ScienceFlow Adds Executable Workspaces to Agent Memory, Reaching a 70.22% Any-Medal Rate on 75 MLE-bench Tasks

Huawei Noah’s Ark Lab has open-sourced [ScienceFlow](https://github.com/huawei-noah/noah-research/tree/master/ScienceFlow), a long-horizon research agent designed to let machine learning experiments resume from a validated state after running for hours or even days, rather than relying on an ever-growing chat context to preserve progress. Each research worker in the system operates in an isolated, executable workspace. After the Evaluator produces normalized evidence and the Stage Gate accepts the result, the framework packages the code, artifacts, compact memory, validation records, and resource usage into an immutable Stage. The ESTRA mechanism can then select either the current Stage or an earlier one and decide whether to continue along the existing direction or re-anchor from that state and pursue a different research path.

This design addresses a common problem in long-horizon agents: the text remembers, but the environment cannot be restored. The memory layer retains recent results, the best validation evidence, and the information required for anchoring while folding older records. The execution controller manages CPU/GPU leases, timeboxes, and stopping conditions. The framework also places artifact formats, scoring direction, and the evaluator inside a task package, preventing the research agent from declaring success on its own.

The [paper](https://arxiv.org/abs/2608.14354) evaluates ScienceFlow on the full 75-task MLE-bench using DeepSeek-V4-Flash-Preview, with up to two GPUs and a 24-hour budget per task. Across three runs, it achieved an Any-Medal rate of 70.22±1.18%, 4.92 percentage points above the strongest reported baseline in the comparison table. Its Any-Medal rate was 74.56% in the Medium category but only 44.44% in the High category. Engineering teams should consider the storage cost of Stage snapshots, whether external data and services can truly be replayed, and whether restoring an old workspace could carry forward outdated prompts or data. The authors also explicitly warn against restoring old workspaces across different task profiles. These figures do not represent perfectly resource-matched experiments: some baselines used only 12 hours and different models, while Any-Medal is merely a coarse-grained measure of whether a medal threshold was crossed. The results therefore do not yet support the conclusion that ScienceFlow is superior across the board in either cost or final model quality.