Apodex Discovery beats SOTA on AAV capsid design by 7%
TL;DR
- Apodex Discovery selected 20 problems from 423 candidates surveyed across 561 industries and 16 sectors for its initial benchmark release.
- On AAV capsid design the system beat the published state of the art by 7% across viability, tropism, structure prediction, and generative design.
- The HDS6 rubric grades Tools, Repair, Alternatives, Coherence, Evidence, and Scope independently of whether the final task succeeded.
Apodex Discovery, published as an arxiv preprint on August 11, 2026, reports that its 'heavy-duty solver' surpassed the published state of the art on AAV capsid design by 7% across viability, tropism, structure prediction, and generative design.
The framework is pitched as a benchmark for what the authors call 'discoverative AI', systems that pursue 'extended, stateful, verifiable investigations' rather than solving predefined tasks with known answers. The initial release holds 20 problems, drawn from a scouting process that, per the paper, 'surveyed 561 industries across 16 sectors' and assembled 423 high-value candidates.
A companion evaluation, HDS6, scores 'Tools, Repair, Alternatives, Coherence, Evidence, and Scope independently of final-task success', a deliberate separation of process quality from outcome. On a biomedical drug repurposing task, a specialized environment 'improved the mean normalized prediction score of GPT-5.5 and GPT-5.6-sol by 2.5 and 7.6 points over the same closed-book backbone,' the abstract reports.
The preprint does not name the AAV baseline the 7% margin is measured against, and the 20 released problems are not itemized in the abstract.
Shared on Bluesky by 1 AI expert
Originally reported by paper
Read the original article →Original headline: Apodex Discovery Defines 'Discoverative AI' Benchmark on 20 Real Scientific Problems, Beats SOTA on Protein Design