KARL: Knowledge Agents via Reinforcement Learning

arXiv 2603.05218·medium signal

Enterprise search agent trained with iterative large-batch off-policy RL across 6 search regimes (entity search, cross-doc synthesis, tabular reasoning, exhaustive retrieval, procedural reasoning, fact aggregation). Achieves Pareto-optimal cost-quality and latency-quality tradeoffs vs Claude 4.6 and GPT 5.2, including on out-of-distribution tasks. Multi-task training generalizes better than single-benchmark optimization. Relevant for builders shipping RAG/search agents in enterprise settings.