Adaptive Switching Between Self-Reinforcement and Mutual Reinforcement Learning for Multiple Large Language Models

Abstract

Large language models (LLMs) can generate candidate answers, evaluate their quality, and use evaluation signals to improve subsequent responses. However, self-evaluation may reinforce a model's own biases and errors, whereas evaluation by other models may introduce evaluator disagreement, capability mismatches, and correlated errors. This paper proposes Adaptive Switching between Self-Reinforcement and Mutual Reinforcement Learning (ASMR), a multi-LLM learning framework that dynamically switches between independent self-evaluation and cross-model evaluation during iterative training. Given the same input problem, multiple LLMs independently generate candidate answers and assign quality scores to those answers. During self-reinforcement phases, each model evaluates its own candidates and learns from its preferred responses. During mutual-reinforcement phases, models evaluate candidates produced by other models, exchange preference information, and learn from cross-model feedback. A switching controller selects the learning mode according to evaluation disagreement, evaluator reliability, model-specific performance trends, and the availability of informative preference signals. The framework preserves model-specific parameters while permitting controlled exchange of training information. To reduce the risk of self-reinforcing errors, ASMR incorporates score calibration, uncertainty-aware selection, independent verification where available, and safeguards against repeated optimization of unreliable evaluations. We formulate the learning process, define a switching policy, and propose an experimental protocol for comparing adaptive switching with self-reinforcement alone, mutual reinforcement alone, and fixed-proportion hybrid learning. This paper presents a research framework and testable hypotheses rather than empirical findings. Its principal contribution is the explicit formulation and evaluation of adaptive switching between two learning pathways in a multi-LLM system.