Structured Scaling of AI Discovery Across Diverse Scientific Domains

arxiv:2604.19341

Original source · Paper JSON

Harness · 2026-04-21

Cites (12) · Cited by (2)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4
5. Part 5
6. Part 6
7. Part 7
8. Part 8
9. Part 9
10. Part 10
11. Part 11
12. Part 12

Contents

• Introduction
• Overview of SimpleTES
• SimpleTES discovers effective algorithms to reduce quantum compilation overhead
• SimpleTES improves optimizers for fuel-efficient deep-space trajectories
• SimpleTES discovers better scientific algorithms
• SimpleTES discovers a geometry-aware lasso solver that surpasses glmnet .
• SimpleTES discovers new state-of-the-art predictor for Zebrafish whole-brain neural-activity.
• SimpleTES improves AI foundations
• GPU kernel optimization.
• Scaling-law discovery.
• Post-training the LLM for SimpleTES leads to new mathematical discoveries
• Discussion
• Code availability
• Supplementary Information
• Contents
• A Authors
• B Related Work
• B.1 Existing Evaluation-Driven Discovery Methods
• AlphaEvolve [ 100 ] .
• OpenEvolve [ 6 ] .
• ShinkaEvolve [ 73 ] .
• ThetaEvolve [ 155 ] .
• TTT-Discover [ 165 ] .
• AdaEvolve [ 23 ] .
• EvoX [ 82 ] .
• GEPA [ 2 ] .
• Summary.
• B.2 Self-evolving AI
• Reward-based Self-Evolution.
• Imitation & Demonstration Learning.
• Population-based & Evolutionary Methods.
• B.3 LLM for Scientific Discovery
• LLM based scientific discovery systems.
• Scientific benchmarks for LLMs.
• Important results proposed by LLM systems.
• B.4 Test-Time Scaling
• Parallel Scaling.
• Sequential Scaling.
• Hybrid Scaling.
• Internal Scaling.
• C Theoretical Modeling of SimpleTES
• Definition .2 (Multidimensional Problem Refinement Trajectory) .
• Number of global independent trials C C .
• Theorem .3 .
• Proof.
• Remark.
• Local sample size K K .
• D Method Analysis
• D.1 Experiments on the Scaling Behavior of SimpleTES
• Scalability of global width ( C C ) and refinement depth ( L L ).
• Scalability of local sample size ( K K ) and refinement depth ( L L ).
• Scalability of the total evaluation budget ( N N ).
• D.2 Experimental Details on Post-Training
• Training improves the trajectory scores on the training tasks.
• Training enables generalization to the held-out tasks.
• Training unlocks stronger solutions that vanilla models fail to uncover.
• D.3 From Golden Metrics to Surrogate: Hacking Analysis
• Hacking in math construction tasks.
• Hacking in kernel optimization tasks.
• Discussion.
• E Variants of Proposal Constructor Φ \Phi
• Balance policy.
• LLM-Elite policy.
• F Ablation Studies
• F.1 Ablations on Different Designs of Φ \Phi
• F.2 Ablation on Reflection and Failure Patterns
• F.3 Efficiency Analysis: Trajectory-level Pruning
• G Wall-clock Time Efficiency
• H Quantum Compilation Tasks
• H.1 Problem definitions
• Superconducting qubit routing.
• Zoned neutral-atom compilation.
• H.2 Experimental scaffolds and evaluation setup
• Superconducting routing setup.
• Zoned neutral-atom setup.
• H.3 Discovered programs
• Discovered superconducting algorithm.
• Discovered zoned neutral-atom algorithm.
• H.4 Discovered superconducting routing policy using Gemini
• I Astrodynamics trajectory design
• I.1 Task formulation and program interface
• I.2 Independent trajectory evaluator
• I.3 Benchmark instances
• I.4 Construction of historical-sequence references
• I.5 How evolution changes mission-design search
• J GPU Kernel Optimization
• Task definitions and benchmark settings.
• TriMul evaluation alignment.
• Cross-device performance generalization.
• K Scaling Law Discovery
• Overview.
• Experimental setting.
• Results analysis.
• Case study.
• L Whole-brain neural-activity forecast on ZAPBench
• Problem definition and benchmark design.
• Forecasting task and data splits.
• Evaluation and search protocol.
• M Single-Cell RNA-Seq Denoising
• N Lasso Regularization Path
• Overview.
• Experimental setting.
• Results analysis.
• Case study.
• O Mathematical Scientific Discovery Tasks
• O.1 Extremal analysis
• Erdős minimum-overlap problem.
• Autocorrelation Inequalities.
• O.2 Combinatorial construction
• Sum–Difference Problem.
• Circle Packing in a Unit Square.
• Hadamard maximum-determinant problem of order 29 29 .
• Instructions for reporting errors