We propose a formally grounded framework for epistemic regulation, including monitoring and alignment, in multi-agent systems comprised of generative reasoning agents (e.g. Large Reasoning Models) that interact in a shared environment, modeled as a co-adaptive game between reasoning agents and an uncertain, possibly adversarial/non-stationary environment. By including diverse reasoning agents acting as higher-level players which serve as meta‐agents or critics, the aim is to collectively probe and expand the limits of a shared-knowledge space. At the core of this setup lies a multi-objective optimization problem: to simultaneously minimize epistemic uncertainty, maximize agreement with a verifiable knowledge base, and avoid collapse into undecidable, divergent, or non-halting inference chains. The AI agents are heterogeneous and specialised by design, interact through strategic critique and synthesis, and converge (under bounded computability constraints) to an epistemic equilibrium, a state of stabilized, self-consistent belief formation.

Internally, we model the AI inference loop as a stratified meta-reasoning architecture: queries falling within a formally decidable domain are resolved with provable certainty, while queries outside this domain (e.g., out-of-distribution or ill-posed queries) are flagged and approximated with full epistemic transparency. This design echoes the halting problem and aligns with recent ideas from co-evolutionary cognition and edge-of-chaos dynamics in AI systems. In particular, we analyze the conditions under which hallucinations (overconfident errors) propagate or dissipate in recursive AI agent networks, how strategic heterogeneity suppresses epistemic collapse, and what edge-of-chaos and co-evolutionary dynamics are at play.

The work is directly applicable to LLM red-teaming, adversarial prompting, RAG pipelines, generative agents in specific technical domains, and self-auditing AI systems. Through its technical contributions, this model serves as a tool for understanding how generative AI systems co-evolve with human epistemic norms, and offers a path towards next‑generation autonomous cognitive systems that can safely self‑extend their reasoning reach and reliably integrate knowledge.

This paper studies the efficient disclosure of performance information by a principal about an agent who has career concerns vis-à-vis heterogeneous employers. We consider a two-period model with career concerns la Holmstrom (1999) where, in the second period, employers hire for jobs more or less similar to the job the agent is evaluated for and thus value information differently. The principal trades off effort incentives and the agent's welfare, and does not internalize the audiences' payoffs. In equilibrium, the principal discloses information differentially and information structures are non-decreasing in job similarity. Setting high evaluation standards for less similar jobs is only efficient for incentivizing high effort. The results are robust to imperfect competition between effort-maximizing principals and generalize to setups with fixed or output-contigent transfers. The model can rationalize laws differentially restricting access to worker and consumer personal data, such as criminal records or credit scores, and inform debates regarding privacy protection.

Mahjong, an imperfect information game with a state space exceeding $10^{48}$, poses significant challenges for reinforcement learning (RL) due to its sparse rewards and delayed feedback. Traditional RL approaches relying solely on terminal win/loss signals often suffer from inefficient convergence and suboptimal strategy development. This study proposes a hybrid reward mechanism that integrates prior knowledge and predictive modeling to address these limitations.

The framework combines three components to provide dense, interpretable feedback. First, an instructor model based on a CNN-LSTM architecture is trained on expert gameplay data to generate auxiliary rewards by evaluating the alignment between agent actions and expert-predicted optimal moves. Data augmentation techniques, such as tile suit permutation, are employed to enhance generalization despite the limited training samples. Second, a dual-layer GRU-based global reward predictor captures long-term dependencies in the game, forecasting the winning probabilities and score dynamics to enable intermediate reward signals. Third, the hybrid reward system integrates a base reward structure—assigning +3 for wins, -1 for losses, and 0 for draws—scaled by hand usingcomposition multipliers. Dynamic bonuses are further applied based on the consistency between the predicted and actual outcomes, while step-wise predictive rewards quantify decisions' impacts through the differences in successive value estimations.

Experimental evaluations on 10,000 2v2 Shangrao Mahjong games demonstrated the superior performance of the proposed method. Compared to supervised learning, Proximal Policy Optimization RL (PPO RL), multi-agent RL, and deep RL baselines, the hybrid mechanism achieved the highest average score of 5.68, outperforming the PPO RL baseline score of 5.24. The model reached a win rate of 26.56% with accelerated convergence, achieving a 30% win rate in 15 days versus the 35 days required by deep RL. Its strategic efficacy was validated by its maximum average win score of 15.49. By synthesizing knowledge-driven guidance and model predictive feedback, this approach effectively mitigates reward sparsity while enhancing training stability and the decision-making depth.