Developed and integrated the VisFactor benchmark task into the EvolvingLMMs-Lab/lmms-eval framework, delivering a rule-based scoring pipeline that normalizes multiple answer types and mirrors the VLMEvalKit standard. The work included implementing two-level aggregation for evaluation parity on thousands of real-model predictions, as well as creating new task configurations and data-processing utilities. Leveraging Python, YAML, and regex, the solution enables reproducible, vendor-agnostic benchmarking of vision-cognition models while reducing reliance on external LLM judges. Comprehensive documentation and parquet data conversion references were provided, streamlining data ingestion and supporting quick adoption within the open-source evaluation ecosystem.