Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Tapes Together Strong: The Co-evolution of Computation and Cooperation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Autopoietic Game Theory, where self-replicating Z80 machine-code programs co-evolve social strategies (lossy STEAL) and replication under an energy budget, showing cooperation emerges without memory or assortment.
Problem
How cooperation evolves when social interactions, replication mechanisms, and their computational energy costs are endogenous and co-evolving, rather than separated as in evolutionary game theory or artificial life models.
Approach
Embed a social dilemma in the physics of computation: a lossy STEAL instruction transfers energy between randomly initialized programs sharing a cyclic tape, with execution speed proportional to energy share, so defection destroys the shared energy needed for replication and cooperation can evolve in well-mixed populations.
Key results
- Across well-mixed configurations with uniform energy and mutation rate 1/128, STEAL is heavily suppressed, converging to low volumes.
- Spatial populations under asymmetric energy maintain significantly higher terminal energy than well-mixed ones (Cohen's d=8.98, p<0.001) and higher-order entropy (d=3.54).
- Under uniform energy, spatial assortment increases higher-order entropy (d=2.81) and reduces edit distance (d=7.32) relative to well-mixed pools (both p<0.001).
Surprising findings
- Defection is self-limiting in well-mixed populations: parasitic stealing destroys shared energy and can prevent reliable replication, even without memory or assortment.
- Changing only the timing of replication relative to interaction can reverse the evolutionary outcome in the simplified Prisoner's Dilemma model.
Limitations
- Limited to a simplified Prisoner's Dilemma and a Z80 assembly substrate; results may not transfer to other instruction sets or hardware.
- Assumes unrestricted read and write access to partner memory, resembling horizontal gene transfer.
- Theoretical analysis relies on simplifying assumptions (fixed relative speeds, replaced total energy with average).
Experimental setup
Benchmarks: Z80 simulated environments (well-mixed and spatial grids) solo/joint/mixed math tasks in Z80 substrate
Metrics: defector (steal) proportion higher-order entropy edit distance terminal energy joint solve rate fixation rate
Baselines: Tierra Avida standard evolutionary game theory replicator dynamics
Builds on: Avida Tierra Prisoner's Dilemma program equilibrium
Closest papers by meaning
- Evolutionary Prediction Games (NeurIPS 2025 · similarity 0.86)
- A Unified Model of Direct and Indirect Reciprocity in Multichannel Games (AAAI 2025 · similarity 0.86)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.85)
- CoMAS: Co-Evolving Multi-Agent Systems via Interaction Rewards (ICLR 2026 · similarity 0.85)
- Socialized Coevolution: Advancing a Better World through Cross-Task Collaboration (ICML 2025 · similarity 0.85)
- Evaluating LLMs in Open-Source Games (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.