TTT-Discover is a new approach that applies reinforcement learning at test time for large language models (LLMs). This method allows the model to adapt based on specific challenges it encounters, leading to improved performance in various tasks. The results claim state-of-the-art achievements in several domains, including mathematics, GPU kernels, algorithms, and biology. For example, in the mathematics category, TTT-Discover achieved a score of 0.380876, just below the best human score of 0.380927. In GPU kernel competitions, it significantly reduced processing times, posting 2198 µs for the A100, compared to the best human time of 4531 µs. Key performance metrics highlight TTT-Discover's capabilities. In algorithm engineering, it outperformed previous AI benchmarks, scoring 567,062 in the AtCoder heuristic contests, which is higher than the previous best AI score of 558,026. In biology applications, it achieved a denoising performance of 0.71 and 0.73 for different datasets, surpassing the best human scores of 0.64. The project emphasizes collaboration with various communities and builds on existing frameworks like AlphaEvolve and Tinker, which are related to GPU optimization and LLM training. The implementation requires specific environment setups, including SLURM and various dependencies outlined in the requirements files. Users can launch experiments using a Python script or a preconfigured bash script. The article also suggests that a simplified API is on the way, indicating ongoing improvements to enhance user experience. Overall, TTT-Discover showcases a promising advancement in LLM adaptability and efficiency across multiple challenging tasks.