Publicity Highlights

Introducing HELIX: From AI Assistance to Autonomous Experimentation

11 Sep 2026

The way we conduct computational research is beginning to change.

Developing an approach, writing and debugging code, running experiments, evaluating results and deciding what to try next: this cycle takes substantial time and effort. Agentic AI can carry out more of that execution autonomously, while scientists set the direction, contribute domain knowledge and judge scientific validity.

At A*STAR IAIC, we developed HELIX to explore this possibility.

HELIX brings domain data, scoring rules and scientific tools into a continuous research loop. Its agentic harness—the system around the foundation model—coordinates how the agent plans, uses tools, conducts experiments and adapts to the evidence it produces.

Paired with Claude Opus 5, HELIX currently ranks first on the full NatureBench leaderboard.

NatureBench evaluates AI coding agents on 90 scientific machine-learning tasks drawn from peer-reviewed papers in Nature-family journals, spanning six scientific domains. Agents receive the problem and data, while the original method is withheld and web search is disabled. Their solutions are evaluated against the papers’ published results.

For each task, HELIX was given a four-hour time budget and operated with no human in the loop. Within that window, it autonomously developed its approach, wrote and debugged code, ran experiments, evaluated results and decided what to try next—without human guidance or intervention during execution.

Under these constraints, across the 90 tasks, HELIX:

  • Surpassed the published state of the art on 26.7% of tasks, using the benchmark’s greater-than-10% improvement threshold.
  • Matched or exceeded it on 62.2%.
  • Achieved a 97.8% valid completion rate.

HELIX also outperformed the Claude Code configuration using the same Opus 5 model on these three measures. These system-level results provide encouraging evidence that the way we harness a model can make a meaningful difference to scientific performance and reliability—even within a fixed time budget and without human intervention. View the leaderboard.

Nature Bench Leaderboard Helix

Nature Bench Match Sota Helix

For IAIC, this is a step towards turning frontier AI models into dependable scientific capability. For Singapore, such systems could help shorten research cycles and enable our researchers to explore more ideas with the resources available.

Our next step is to demonstrate HELIX on scientific and engineering challenges beyond the benchmark and progressively connect computational experimentation with the physical world.

Follow us on LinkedIn for more innovations.