Worked on the vllm-project/aibrix repository to improve benchmarking reliability for Chain-of-Thought large language models. Addressed a critical issue in Time To First Token (TTFT) measurement by ensuring reasoning_content was accurately captured during streaming, even when content was not immediately available. This bug fix enhanced the accuracy and credibility of benchmarking data, enabling more reliable model comparisons and data-driven evaluation. Utilized Python for debugging and instrumentation, applying skills in benchmarking and LLM integration to strengthen the benchmarking pipeline. The work focused on refining measurement processes, resulting in more trustworthy metrics for evaluating reasoning capabilities in advanced language models.