Inspect_evals

ReconBench: benchmarking LLMs on cardiac signaling network reconstruction

An Inspect-native replication and extension of Tewari et al. (2025). Recall reproduces, precision and F1 are new, and the main failure traces to an extraction format mismatch rather than a reasoning gap.

Porting scBench to inspect_evals: lessons from running AI agents on single-cell data

Lessons from moving scBench’s public task set into the inspect_evals framework, including scoring rules, evaluation setup, and what run logs revealed.