MIT researchers released experiment logs, code, and prompts on GitHub, published the Qwen3.5-0.8B model edited to eliminate repetitive reasoning on Hugging Face, and announced ongoing benchmark development