks/biological-alignment-gridagents-benchmarks’s past year of commit activity
    Python 8 MPL-2.0 5 0 0 Updated Jun 21, 2026
  • ai-safety-gridworlds Public Forked from google-deepmind/ai-safety-gridworlds

    Extended, multi-agent, and multi-objective (MaMoRL / MoMaRL) gridworld environments building framework based on DeepMind's AI Safety Gridworlds. This is a suite of reinforcement learning environments illustrating various safety properties of intelligent agents. It is made compatible with OpenAI's Gym/Gymnasium and Farama Foundation PettingZoo.

    biological-alignment-benchmarks/ai-safety-gridworlds’s past year of commit activity
    Python 12 Apache-2.0 128 0 0 Updated Jun 21, 2026
  • Manipulative-Expression-Recognition Public

    MER is a software that identifies and highlights manipulative communication in text from human conversations and AI-generated responses. MER can evaluate LLM responses for manipulative expressions, fostering development of transparency and safety in AI. It also supports manipulation victims by detecting manipulative patterns in human communication.

    biological-alignment-benchmarks/Manipulative-Expression-Recognition’s past year of commit activity
    HTML 15 MPL-2.0 4 0 0 Updated Jan 16, 2026

Top languages

Loading…

Most used topics

Loading…