GECCO 2026 · SpOC4 optimization challenge
Ten-day validation-guided campaign · third place, 393.229.
Open-source research project
ScienceFlow: A Long-Horizon Agent for ML Research, Scientific Discovery, and Beyond
ScienceFlow Team · Noah's Ark Lab, Huawei
Research Highlights
MLE-bench · Machine learning engineering
Developing complete solutions for diverse machine learning competitions.
70.22%
Overall · 24 h
MLE-bench
Any-Medal rate (%) · completed 2026-07-25 · higher is better.
| Agent | LLM(s) used | Lite (%) | Medium (%) | High (%) | All (%) | Running Time (hours) | Code & Reports Available |
|---|---|---|---|---|---|---|---|
| ScienceFlow | DeepSeek-V4-Flash-Preview | 80.30± 1.52 | 74.56± 0.88 | 44.44± 2.22 | 70.22± 1.18 | 24 | Open source |
| MLEvolve | Gemini-3.1-Pro-Preview | 80.30± 1.50 | 64.00± 0.90 | 46.70± 0.00 | 65.30± 0.80 | 12 | Code · Paper |
| Iris | Claude-Opus-4.6 | 80.30± 1.50 | 64.00± 0.90 | 44.40± 2.20 | 64.90± 0.40 | 12 | Paper |
| Famou-Agent 2.0 | Gemini-3-Pro-Preview | 80.30± 1.52 | 64.04± 2.32 | 42.22± 2.22 | 64.44± 1.18 | 24 | Project code · Paper |
| AIBuildAI | Claude-Opus-4.6 | 77.27± 0.00 | 61.40± 0.88 | 46.67± 0.00 | 63.11± 0.44 | 24 | Project software · Paper |
| CAIR MARS+ | Gemini-3-Pro-Preview | 78.79± 1.52 | 60.53± 1.52 | 44.44± 2.22 | 62.67± 0.77 | 24 | Project code · Paper |
Math / Optimization · Fourier uncertainty
Constructing a function and its Fourier transform with smaller sign-change radii.
0.343293122432
C4 upper bound
Uncertainty task
AlphaEvolve B.4 Hermite · completed 2026-07-07 · lower is better.
| Work | Model | Best C4 upper bound | Evidence / verification |
|---|---|---|---|
| ScienceFlow | openPangu-2.0-Pro | 0.343293122432 | Verified · public verifier |
| FM Agent | Gemini-2.5-Pro | 0.352099104416 | Reported · paper |
| LoongFlow | DeepSeek-R1 | 0.352099104422 | Reported · project |
| AlphaEvolve | Gemini Pro + Flash | 0.352099104423 | Verified · public verifier |
| MLEvolve | Gemini-3.1-Pro-Preview | 0.352099104416 | Reported · project |
SciModelingBench · Scientific design optimization
Optimizing candidate designs across 12 scientific discovery tasks.
54.41
Group-balanced score
SciModelingBench
7 groups · 12 tasks · completed 2026-07-22 · higher is better.
| Task | ScienceFlow | Claude Code | Codex | OpenCode | Pi | Random |
|---|---|---|---|---|---|---|
| Standalone scientific design · 6 task groups | ||||||
| TFBind8 | 0.9769 | 0.9648 | 0.9718 | 0.9764 | 0.9730 | 0.7533 |
| Superconductor | 0.8583 | 0.8132 | 0.8188 | 0.4929 | 0.8671 | 0.2800 |
| UTR MRL | 0.7605 | 0.7367 | 0.7424 | 0.6523 | 0.7533 | -0.0006 |
| TFBind10 Pho4 | 0.2543 | 0.3197 | 0.2024 | 0.2467 | 0.2473 | 0.0000 |
| GFP | 0.2732 | 0.1977 | 0.2586 | 0.2600 | 0.2692 | 0.0001 |
| Hopper Controller | 0.3756 | 0.3718 | 0.3755 | 0.3917 | 0.3248 | 0.1698 |
| Rat clinical pathology (DrugMatrix) · 6 endpoints, one task group | ||||||
| MCHC | 0.7220 | 0.6503 | 0.6178 | 0.7344 | 0.3720 | 0.1733 |
| MCH | 0.4793 | 0.3623 | 0.4242 | 0.4147 | 0.3673 | 0.2669 |
| Creatinine | 0.6995 | 0.8038 | 0.5681 | 0.6081 | 0.7638 | 0.1955 |
| Sodium | 0.8488 | 0.8498 | 0.7696 | 0.7619 | 0.8283 | 0.1519 |
| Chloride | 0.5488 | 0.5397 | 0.5589 | 0.5106 | 0.4551 | 0.1689 |
| Phosphorus | 0.5985 | 0.3999 | 0.3749 | 0.4130 | 0.4917 | 0.2040 |
| Standalone Design (6) | 54.16 | 51.65 | 51.49 | 43.84 | 52.77 | 0.0000 |
| Rat Clinical Pathology (6) | 55.91 | 49.64 | 43.79 | 46.44 | 43.12 | 0.0000 |
| Group-balanced Score ↑ | 54.41 | 51.37 | 50.39 | 44.21 | 51.39 | 0.0000 |
| 7-group Avg. Rank ↓ | 1.55 | 3.69 | 3.64 | 3.33 | 2.79 | — |
| Task Wins | 5/12 | 3/12 | 1/12 | 2/12 | 1/12 | — |
DeepSeek-V4-Flash-Preview for all agents · Group-balanced scores map Random to 0 and the attainable optimum to 100 across 7 equally weighted groups.
GECCO 2026 · SpOC4 optimization challenge
Finding the fastest ΔV-constrained route to collect every target orbiting the Moon.
🥉
KTTSP-hard · 393.229 ↓
Keplerian Tomato Traveling Salesperson · KTTSP-hard
3,151 dimensions · 1 objective · submitted 2026-07-01.
| Rank | Team | Model | Score ↓ | Submitted (UTC+8) |
|---|---|---|---|---|
| 🥇 | TGMA ↗ | — | 337.646 | 2026-07-01 13:50 |
| 🥈 | AC_TUWien ↗ | — | 337.700 | 2026-07-01 01:05 |
| 🥉 | ScienceFlow | DeepSeek-V4-Pro-Preview / DeepSeek-V4-Flash-Preview + GLM-5.1 | 393.229 | 2026-07-01 09:33 |
| 4 | fcmaes ↗ | — | 436.015 | 2026-06-30 14:01 |
| 5 | Team HRI ↗ | — | 526.078 | 2026-06-30 21:17 |
Score: total mission time in days; lower is better. Models are shown only when publicly disclosed. Official challenge and leaderboard ↗
Long-horizon Tasks
Click any panel to enlarge.
MLE-bench · Machine learning engineering
Workspace, validation, assets, and resource use.
OpenPyRo-A1 · Long-horizon embodied manipulation
Four long-horizon manipulation tasks.
SciModelingBench · Scientific design optimization
DrugMatrix MCHC candidate discovery.
Code and technical report
ScienceFlow is open source — framework code, task packages, and documentation are available on GitHub, and the technical report is on arXiv.