Tag
This paper presents BudgetBench, a protocol and harness for evaluating memory strategies in local large language model agents by varying per-call input-token budgets, offering a reusable measurement surface for community benchmarking.
This paper introduces answer-in-context, a diagnostic metric for budget-constrained multi-hop RAG that measures whether the gold answer survives in the packed reader context, and proposes a submodular evidence packing method that improves over heuristics under specific conditions.