The paper identifies four system limitations: BM25 may bottleneck as the bank becomes large and diverse, the curator adds one LLM call per task, payload formats are fixed and benchmark-specific, and curation occurs only once per task rather than adapting at each turn or step. · CiteArk
The paper identifies four system limitations: BM25 may bottleneck as the bank becomes large and diverse, the curator adds one LLM call per task, payload formats are fixed and benchmark-specific, and curation occurs only once per task rather than adapting at each turn or step.