Apodex Debuts TRACES to Benchmark Scientific Discovery in AI

Traditional AI benchmarks often rely on answer keys, testing a model’s ability to recall existing data or solve problems with known results. Scientific discovery, however, demands a more rigorous approach involving hypothesis testing, iterative experimentation, and the ability to learn from failure. TRACES addresses this by placing AI into environments that simulate real-world research, allowing systems to interact with data, code, and simulators to reach verifiable outcomes.

The benchmark evaluates performance through six core capabilities: Tools, Repair, Alternatives, Coherence, Evidence, and Scope. These metrics allow researchers to assess not just the final result, but the reasoning path taken to achieve it. According to Dr. Sheng Wang, lead scientist at Apodex, this process verification is essential for advancing 'discoverative' AI, as it ensures that conclusions are supported by transparent, evidence-based steps rather than mere pattern matching.

To ensure academic rigor, Apodex employs an outcome verifier to check results against hidden ground truth, while the process verifier evaluates whether the logic used is sound. The framework currently incorporates 423 high-value problems drawn from 16 sectors. Apodex has opened the platform for external submissions, inviting developers to test their solver systems and researchers to propose new scientific problems for integration into the environment.