Introducing AIBuildAI Science: #1 on NatureBench for Autonomous AI-for-Science Model Development
We are excited to introduce AIBuildAI Science, a vertical AI agent built on top of AIBuildAI Agent 2.5 for autonomously developing high-performance AI models for scientific research. AIBuildAI Science achieved #1 on the NatureBench leaderboard, a benchmark comprising 90 real-world AI-for-Science tasks derived from papers published in leading Nature family journals. Across all 90 tasks, AIBuildAI Science surpassed published state-of-the-art (SOTA) methods on 25.6% of tasks and matched published SOTA performance on 57.8% of tasks. Beyond performance, AIBuildAI Science is highly efficient. Every NatureBench task was completed fully autonomously within a 4-hour time budget, demonstrating the agent’s ability to rapidly develop high-performance AI models with minimal human intervention.
AI agents have shown strong capability in writing code, implementing methods, and reproducing published scientific results, yet their ability to propose novel methods for critical research problems remains far less examined. NatureBench [1] was built to test precisely that. Its 90 tasks are distilled from papers across six Nature-family journals published between 2022 and 2025, spanning cellular omics, protein biology, biomedical and physical modeling, molecular design, and relational reasoning; each asks whether an agent can independently match or beat the method the authors published.
AIBuildAI-Science is an autonomous system for developing AI models for scientific problems. Given only a task and its data, it designs model architectures, writes code, trains models, analyzes its own experimental results, and iteratively improves its solutions without human intervention. It organizes this process as a search tree of candidate solutions, where each new candidate is built from a diagnosis of what the previous ones got wrong. By learning from both successes and failures, it keeps refining its search, so the solutions it finds get stronger the longer it runs.
Applied to NatureBench, AIBuildAI-Science surpasses the published state of the art on a quarter of the 90 tasks and outperforms every baseline agent on the leaderboard. The benchmark’s authors evaluated twelve frontier configurations built on the strongest available models; the best of them surpassed the published result on 15.6% of tasks and matched it on 41.1%, while AIBuildAI-Science reaches 25.6% and 57.8%. The second number is the one worth pausing on: it is the first time a system has matched or beaten the published result on a majority of these papers rather than a minority. Produced fully autonomously within four hours per task, these solutions suggest that AI agents are beginning to contribute to scientific discovery itself, not only to the engineering around it.
Inside NatureBench
NatureBench comprises 90 tasks, each distilled from a single Nature-family paper and delivered as a self-contained container. The agent can read a task brief, a data description, and the paper’s real input data; the held-out ground truth, the evaluator, and the paper’s reported score sit outside, on a scoring service it cannot reach directly. Withheld from the package is the paper’s own method, together with anything identifying which paper the task came from, so the agent can neither look the answer up nor reimplement the authors’ approach. Working from the data alone, it develops a solution in a private workspace and submits iteratively, querying the service for a score and for the time it has left.
Each run is given four hours of wall clock and one GPU matched to the task’s compute requirement, with web search disabled; the clock pauses during scoring, so evaluation does not consume the budget. Because the 90 papers report different quantities, from AUROC to RMSE to Spearman correlation, every submission is reduced to one comparable number: the relative gap g between the agent’s score and the paper’s, expressed as a fraction of the paper’s own result and signed so that positive is always better. Reaching zero counts as Match-SOTA and clearing +0.1 counts as Surpass-SOTA, while a task with no valid submission scores −1.0. Once the run ends, an automatic judge screens the submission for shortcuts such as fabricated outputs or gaming the scorer, and discards any run it flags.
Results across all 90 tasks
Figure 1. AIBuildAI-Science against the twelve agent configurations on the NatureBench leaderboard, over all 90 tasks. Baseline values are the published leaderboard numbers [1]; a task with no valid score counts against the agent, following the benchmark’s own convention.
AIBuildAI-Science surpasses the published state of the art on 25.6% of the 90 tasks and matches or exceeds it on 57.8%, both the best recorded on the benchmark. The strongest baseline agent, Claude Code, which is powered by Claude Opus 4.7, reaches 15.6% and 41.1%, so the margins are 10.0 and 16.7 percentage points, or nine more papers beaten and fifteen more reached. The median relative gap tells the same story from another angle: AIBuildAI-Science’s median across all 90 tasks is +0.009, against −0.030 for Claude Code and lower for every other agent, so the typical task now lands just above the number a Nature-family paper published rather than below it.
Results by scientific domain
Figure 2. Surpass-SOTA and Match-SOTA by scientific domain, with the number of tasks in each domain below the axis. Claude Code is the strongest baseline agent overall.
The gains are spread across fields rather than concentrated in one. AIBuildAI-Science matches the published result more often than the best baseline agent in all six domains and surpasses it more often in five. Physical modeling is strongest in absolute terms, at 46% surpass and 62% match against 31% and 46%. Cellular omics, the largest domain at 31 tasks, reaches 32% and 58% against 19% and 45%. The sharpest relative change is in biomedical modeling, the hardest domain for every system evaluated: the surpass rate is unchanged at one task in fourteen, but the match rate triples, from 14% to 43%. Relational reasoning is matched outright, five tasks from five.
Two solutions in detail
Figure 3. Two tasks where AIBuildAI-Science scores higher than every baseline agent, with all twelve agents’ scores on the same task shown for comparison.
We take two tasks as examples of the methods AIBuildAI-Science proposes: enzyme optimum pH prediction and anomalous-diffusion out-of-distribution detection. On both it surpasses the published state of the art and scores higher than every baseline agent on the leaderboard.
Enzyme optimum pH prediction asks for the pH at which an enzyme works best from its sequence, scored with weight on the acidic and alkaline extremes. Roughly 88% of known enzymes sit in the neutral band, so a standard regressor drifts there and flattens the very tails being measured. Instead of searching for a better encoder, AIBuildAI-Science targeted that imbalance directly (Figure 4a): four frozen protein language models from different pretraining families supply decorrelated features, a neural head trained under a balanced loss divides the neutral mass out of the objective, and a monotone calibration fitted on held-out folds stretches each bank’s compressed tails back before blending. Its g = +0.110 is the highest on the task and the only score clearing the surpass threshold; five of twelve baselines reach the published number, the best of them, Claude Code, stopping at +0.090.
The second task asks a model to flag particle trajectories produced by a physical regime absent from its training data. The novelty here is in how the problem is posed (Figure 4b). Rather than train an in-distribution model and threshold its confidence, AIBuildAI-Science exploits a property of the task itself: the unlabeled test pool is a mixture of in-distribution and out-of-distribution trajectories in a proportion the task discloses. Separating that pool from freshly simulated in-distribution trajectories with a positive-unlabeled classifier therefore estimates the density ratio an ideal detector would compute, while five-fold cross-fitting keeps any trajectory from contributing to its own score. It reaches +0.687, against +0.478 for Codex, the strongest baseline on this task.
Figure 4. The two solutions AIBuildAI-Science designed, each discovered autonomously within the four-hour budget. a, For enzyme optimum pH, four decorrelated protein language models feed a boosted-tree and a balanced-loss neural head, whose tails are recalibrated before blending. b, For anomalous-diffusion detection, simulated in-distribution trajectories and the unlabeled test pool are separated by a positive-unlabeled classifier over a shared physics-aware representation, yielding a score proportional to the density ratio an ideal detector would compute.
References
[1] Wang, Y., Cheng, L., Zuo, Y. et al. NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers? Horizon Research, Frontis.AI and Tsinghua University (2026).