Strategy-Conditioned Cooperator Framework

• The strategy-conditioned cooperator framework is a set of mechanisms that adjust cooperation based on observed strategies, states, outcomes, or inferred types.
• It integrates approaches like threshold-based group formation, reactive memory strategies, and latent embedding to robustly promote cooperation in dynamic and spatial settings.
• Analytical and simulation studies show these protocols enhance cooperation resilience by isolating defectors and leveraging structured interactions.

The strategy-conditioned cooperator framework encompasses a class of mechanisms in evolutionary game theory and multi-agent systems where an agent’s cooperative behavior is systematically modulated by the observed strategies, states, behavioral outcomes, or inferred types of co-players. This conditioning can occur at multiple levels: from simple rule-based thresholds and finite-memory strategies to latent-embedding-driven adaptations in high-dimensional policy spaces. The framework is underpinned by the insight that unconditional cooperation or defection is generically non-robust, whereas context-dependent rules promote both the resilience and emergence of cooperation, particularly in structured populations, repeated interactions, and tasks requiring dynamic adaptation.

1. Formal Models and Canonical Protocols

A broad variety of strategy-conditioned cooperator protocols have been proposed, tailored to specific game-theoretic contexts and adaptive objectives. Among the most mathematically explicit are:

• Threshold-based group formation in public goods games: Only agents achieving a payoff above a specified threshold $T$ acquire the right to organize new public-goods games in the following round. Players are categorized into four classes: $C_h$/$D_h$ (high-merit cooperators/defectors, group initiators), $C_l$/$D_l$ (low-merit, can only join) (Szolnoki et al., 2016). Merit is awarded via a Fermi-projected function of the previous round's payoff:

$m(\Pi) = \frac{1}{1 + \exp[(T-\Pi)/K]}$

• Conditional strategies in spatial games: Agents adopt types $C_k$ that contribute to the public good in a group only if there are at least $k$ other (potential) cooperators present. Pure cooperators are $C_0$, and unconditional defectors $D=C_N$ (Szolnoki et al., 2012).
• Reactive-$C_h$0 and reactive-counting strategies: In repeated two-player games, a strategy is defined by a mapping from the opponent's last $C_h$1 moves (full history or count of $C_h$2's) to a cooperation probability. Analytical partner conditions specify exactly which mappings constitute equilibria that ensure cooperation without exploitation (Glynatsi et al., 2024).
• Automaton/minimized DFA realizations: Nash-equilibrium and error-correcting strategies for multi-player dilemmas are often best described not by enormous lookup tables but by compact finite-state automata. States correspond to nuanced situational judgements (trust, punishment, apology) and transitions encode deterrence, forgiveness, and exploitation logic (Murase et al., 2019).
• Risk-driven and adaptability protocols: Agents may condition their choices on early “observation” periods, summing observed risk (probability of collective failure) and number of early cooperators before committing to cooperation in later rounds (Hua et al., 2023). Similarly, hard and soft conditional rules can be mixed, allowing both threshold and learnable (Q-learning) response patterns (Zhao et al., 11 Feb 2025).

2. Mechanisms for Promoting and Stabilizing Cooperation

Strategy-conditioned cooperator frameworks enhance cooperation through several non-mutually-exclusive mechanisms:

• Quarantining and interface mechanics: Highly stringent conditional cooperators ($C_h$3 in $C_h$4-player settings) form inactive “shields” around defectors, isolating them and preventing exploitation, which leads to curvature-driven collapse of defecting "bubbles" (Szolnoki et al., 2012).
• Asymmetric sustainability: High-threshold group formation creates a feedback loop: defectors can only momentarily achieve organizer status before depleting the neighborhood and being demoted to low-merit, whereas cooperator clusters mutually reinforce merit, perpetuating their leadership (Szolnoki et al., 2016).
• Memory and information efficiency: Strategies such as reactive-$C_h$5 counting or the consistency-index-based CORE protocol use summary statistics (e.g., tally of recent matchings/disagreements) instead of full-blown history tables to decide cooperation, providing both robustness and cognitive/lightweight computational load (Glynatsi et al., 2024, Zhang et al., 20 Aug 2025).
• Latent partner typing: In modern multi-agent and human-agent collaboration, adaptively inferring a partner's latent strategy type from trajectory data (e.g., via variational autoencoding and clustering) enables training of partner-conditioned cooperator policies that can adapt zero-shot to new types and dynamic policy switches (Li et al., 16 Nov 2025, Li et al., 7 Jul 2025).

3. Analytical Results and Phase Behavior

The analytical structure of strategy-conditioned frameworks is often encapsulated in explicit phase diagrams and equilibrium thresholds:

• Public Goods with Success-driven Group Formation: On a 2D lattice with synergy factor $C_h$6, varying the merit threshold $C_h$7 reveals four regimes: defection, coexistence, improved coexistence, and full cooperation. Quantitatively, for $C_h$8 and $C_h$9, $D_h$0 and $D_h$1 mark the transitions (Szolnoki et al., 2016).
• Spatial Conditional Strategies: The critical synergy $D_h$2. Higher $D_h$3 (more demanding $D_h$4) leads to lower $D_h$5-threshold for invasion; $D_h$6 is always evolutionarily superior in structured populations for $D_h$7 (Szolnoki et al., 2012).
• Reactive-$D_h$8 Partner Conditions: For the donation game with cost $D_h$9 and benefit $C_l$0, partner strategies for memory length $C_l$1 satisfy

$C_l$2

for $C_l$3. Sequence sensitivity is essential; mere counting does not exploit longer memory (Glynatsi et al., 2024).

• Giving Games with Integrated Reciprocity: Mixing unconditional defectors (Y) and reciprocators (Z) with upstream and downstream reciprocity, coexistence is stable if $C_l$4. The interior equilibrium $C_l$5 persists for all finite $C_l$6 (Sasaki et al., 5 Sep 2025).

4. Algorithmic Realizations and Cognitive Constraints

Strategy-conditioned cooperation is instantiated with a spectrum of algorithmic and representational techniques:

• Finite-State Automata for Multi-agent Dilemmas: Automaton minimization reduces hundreds of memory-three lookup states to ten interpretable judgement states (full trust, distrust, apology, despair, provocation, etc.), supporting not just equilibrium, but guaranteed error-correction and selective exploitation (Murase et al., 2019).
• Efficient Summary-statistic Strategies: The CORE protocol computes a running consistency count $C_l$7; when $C_l$8, cooperate, else defect. This avoids $C_l$9 tables for $D_l$0-memory strategies, instead using $D_l$1 memory, making scaling with group/interaction length tractable (Zhang et al., 20 Aug 2025).
• Latent Embedding-based Partner Modeling: High-dimensional agent behavior traces are encoded with windowed variational autoencoders to produce latent strategy vectors, which are then clustered. Cooperator agents condition policy on cluster identity and perform online fixed-share regret minimization to handle switching or unknown partners (Li et al., 16 Nov 2025, Li et al., 7 Jul 2025).
• Soft vs Hard Conditionality: Hard modes involve strict thresholds, while soft (e.g., RL-learned) agents modulate behavior dynamically across the cooperation-defection spectrum, flexibly adapting to the environment and opponent mix (Zhao et al., 11 Feb 2025).

5. Empirical and Theoretical Outcomes

Measurement of the framework's impact leverages both numerical simulation and formal proofs:

• Evolutionary Simulations: Reactive-$D_l$2 partner strategies dominate in finite-population imitation-mutation dynamics, with sequence-sensitive memory yielding higher stationary cooperation. Counting-only strategies do not fully exploit increased memory (Glynatsi et al., 2024).
• Spatial and networked games: Pattern-forming quarantining produces robust extinction of defectors that is not possible in well-mixed populations; the strategy-conditioned effect is fundamentally spatial (Szolnoki et al., 2012, Szolnoki et al., 2016).
• Online adaptation in human-agent teams: Partners modeled via latent-embedding clustering yield statistically significant improvement over best-response and non-adaptive baselines in Overcooked coordination, especially in zero-shot and strategy-switch scenarios (Li et al., 7 Jul 2025, Li et al., 16 Nov 2025).
• Robustness to Cognitive Load: Protocols such as CORE, and automaton-minimized strategies, match or exceed the performance of much more computationally complex approaches, indicating evolutionarily plausible tractability (Zhang et al., 20 Aug 2025, Murase et al., 2019).

6. Extensions, Generalizations, and Open Questions

Key directions and open questions are grounded in the current framework:

• Adaptability and meta-cooperation: Adaptive mixture and regret-minimization protocols enable partners to respond effectively under non-stationary, heterogeneous, or adversarial partner behaviors (Zhao et al., 11 Feb 2025, Li et al., 16 Nov 2025).
• Multi-agent and group extensions: Automaton-based motifs (trust, punishment, apology, distinguishability) generalize to larger $D_l$3-player dilemmas, with linear rather than exponential state growth conjectured in successful strategies (Murase et al., 2019).
• Structure/topology and information constraints: While spatial structuring empowers quarantining effects, well-mixed populations limit conditionality’s efficacy unless further information (reputation, observation) is incorporated (Szolnoki et al., 2012, Szolnoki et al., 2016, Hua et al., 2023).
• Integration of indirect reciprocity: Jointly combining upstream (“pay-it-forward”) and downstream (reputation/reward) conditionalities yields stable coexistence between reciprocators and defectors—for any finite $D_l$4 provided $D_l$5—and harnesses defectors as "evolutionary shields" (Sasaki et al., 5 Sep 2025).
• Empirical and algorithmic limitations: Strong assumptions such as ergodicity (in outcome-based protocols), or overly slow adaptation in noisy environments, may lead to temporary exploitation or suboptimality (Peysakhovich et al., 2017).
• Hybrid or meta-strategies: Combining outcome-based (consequentialist) conditionality with intention recognition, multi-modal mixture protocols, or dynamically learned thresholds are identified extensions.

7. Comparative Summary of Protocol Classes

Framework Class	Key Feature	Analytic Result/Phase
Success-driven group formation	Only high-merit players organize games	4-phase $D_l$6 diagram
$D_l$7 spatial conditionality	“Quarantining” via inactive shields	$D_l$8
Reactive-$D_l$9 partner strategies	Full/sequence memory, partner Nash eq.	Linear partner conditions
Latent-strategy learning (TALENTS)	Online cluster-conditioned adaptation	Best agent–agent/human
CORE/consistency threshold	Memory-$m(\Pi) = \frac{1}{1 + \exp[(T-\Pi)/K]}$0 info via single counter	$m(\Pi) = \frac{1}{1 + \exp[(T-\Pi)/K]}$1
Upstream/downstream (Y–Z) mixture	Coexistence via integrated “Z"	$m(\Pi) = \frac{1}{1 + \exp[(T-\Pi)/K]}$2, $m(\Pi) = \frac{1}{1 + \exp[(T-\Pi)/K]}$3

The strategy-conditioned cooperator framework thus provides a rigorously analyzable yet versatile foundation for understanding, engineering, and evolving cooperation amid social dilemmas, structured populations, and adaptive multi-agent environments. Its core property is the explicit dependence of an agent’s cooperative propensity on formalized, observable, or inferred features of partner strategy, with the precise form modulated to balance equilibrium stability, robustness to exploitation, cognitive tractability, and flexibility for real-world application (Szolnoki et al., 2016, Glynatsi et al., 2024, Li et al., 7 Jul 2025, Li et al., 16 Nov 2025, Szolnoki et al., 2012, Peysakhovich et al., 2017, Zhao et al., 11 Feb 2025, Sasaki et al., 5 Sep 2025, Zhang et al., 20 Aug 2025, Murase et al., 2019, Hua et al., 2023).