- Benchmarking is a form of data activation, transforming domain data into measurable tasks for models to be evaluated and trained on.
- Clear, verifiable metrics are essential for model improvement, as seen in coding and math, but complex domains like medicine lack inherent benchmarks.
- Activation starts with measurement: converting health data into scores for models can reveal knowledge gaps and drive improvement, even without model changes.
- Verifiers integrate benchmarking with reinforcement learning, turning scores into rewards and merging measurement with optimization, which amplifies both benefits and risks.
- The key challenge is making domains amenable to scale by converting messy data into verifiable tasks, with benchmarking as a conversion method.
- Different benchmark approaches exist on a spectrum: from expensive, raw-data-based methods (e.g., latchbio) to scalable, rubric-based ones (e.g., HealthBench) and lightweight multiple-choice setups (e.g., MedMarks).
- QuestBench highlights the importance of noticing missing information, which mirrors real-world tasks like medical diagnosis where identifying gaps is crucial.
- Curation remains a core challenge; expert judgment shapes benchmarks, but making it explicit in reward functions increases inspectability and transparency.
- Benchmark flaws become incentives when used as rewards, potentially leading models to learn distorted behaviors, emphasizing the need for careful design.
- Building benchmarks is a concrete form of data activation, involving creating verifiable tasks that enable scale to drive progress in complex domains.