Alignment Science in the Dawn of Cooperative Intelligence

This paper reviews the state of AI alignment and decision-making processes. It formally introduces the Third-Way Alignment (3WA) framework for researching, developing, and deploying aligned AI systems. 3WA moves beyond control-based paradigms toward cooperative intelligence and ethical partnership. This model is grounded in Mutual Respect, Shared Flourishing, and Ethical Coexistence. Technically, we synthesize three families of methods: (a) preference learning, (b) process-based supervision, and (c) verifiable-feedback reinforcement learning. We analyze technical evidence showing outcome-only supervision (like RLVR) struggles with sparse success signals and can reward flawed reasoning. In contrast, process-based supervision (PBS) provides dense, step-wise feedback. PBS significantly outperforms outcome-only methods on complex reasoning tasks. We integrate these findings with Supervised Reinforcement Learning (SRL), which uses step-wise action rewards to build procedural competence (Deng et al., 2025). We propose an SRL to RLVR pipeline as a robust training regimen. We contribute: (1) a taxonomy integrating cooperative intelligence with technical alignment methods, (2) a formalization of 3WA technical and organizational pillars, and (3) an evaluation protocol combining process, outcome, safety, and interpretability metrics. We argue this pipeline is effective for smaller models trained on high-quality data. SRL avoids the performance degradation SFT exhibits on complex datasets (Deng et al., 2025).