Selective Reincarnation introduces a dynamic population-based training mechanism tailored for multi-agent reinforcement learning. Each agent’s performance is regularly evaluated against predefined thresholds. When an agent’s performance falls below its peers, its weights are reset to those of the current top performer, effectively reincarnating it with proven behaviors. This approach maintains diversity by only resetting underperformers, minimizing destructive resets while guiding exploration toward high-reward policies. By enabling targeted heredity of neural network parameters, the pipeline reduces variance and accelerates convergence across cooperative or competitive multi-agent environments. Compatible with any policy gradient-based MARL algorithm, the implementation integrates seamlessly into PyTorch-based workflows and includes configurable hyperparameters for evaluation frequency, selection criteria, and reset strategy tuning.

Text-to-Reward provides a pipeline to train reward models that map text-based task descriptions or feedback into scalar reward values for RL agents. Leveraging transformer-based architectures and fine-tuning on collected human preference data, the framework automatically learns to interpret natural language instructions as reward signals. Users can define arbitrary tasks via text prompts, train the model, and then incorporate the learned reward function into any RL algorithm. This approach eliminates manual reward shaping, boosts sample efficiency, and enables agents to follow complex multi-step instructions in simulated or real-world environments.