Large Language Models (LLMs) frequently face limitations, such as judgment biases and cognitive deadlocks, in complex reasoning due to “Degradation of Thought” in self-reflection. While Multi-Agent Debate (MAD) approaches attempt to address this issue, their inherently adversarial nature suppresses diverse perspectives by overemphasizing antagonism, which results in the loss of partially valid reasoning. In this paper, we propose Multi-Agent Diverse Thinking (MADT), a novel cognitive framework that redefines multi-agent collaboration by shifting from adversarial confrontation toward constructive synergy. To break the oppositional deadlock found in traditional debates, MADT decomposes the thinking process into fine-grained modules: an Affirmative Thinker to preserve rational components, a Critical Thinker to rectify errors, and a Growth-Minded Thinker to provide optimization suggestions. These agents jointly foster a synergistic environment where specialized roles protect valid logic while iteratively refining flaws. Meanwhile, a Leader agent coordinates the collaborations and synthesizes the feedback from multiple thinkers to produce the final result. Extensive experiments on Common Machine Translation and Counter-Intuitive Arithmetic Reasoning tasks show that MADT consistently outperforms Self-Reflection and MAD baselines, validating the superior ability to enhance complex logical and reasoning performance in LLMs.

Recent advancements in Large Language Models (LLMs) are increasingly driven by using LLMs themselves to synthesize large amounts of instruction-tuning data. To transcend the limitations of single model synthesis, recent research has shifted toward multiple model data synthesis, where distinct models assume specialized roles to synthesize data through interaction. However, existing feedback cannot distinguish truly high-quality instructions, those that promote complex reasoning, deviate from the learned distribution, and require multi-source knowledge, due to coarse granularity and signal dilution. In human cognition, truly high-quality instructions require greater mental effort—the **Cognitive Load**—to address. In this paper, we propose a paradigm shift in feedback. Instead of external, result-oriented evaluation, we propose an internal, process-aware feedback based on target models’ computational effort. The novel feedback, namely **Inference Stress Index (ISI)**, quantifies instruction quality in a fine-grained, dilution-resistant manner by focusing on high-uncertainty tokens. Building on this, we present the **Co**gnitive-**F**eedback **A**daptive **G**eneration (**CoFAG**) framework, a self-evolving multiple model collaboration paradigm involving a synthesizer model, a fixer model, and a collaborator model. Extensive experiments demonstrate that CoFAG significantly outperforms other data synthesis techniques on complex reasoning benchmarks, validating the superiority of using internal cognitive feedback for generating high-quality data.