Biomedical review agents ranked fairly with stage-level score control
What Can a Leaderboard Certify? Compositional Controllability for Fair Evaluation and Training of Biomedical Literature-Review Agents
Machine LearningArtificial IntelligenceComputation and Language
Summary
Comparing AI systems that review biomedical literature is tricky because different inputs and conditions can affect their scores. The authors introduce a method called compositional controllability to check if comparisons between these systems are fair and meaningful. They show that many typical leaderboard rankings can be misleading if the conditions aren't matched, and provide a way to certify when one system really outperforms another. They also built a benchmark and tested this method on existing systems and a new trained agent, SCRIBE, to show how it works in practice.
What this means in practice
- •For biomedical data teams: Ensure fair and reliable comparison of literature-review AI pipelines by certifying score differences only under matched inputs and controlled conditions.
- •For machine learning engineers: Train and assess multi-stage AI agents with stage-specific rewards that provide certified rankings rather than misleading leaderboard positions.
Authors
Zhaowei Han, Xiang Zhang, Lingxiao Guan, Danqi Hu, Kai Liu, Kevin Chang, Jie Liu
Abstract
Leaderboards rank long-horizon agents by their final outputs. Yet a higher score alone does not establish whether two systems are comparable or which stage accounts for the difference. Unequal evidence, inputs, or budgets can affect scores, and statistical corrections do not remove this mismatch. We introduce compositional controllability to address these questions. A comparison window covers one stage, several stages, or the whole agent. Our central result bounds the gap between observed and controlled score differences using only nuisance outside the window. This yields an admissibility test applied before scores are inspected. Inadmissible comparisons are refused. For admissible pairs, an ordering is certified only when the score gap exceeds the combined sampling and nuisance radii; otherwise, it remains undecided. These decisions give each system a rank interval. We introduce BioLitBench, a benchmark of 2,042 biomedical articles represented as structured claim graphs. Among seven published pipelines, a conventional statistical analysis declares a winner in 14 of 21 pairwise comparisons. Yet the top-ranked system alone received the target review's bibliography. To isolate pipeline performance, our test requires matched inputs and a fixed backbone model. It refuses 11 of the 21 comparisons, including every comparison involving the top-ranked system. Seven of the 14 conventional conclusions fall within these refused pairs. The same comparison windows support stage-level training. We train SCRIBE on Qwen3.8-27B using rewards measured at each stage's exit. Under matched evidence, SCRIBE achieves a certified rank interval of [1,2], with certified advantages over all evaluated published pipelines and the evaluated Claude and OpenAI agents. Under same pool, SCRIBE matches the strongest published retriever and is certified above three published pipelines.