Apodex Launches TRACES Benchmark for Scientific Discovery AI

Apodex announced in a press release TRACES, a benchmark for evaluating AI systems on scientific discovery tasks where answers may not be known in advance. The benchmark uses executable environments instead of static datasets with answer keys.

TRACES gives an AI system a work environment that can include scientific literature, structured datasets, code execution, specialized tools, simulators, folding engines, experimental feedback, or other interfaces. The system must choose actions, interpret results, adjust its approach, and work toward a verifiable outcome.

The benchmark evaluates both final results and the process used to reach them. Apodex said TRACES measures six capabilities: tool use, error repair, alternative hypotheses, coherence across long work sequences, evidence grounding, and scope limits for conclusions.

Apodex said TRACES is open for participation. Teams can submit solver systems for evaluation, and researchers or organizations can propose scientific problems to be turned into executable benchmark environments.