A new training method recycles benchmark failure data instead of discarding it, and its creators say the resulting medical AI agent beats a Claude Code setup on a niche benchmark.
Researchers built Benchmark-as-Teacher (BaT), a system that pairs an asynchronous data pipeline called Stage Bank with a training method called BiCuRL, short for Bilevel Curriculum Reinforcement Learning. Stage Bank generates isolated training scenarios outside the main training loop. BiCuRL then picks the next curriculum from a held-out evaluation set, checks each attempt against stage-by-stage rubrics, and updates the model using GRPO, a reinforcement-learning technique. On the AutoMedBench-Lite benchmark, the resulting BaT-4B and BaT-9B agents more than doubled the Overall scores of their base Qwen Instruct models. BaT-9B hit an Overall score of 79.6, edging out a comparison score of 77.5 attributed to what the paper calls Claude Opus 4.6 with Claude Code.
That's notable because medical imaging data is scarce and hard to share for privacy reasons, and expert task trajectories for training agents are even harder to come by. A cheap self-improvement loop that squeezes more performance out of small, open 4-billion and 9-billion parameter models, rather than requiring ever-bigger frontier systems, would matter a lot for hospitals and researchers who cannot ship patient data to a third-party API.
One problem: there is no Claude model called Opus 4.6 in Anthropic's published lineup. The closest real releases are Opus 4.8 and Sonnet 4.6, an easy detail to garble in a lightly-reviewed preprint, and reason enough to treat that specific comparison number with some caution.