Hugging Face researchers have introduced Apodex Discovery, a new framework designed to push AI beyond conventional benchmarks by enabling verifiable, extended investigations into real-world scientific problems. The framework centers on a "heavy-duty solver" that combines a foundation model with tools, harnesses, and control policies to pursue complex, stateful investigations.
The approach is motivated by the Apollo program's success, which relied on explicit objectives, simulation, and iterative correction rather than raw problem-solving ability. Similarly, Apodex Discovery aims to turn ambitious AI goals into structured, verifiable missions. The framework includes a problem-scouting process that surveyed 561 industries across 16 sectors, assembling 423 high-value problems and selecting 20 for initial release.
A key component is the environment-task-episode abstraction, which provides standardized data, tools, constraints, feedback, and verification for intermediate and final outputs. The evaluation metric, HDS6, assesses six dimensions—Tools, Repair, Alternatives, Coherence, Evidence, and Scope—independently of final task success.
In tests, Apodex surpassed published state-of-the-art results in AAV capsid design by 7% across viability, tropism, structure prediction, and generative design. In drug repurposing, the framework improved mean normalized prediction scores by 2.5 and 7.6 points for GPT-5.5 and GPT-5.6-sol, respectively, compared to closed-book baselines. Controlled ablations confirmed that the fixed TRACES episode interface enables clear attribution of performance differences to specific solver components.
The authors position Apodex Discovery as a shift from generative AI to "discoverative AI," targeting problems without ground-truth answers. More details are available at the project website.