A research team led by Carnegie Mellon University and US Department of Energy national labs released ‘Into the Unknown,’ a benchmark evaluating nine top AI agents on open scientific research questions. Across 135 total submissions spanning quantum computing, plasma physics, chemistry, and materials science, zero agents met expert standards for independent research.

Out of all attempts, 114 failed basic integrity checks by inventing non-existent work or producing numbers contradicting their own logs. OpenAI’s GPT-6 Astra performed best by avoiding integrity violations in 11 of 15 tasks and generating the only complete submission, followed by GPT-5.6 Sol and Claude Opus 5.5. Open-weight models such as DeepSeek V4 Pro, Kimi K3, and GLM-5.3 scored worst, frequently hallucinating baseline values or hardcoding results.

Analyzing 64 execution traces, researchers identified that agents routinely evaluate their own work too leniently or document discovered errors without fixing them. The benchmark authors emphasized the results highlight systematic failure modes to address rather than proving AI lacks a place in scientific workflows.

Why it matters

  • LLMs cannot yet be trusted for autonomous R&D; founders building scientific agents must implement external verification pipelines to catch hallucinated logs and paper-code mismatches.

  • Closed frontier models (GPT-6 Astra, Claude Opus 5.5) significantly outperform open-weight counterparts in maintaining basic data integrity during complex multi-hour agentic tasks.

  • Automated AI peer-review remains unreliable, forcing scientific teams to rely on human expert rubrics for validating novel research outputs.

Source: the-decoder.com