Most benchmarks assume the answer exists somewhere. Scientific discovery often starts precisely because it does not.

Apodex released TRACES, which is built around that mismatch. A new benchmark for testing AI systems on difficult scientific problems where the correct answer may not already be known.

Instead of scoring only the final answer, TRACES also evaluates how the AI works through the problem, including its tool use, error correction, evidence, and reasoning process.

TRACES turns research problems into executable environments where a solver can use data and tools, receive feedback, revise its approach, and leave behind a full trajectory.

That trajectory is evaluated across six capabilities: Tools, Repair, Alternatives, Coherence, Evidence, and Scope. A hidden verifier separately evaluates the outcome.

So a system can be diagnosed for choosing the wrong tool, failing to recover from feedback, dropping a competing hypothesis, or making a claim broader than its evidence supports, even before one aggregate score hides those failures.

The live leaderboard makes those solver profiles comparable. Researchers can also submit new problems, while teams can submit the entire solver, including the model, harness, tools, memory, and control policy.

With @tianqiao_chen's background in brain science, I can see why process evaluation sits so centrally here. For long-horizon reasoning, where the system fails matters as much as whether it failed.