Evaluating large language models typically requires thousands of benchmark items, making the process expensive and increasingly impractical. Existing protocols rely on average accuracy over fixed item sets, treating all items as equally informative—yet many items are too easy or too hard to distinguish between models, making evaluation both wasteful and imprecise. Item Response Theory (IRT) with Computerized Adaptive Testing (CAT) offers a principled alternative—separately estimating model ability and item difficulty, then adaptively selecting the most informative items. However, practical deployment is bottlenecked by costly item calibration, and standard greedy selection causes severe over-exposure of frequently picked items while leaving the rest of the bank under-utilized, which risks data contamination over time. We propose FAIR-Eval with two components: Feature-Driven Calibration (FDC), which jointly learns item parameters from text features and sparse responses, enabling low-cost bank expansion requiring only minimal model responses; and Maximum Fisher Information Plus (MFI+), which counteracts greedy selection by incorporating exposure control, while simultaneously enforcing difficulty coverage and domain diversity to systematically activate under-utilized items. On a 6,796-item bank (193 calibration models, 50 held-out), FAIR-Eval achieves Pearson r = 0.9882 and Spearman 𝜌 = 0.9828 against full evaluation using around 3% of items. Crucially, it broadens difficulty coverage from 60.0% to 68.7%, achieving a more comprehensive and robust assessment of model capabilities with minimal test length.

Reinforcement learning with outcome-based objectives such as GRPO enables LLM-based agents to solve complex, long-horizon tasks, yet the reusable exploration patterns embedded in interaction trajectories are largely discarded after a single policy update. Existing experience-augmented approaches retrieve historical guidance at inference time, but they apply experiences without accounting for the policy’s evolving capability and create persistent dependencies on external retrieval. We propose EDGE (Experience-Distillation for Guided Exploration), a framework that treats retrieved experiences as temporary training-time scaffolds and progressively internalizes their benefits into the parametric policy. Concretely, EDGE partitions each rollout group into experience-conditioned and experience-free trajectories to estimate and admit only positive marginal gains without extra sampling, then distills the induced behavior into the base policy via a reverse-KL objective on its own empirical support. A co-evolutionary experience bank further synthesizes guidance from emerging failure modes and prunes obsolete entries as the policy evolves. Across embodied, web, and search-based QA tasks, EDGE improves over strong RL baselines by up to 12.5 points and remains effective without inference-time scaffolds or a proprietary reflector. The code is available at https://github.com/xvolcano02/EDGE.