Organizations: National Laboratory of Solid State Microstructures, School of Physics and, Collaborative Innovation Center of Advanced Microstructures,, Nanjing University, Nanjing 210093, China · Shanghai Innovation Institute, Shanghai 200231, China · College of Computer Science and Artificial Intelligence,, Fudan University, Shanghai 200433, China · BEDI Cloud, Beijing, China · Shanghai Research Institute for Intelligent Autonomous Systems,, Tongji University, Shanghai 201210, China
Scientific discovery often requires identifying a small number of high-performing designs in vast, constrained search spaces. This is difficult when evaluations are costly, feasible solutions are sparse, parameters are strongly coupled, and physical constraints are strict. Recent AI for AI methods can automatically improve algorithms and optimization procedures. However, general purpose approaches are not designed around scientific priors, physical constraints, and limited evaluation budgets. We introduce ADAPT for Science, a multi-agent framework that treats scientific optimization as two coupled decisions: where to search and how to search. ADAPT uses historical designs to identify promising and diverse initial conditions, progressively concentrates computation on favorable search directions, and evolves the optimization procedure from execution feedback. We evaluate ADAPT on the design of gravitational wave detectors using a differentiable physics simulator and eight unseen detector topologies. Under matched scientific evaluation budgets, ADAPT achieves a mean loss of 0.204 +/- 0.101, reducing loss by 54.9% relative to the best classical optimizer and by 62.2% relative to the best general purpose agentic evolution baseline. ADAPT achieves the lowest loss on all unseen topologies among the evaluated baselines. ADAPT provides a step toward scientific optimization systems that continuously learn from prior experience and their own search outcomes.
Figures & tables
Figure 1: Overview of ADAPT for Science. A target-aware matcher retrieves historical designs and reference optimizers to form a target-specific search prior. The selection evolver determines where to search , while the optimizer evolver determines how to search from execution feedback. Together, they define a target-specific search policy that is evaluated on held-out targets under the original scientific objective and physical constraints.
Figure 2: Evolution of the selection policy. Mean loss across successive selection-policy iterations. Starting from direct selection after 64 forward evaluations, evolution introduces quality–diversity shortlisting, staged probing, and progressive budget reallocation. The final policy reaches a mean loss of 3.0054 .
Figure 3: Selection evolution discovers progressive search. Starting from a 64 -to- 1 objective, the evolved policy selects four high-quality feasible candidates and four diversity-preserving candidates, then progressively concentrates evaluation as 8→2→1 while retaining protected exploration.
Figure 4: Optimizer evolution discovers budget-aware search dynamics. The evolved procedure retains the Adam structure while using the selected initialization, shorter first-moment memory, time-dependent updates, and explicit rejection of invalid evaluations and proposals under a fixed scientific budget.
Figure 5: Evolution of the optimization procedure. Mean loss across successive optimizer-evolution iterations, with annotations summarizing the changes tested at each stage. Starting from the reference AdamGD implementation, evolution tests numerical safeguards, stopping behavior, update scaling, and optimizer memory. The final procedure reaches a mean loss of 0.6534 during strategy evolution.
Method
T1
T2
T3
T4
T5
T6
T7
T8
Mean ± Std.
Adam
0.423 +36.0%
0.887 +117.4%
1.833 +2344.0%
0.332 +45.6%
0.379 +152.7%
0.692 +397.8%
0.993 +509.2%
0.666 +329.7%
0.776±0.458 +280.4%
NAAdam
0.423 +36.0%
0.696 +70.6%
0.423 +464.0%
0.334 +46.5%
0.415 +176.7%
0.375 +169.8%
0.695 +326.4%
0.255 +64.5%
0.452±0.150 +121.6%
Lion
4.358 +1301.3%
3.841 +841.4%
2.028 +2604.0%
3.028 +1228.1%
2.004 +1236.0%
2.641 +1800.0%
2.663 +1533.7%
2.325 +1400.0%
2.861±0.793 +1302.5%
SGDM
4.151 +1234.7%
0.445 +9.1%
0.468 +524.0%
0.473 +107.5%
0.581 +287.3%
0.448 +222.3%
0.196 +20.2%
0.415 +167.7%
0.897±1.234 +339.7%
Sophia
0.506 +62.7%
0.412 +1.0%
0.627 +736.0%
1.749 +667.1%
0.653 +335.3%
0.574 +312.9%
0.430 +163.8%
1.659 +970.3%
0.826±0.514 +304.9%
Adan
1.404 +351.4%
1.321 +223.8%
1.376 +1734.7%
2.783 +1120.6%
2.212 +1374.7%
0.673 +384.2%
1.355 +731.3%
1.387 +794.8%
1.564±0.602 +666.7%
Table 1: Optimization performance of classical optimizers across eight held-out gravitational-wave detector targets. Lower loss is better. Red percentages indicate the relative increase in loss compared with Ours. The final column reports mean ± standard deviation.
Figure 6: Representative held-out detector optimization. (a) Initial detector configuration. (b) Design obtained with standard Adam and its optimized carrier-power distribution. (c) Final ADAPT design. ADAPT suppresses three optical branches with near-zero optimized contributions and replaces them with vacuum inputs while preserving the remaining detector structure. (d) Loss convergence under the same scientific evaluation budget. ADAPT reaches a final reported loss of 0.155 .
Target
Direct+GEPA
Fixed+Adam
No-Prompt
Ours
T1
0.416 +33.8%
0.320 +2.9%
0.424 +36.3%
0.311
T2
0.986 +141.7%
0.371 -9.1%
0.443 +8.6%
0.408
T3
0.373 +397.3%
0.180 +140.0%
0.145 +93.3%
0.075
T4
0.476 +108.8%
0.246 +7.9%
0.287 +25.9%
0.228
T5
0.164 +9.3%
0.200 +33.3%
0.264 +76.0%
0.150
T6
0.138 -0.7%
0.143 +2.9%
0.148 +6.5%
0.139
Table 3: Ablation study. Mean ± std. over eight held-out topologies (lower is better).
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Range
Unit
Reflectivity
[10−12,1−10−12]
dimensionless
Mirror tuning
[−360,360]
degree
Squeezing level
[0,10]
dB
Squeezing angle
[−360,360]
degree
Laser power
[0,200]
W
Mass
[0.01,200]
kg
Appendix
Table 4: Continuous detector-design variables and default physical bounds. The active parameter dimension dτ depends on the detector topology. Some physical properties may share one optimization scalar and are therefore optimized jointly.
Figure 7: Selection Evolver prompt (page 1/3). The prompt defines the frozen scientific contract, the 64 -to- 1 selection objective, development and held-out evaluation regimes, and the required selector interface.
Figure 7 , continued (page 2/3).
Figure 7 , continued (page 3/3).
Figure 8: Optimizer Evolver prompt (page 1/3). The prompt specifies the composite development objective, held-out evaluation protocol, optimizer interface, and permitted execution environment.
Figure 8 , continued (page 2/3).
Figure 8 , continued (page 3/3).
Stage
Hardware
Budget
Used for evolution?
Selection development
RTX 4090
up to 2 h / target
Yes
Optimizer development
Development hardware
fixed evaluator budget
Yes
Held-out validation
NVIDIA H100
4 h / target
No
Appendix
Table 5: Development and held-out evaluation regimes. Development feedback is obtained only from the development panel. The H100 evaluation is performed only after the final algorithms have been frozen.
Method
T1
T2
T3
T4
T5
T6
T7
T8
Mean ± Std.
Adam
0.423
0.887
1.833
0.332
0.379
0.692
0.993
0.666
0.776±0.458
NAAdam
0.423
0.696
0.423
0.334
0.415
0.375
0.695
0.255
0.452±0.150
Lion
4.358
3.841
2.028
3.028
2.004
2.641
2.663
2.325
2.861±0.793
SGDM
4.151
0.445
0.468
0.473
0.581
0.448
0.196
0.415
0.897±1.234
Sophia
0.506
0.412
0.627
1.749
0.653
0.574
0.430
1.659
0.826±0.514
Adan
1.404
1.321
1.376
2.783
2.212
0.673
1.355
1.387
1.564±0.602
Appendix
Table 6: Complete results for classical optimization baselines. All entries are the best reported held-out loss under the matched scientific evaluation budget. Lower is better.
Target
GEPA
OpenEvolve
ShinkaEvolve
ASI
ADAPT
T1
0.557
0.707
0.561
0.725
0.311
T2
0.410
0.492
0.440
0.483
0.408
T3
0.480
0.449
0.674
0.624
0.075
T4
1.218
0.261
2.632
0.653
0.228
T5
0.379
0.665
0.366
1.309
0.150
T6
0.435
0.426
0.210
0.450
0.139
Appendix
Table 7: Complete results for general-purpose agentic evolution baselines. Lower is better.
Target
Direct+GEPA
Fixed+Adam
No-Prompt
ADAPT
T1
0.416
0.320
0.424
0.311
T2
0.986
0.371
0.443
0.408
T3
0.373
0.180
0.145
0.075
T4
0.476
0.246
0.287
0.228
T5
0.164
0.200
0.264
0.150
T6
0.138
0.143
0.148
0.139
Appendix
Table 8: Complete ablation results. Lower is better. Mean ± standard deviation is computed across the eight held-out topologies.
Target
Internal topology
Boundary topology
T1
DDDBBBDDC
LSLLSSSLSHSS
T2
EAFBHDGGG
LLSSSSDLSLSS
T3
FADGHCFEC
LLLLLLLLLLLH
T4
FGHHBGEFE
DLLLLLLLLLLS
T5
HAFHBEBFB
SLLLDLSLSLLL
T6
HHCDHEHBF
SDSSSSSLSSSS
Appendix
Table 9: Eight held-out detector topologies. The first string specifies the 3×3 internal optical layout and the second string specifies the boundary components.
Figure 9: Complete comparison across the eight held-out detector topologies. Each row corresponds to one target. Column (a) shows the initial detector configuration, column (b) shows the optimized standard-Adam baseline, and column (c) shows the final ADAPT design. Carrier power is represented by line width on a logarithmic scale. Components with effectively vanishing optimized contribution are shown as vacuum inputs where applicable.
AI agents for autonomous scientific discovery must decide which experiments to run under a limited budget. Most benchmarks for science agents provide fixed datasets, while others that support active experimentation largely rely on simulation. We introduce Autonomous Science Bench, an initial release comprising six budgeted design-build-test-learn campaigns grounded in published experiments across biology, chemistry, and materials science. Agents select batches of designs, receive measurements from a lab service, and use the results to guide subsequent experiments before delivering a predictive model. Success requires passing two calibrated gates: a discovery gate, which evaluates the designs the agent measured, and a learning gate, which evaluates its model's ranking of held-out designs. In a pilot evaluation of eight agent-model configurations and 144 trials, the best configuration succeeds in 67% of trials, while the hardest campaign is solved in only three of 24 attempts. The two gates reveal different weaknesses: some agents build accurate rankings but find too few promising designs, while others find promising designs but fail to predict which unseen designs will perform well. These differences are obscured by overall success rates alone. We release version 0.1 of the campaigns, lab runtime, and public leaderboard at https://yibow.me/autonomous-science-bench.
The increasing adoption of Large Language Models (LLMs) has enabled AI scientists to perform complex end-to-end scientific discovery tasks requiring coordination of specialized roles, including idea generation and experimental execution. However, most state-of-the-art AI scientist systems rely on static, hand-designed pipelines and fail to adapt based on accumulated interaction histories. As a result, these systems overlook promising research directions, repeat failed experiments, and pursue infeasible ideas. To address this, we introduce EvoScientist, an evolving multi-agent AI scientist framework that continuously improves research strategies through persistent memory and self-evolution. EvoScientist comprises three specialized agents: a Researcher Agent (RA) for scientific idea generation, an Engineer Agent (EA) for experiment implementation and execution, and an Evolution Manager Agent (EMA) that distills insights from prior interactions into reusable knowledge. EvoScientist contains two persistent memory modules: (i) an ideation memory, which summarizes feasible research directions from top-ranked ideas while recording previously unsuccessful directions; and (ii) an experimentation memory, which captures effective data processing and model training strategies derived from code search trajectories and best-performing implementations. These modules enable the RA and EA to retrieve relevant prior strategies, improving idea quality and code execution success rates over time. Experiments show that EvoScientist outperforms 7 open-source and commercial state-of-the-art systems in scientific idea generation, achieving higher novelty, feasibility, relevance, and clarity via automatic and human evaluation. EvoScientist also substantially improves code execution success rates through multi-agent evolution, demonstrating persistent memory's effectiveness for end-to-end scientific discovery.
Scientific research proceeds through iterative cycles of hypothesis generation, experiment design, execution, and revision. AI agents can automate parts of this process, but existing approaches typically follow a single research trajectory or coordinate through a central planner with fixed objectives. As a result, they struggle to sustain parallel exploration, adapt as experimental evidence changes, or preserve knowledge of failed directions over long-running experiments. We introduce AutoScientists, a decentralized team of AI agents for long-running computational scientific experimentation. Agents interpret a shared experimental state, self-organize into teams around promising hypotheses, critique proposals before using experimental compute, and share successes and failures to reduce redundant exploration. Under matched experimental budgets, AutoScientists improves over prior AI agents across biomedical machine learning, language-model training optimization, and protein fitness prediction. On BioML-Bench, spanning biomedical imaging, protein engineering, single-cell omics, and drug discovery, AutoScientists achieves a mean leaderboard percentile of 74.4% across 24 tasks, improving over the strongest AI agent by +8.33%. On GPT training optimization, AutoScientists reaches a target validation bits-per-byte 1.9x faster than Autoresearch and continues discovering improvements from a starting champion where the single-agent approach finds none (7 vs. 0 accepted improvements). On ProteinGym fitness prediction, AutoScientists discovers a method for ACE2-Spike binding that improves over the current state-of-the-art model by +12.5% in Spearman correlation. Applied without modification across all 217 ProteinGym assays, the same method improves over the prior state of the art by +6.5% (Spearman correlation).