KARL: Knowledge Agents via Reinforcement Learning

Dimensions overall score 8.0

• KARL achieves state-of-the-art performance across a diverse suite of hard-to-verify agentic search tasks.result 90%
  Evidencepartial

  We present a system for training enterprise search agents via reinforcement learning that achieves state-of-the-art performance across a diverse suite of hard-to-verify agentic search tasks.

  Implicationpartial

  Explicitly stated in the abstract as a core finding of the paper

  Verificationpartial

  partial

• Models trained across heterogeneous search behaviors generalize substantially better than those optimized for any single benchmark.result 85%
  Evidencepartial

  Second, we show that models trained across heterogeneous search behaviors generalize substantially better than those optimized for any single benchmark.

  Implicationpartial

  Directly stated as a core contribution in the abstract with supporting experimental results implied

  Verificationpartial

  partial

• KARL is Pareto-optimal on KARLBench across cost-quality and latency-quality trade-offs compared to Claude 4.6 and GPT 5.2.result 90%
  Evidencepartial

  Compared to Claude 4.6 and GPT 5.2, KARL is Pareto-optimal on KARLBench across cost-quality and latency-quality trade-offs, including tasks that were out-of-distribution during training.

  Implicationpartial

  Explicitly stated in the abstract with clear comparative metrics

  Verificationpartial

  partial

• The system introduces KARLBench, a multi-capability evaluation suite spanning six distinct search regimes.method 95%
  Evidencepartial

  First, we introduce KARLBench, a multi-capability evaluation suite spanning six distinct search regimes, including constraint-driven entity search, cross-document report synthesis, tabular numerical reasoning, exhaustive entity retrieval, procedural reasoning over technical documentation, and fact aggregation over internal enterprise notes.

  Implicationpartial

  Explicitly stated as the first core contribution with specific details provided

  Verificationpartial

  partial

• The system uses an agentic synthesis pipeline that employs long-horizon reasoning and tool use to generate diverse, grounded, and high-quality training data.method 85%
  Evidencepartial

  Third, we develop an agentic synthesis pipeline that employs long-horizon reasoning and tool use to generate diverse, grounded, and high-quality training data, with iterative bootstrapping from increasingly capable models.

  Implicationpartial

  Directly stated as a core contribution in the abstract with specific methodology described

  Verificationpartial

  partial

• The system proposes a new post-training paradigm based on iterative large-batch off-policy RL that is sample efficient and robust to train-inference engine discrepancies.technical 85%
  Evidencepartial

  Fourth, we propose a new post-training paradigm based on iterative large-batch off-policy RL that is sample efficient, robust to train-inference engine discrepancies, and naturally extends to multi-task training with out-of-distribution generalization.

  Implicationpartial

  Directly stated as a core contribution with specific technical approach described

  Verificationpartial

  partial

• With sufficient test-time compute, KARL surpasses the strongest closed models.result 80%
  Evidencepartial

  With sufficient test-time compute, it surpasses the strongest closed models.

  Implicationpartial

  Explicitly stated in the abstract but conditional on sufficient compute resources

  Verificationpartial

  partial

• The primary limitation is reliance on proprietary datasets which might limit applicability in contexts where such data isn't available.limitation 90%
  Evidencepartial

  The primary limitation is the reliance on proprietary datasets which might limit applicability in contexts where such data isn't available.

  Implicationpartial

  Explicitly stated in the analysis section as a caveat/limitation

  Verificationpartial

  partial

Competitive landscape

KARL leverages innovative reinforcement learning for affordable, high-performance enterprise search agents.