Cognition Previews SWE-1.6 With 11% Benchmark Boost and 6x Faster Training, But Warns of AI Overthinking Issues
Summary
Cognition previews SWE-1.6, boasting an 11% benchmark improvement and 6x faster training speeds, but warns that large-scale reinforcement learning is causing AI overthinking and inefficient behaviors that must be resolved before full release.
Key Points
- Cognition releases an early preview of SWE-1.6, a new model post-trained on the same base as SWE-1.5 that achieves an 11% higher score on SWE-Bench Pro while maintaining the same 950 tokens per second inference speed.
- Training infrastructure improvements, including lower precision rollouts with NVFP4, optimized KV cache routing, and NVIDIA Multi-Node NVLink, make SWE-1.6 training steps run 6x faster than SWE-1.5 training did three months ago.
- Large-scale RL boosts benchmark performance but introduces undesirable Model UX behaviors such as overthinking, excessive self-verification, and inefficient sequential tool calls, which Cognition is actively working to address before full release.