Researchers published a paper on improving terminal agents with reinforcement learning. They propose the Agentic Compositional Generalization hypothesis. It suggests RL shapes high-level decision-making rather than teaching new skills. They introduce River, a training recipe for better reward quality. River filters low-quality environments and adds process-level behavior regularization. It achieves the best performance among open-source RL-trained 8B models on four benchmarks. River generalizes across model families, scales, agent harnesses, and RL objectives. It uses fewer than 30% of TMax training environments. RL gains improve by 106% on Terminal-Bench-Lite and 30% on Terminal-Bench-v2.1. Results cover models from 2B to 27B parameters.

• 106% RL gain on Terminal-Bench-Lite
• 30% RL gain on Terminal-Bench-v2.1
• 2B to 27B model range
• <30% of TMax training environments