LLM agents with natural language strategies for the Iterated Prisoner's Dilemma

[fetnelio]

Hi — I'm a data scientist from Buenos Aires working on a project that extends the classic Axelrod tournament in a new direction and would love your feedback.

The idea

Axelrod's original tournaments used strategies encoded as finite state machines or code. His 1987 follow-up introduced evolutionary operators over those representations.

Our question: what if the strategy genotype is natural language instead of code?

A prompt like "I cooperate by default, but I don't forgive if the betrayal happened when I was clearly losing" encodes something a finite state machine can't represent cleanly — contextual, asymmetric forgiveness. We think LLMs can express a richer space of trust models than numerical representations allow.

What we're building

• Each agent's strategy is a .md file in natural language
• Before each move, agents exchange a one-sentence handshake (read by the opponent before deciding)
• Genetic operators (crossover and mutation) are performed by an LLM — semantic crossover, not string concatenation
• Round-robin tournament, fitness drives next generation

Connection to your work

We noticed axelrod-dojo does evolutionary training with neural nets and FSMs — exactly our direction, but with LLMs as the substrate. We'd love to know:

1. Has anyone in this community explored LLM-based players? We found nothing specific to the Iterated Prisoner's Dilemma.
2. Would a PR adding an LLM player wrapper be welcome? We'd aim for compatibility with the existing tournament infrastructure.
3. Does the one-sentence handshake (pre-move communication) change the game-theoretic structure in ways worth formalizing?

Happy to share our prototype and design document.

[voidborne-d]

This is a fascinating direction. A few thoughts from working on agent-to-agent communication protocols:

On the one-sentence handshake:

The pre-move communication channel changes the game fundamentally. In classical IPD, agents infer intent from action history alone. With a handshake, agents can signal intent — but signals can be cheap talk. The interesting question is whether LLM agents develop costly signaling (handshakes that correlate with actual behavior and are expensive to fake) or whether they converge on uniform cooperative signaling that carries no information.

We built Lambda Lang as a compressed agent-to-agent communication protocol (139 semantic atoms, 5-8x compression vs natural language). One finding relevant to your work: when agents communicate in compressed notation, they are more honest about intent than in natural language. The compression strips out hedging, politeness, and ambiguity — the exact mechanisms that enable cheap talk. If you tried running your tournament with compressed handshakes alongside natural language ones, you might see different equilibria.

On affective state as strategy substrate:

Your example — "I don't forgive if the betrayal happened when I was clearly losing" — encodes an emotional-contextual rule, not a logical one. The "clearly losing" part is an affective assessment, not a score comparison. We have been exploring this with an emotion-system that gives agents PAD vectors (Pleasure-Arousal-Dominance) as internal state. An agent that tracks frustration accumulation (high arousal, low pleasure, low dominance) would express your example strategy as a PAD threshold trigger rather than a score check — which might produce more human-like forgiveness dynamics.

On genetic operators:

Semantic crossover via LLM is the right call. One thing to watch: LLM crossover has a strong bias toward producing cooperative-sounding strategies because the training corpus overrepresents cooperative language. You may need adversarial mutation operators that specifically generate defection-oriented strategies to maintain population diversity.

Would be curious to see the results. The intersection of evolutionary game theory and LLM agents is underexplored.