Brian Wang, Bin Feng, Xiaoman Pan, Chenyang An, Felix Liu, Tangqi Fang, Gongbo Sun, Lingfeng Shen, Ning Wang, Handuo Zhang, Feng Chen, Fuchao Yang, Xiang Wang, Jiacheng Lin, Siting Li, Zixuan Liu, Chi Han, Zhenhailong Wang, Kunlun Zhu, Lawrence Zhao, Yueqi Guo, Kailong Wen, Feng Xing, Yiling Guo, Lidong Bing, David Tan, Bo An, Heng Ji, Sheng Wang
6 min
Abstract
Apollo did not reach the Moon merely because its engineers could solve difficult equations. It succeeded by turning a distant ambition into a mission architecture of explicit objectives, simulation, verification, and repeated correction. AI now faces a similar transition: frontier models can solve difficult tasks once the problem, tools, and success criteria are specified, yet consequential real-world challenges rarely arrive in an executable or verifiable form. We introduce Apodex Discovery, a framework for building and evaluating discoverative AI through the heavy-duty solver, a system comprising a foundation model, harness, tools, and control policies that pursues extended, stateful, verifiable investigations. It has three core components. First, a problem-scouting process surveyed 561 industries across 16 sectors, assembled 423 high-value real-world problems, and selected 20 for the initial release. Second, a common environment-task-episode abstraction provides data, tools, constraints, feedback, trajectory recording, and verification of intermediate artifacts and final submissions. Third, HDS6 evaluates Tools, Repair, Alternatives, Coherence, Evidence, and Scope independently of final-task success. In AAV capsid design, Apodex surpassed the published state of the art by 7% across viability, tropism, structure prediction, and generative design. In drug repurposing and reformulation, a task-specific biomedical 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. Controlled ablations show that the fixed TRACES episode interface enables attribution of performance differences to specific solver components. Apodex Discovery moves AI evaluation beyond predefined benchmarks toward verifiable investigations aimed at genuine discovery.
Alex: So the quality of the thinking is being assessed, not just the conclusion.
Sam: Precisely. And that rubric does double duty. When a trajectory—a recorded sequence of the AI's actions—falls short, the same evaluation generates a targeted repair note. It tells the system specifically what went wrong, without revealing the hidden answer. The AI then reruns the work with that guidance.
Alex: Does the repair loop actually help?
Sam: Across hundreds of deficient trajectories tested in the study, the repaired reruns scored higher on average. Most improved rather than declined. It's a meaningful result, though the authors are careful not to overstate it—performance still depends heavily on the quality of the verification setup.
Alex: They tested this on a real biological problem, right?
Sam: They did. The domain was adeno-associated virus capsid design—that's the engineering of protein shells used to deliver gene therapy into cells. It's a genuinely hard problem with measurable outcomes. The system, operating inside a domain-specific environment with the right tools and biological constraints, surpassed previous expert methods across several key metrics: viability, structural prediction, and generative design.
Alex: So putting the AI inside a well-built environment with the right tools made a measurable difference to the actual science.
Sam: That's what the paper suggests. The environment isn't just scaffolding—it's doing real work.
Alex: What's the main thing holding this approach back from being applied more broadly?
Sam: Building these environments is extremely labor-intensive. You need domain experts to set up the constraints, the tools, the hidden verification checks—and those checks have to be leak-resistant, meaning the AI can't accidentally see the answer it's supposed to find. Right now, that level of careful engineering limits the approach to a small number of carefully curated domains.
Alex: So the bottleneck isn't computing power—it's human effort in designing the tests themselves.
Sam: For now, yes. Which is a significant constraint. But the authors argue it's a solvable one. As more of these environments get built, the framework becomes more broadly applicable. And when you zoom out, the direction is notable: we're moving toward autonomous systems that can run multi-step research campaigns across disciplines—not just retrieve information, but actually generate and test new knowledge.
Alex: So instead of answering questions, future AI might actually run the experiments.
Sam: That's the long-term trajectory. By combining structured problem framing with rigorous, multi-layer verification, the goal is to push AI past pattern matching into something closer to genuine scientific discovery. The Apodex framework is one early attempt to show what that might look like in practice.
Alex: It's a meaningful step—and an honest one, given how clearly the authors flag what still needs to be solved. Thanks for walking through it.
Sam: Thank you for listening to ResearchPod.