Recent breakthroughs in artificial-intelligence-assisted mathematics show that open-ended machine search can outperform human mathematicians, yet transferring this paradigm to physics requires refactoring the research workflow. We introduce Simjecture, an open-source framework in which language-model agents pursue human-defined physical conjectures. The framework tackles two central challenges in this transfer: first, because physics is empirical and no model is final, it structures inquiry around counterexample-driven falsification and prospective hypothesis repair; second, because physical theories must interact with the real world, it uses computational physics as an executable testbed while separating the qualification of numerical instruments from scientific claim review so solvers and diagnostics can be adapted within a recorded investigation. We apply the framework in a 136.3 h investigation to tackle the long-standing problem of radiated energy exceeding the imploding-shell kinetic energy in a Z-pinch. The investigation reveals competing effects of energy conversion, retention, and release timing. Stronger coherent helical deformation increases resistive conversion while reducing radiation and increasing retained energy. In a matched fine-grid test, broadening the trailing-plasma profile delays the later radiation maximum by 1.80 ns and recovers an early 153 J bolometric deficit to within 0.08% by 44 ns, with a spectral response that shifts during the pulse. These results provide new insight into energy conversion in high-energy-density plasma physics and demonstrate a new paradigm for physics research.
Figures & tables
Figure 1: Coupled scientific (teal) and instrument (orange) workflows in Simjecture .
Figure 3: Investigation progression through (a) density-seed screening, (b) single-helix tests, (c) multimode comparisons, and (d) timing and numerical controls. Bars show recorded job intervals in accumulated campaign hours; horizontal scales differ between panels. Right-hand panels show x – z density slices near 32 ns on a shared scale. RMS denotes the mass-weighted displacement of axial slice centroids from the original axis.
Figure 4: Plasma evolution in the Stage IV realization of 3D-128-M4/4 . Within each block, columns show (a) three-dimensional density surfaces at 0.1 and 1kg/m3 , with the inner surface colored by electron temperature using the scale in (b); (b) electron temperature; (c) magnetic-field magnitude; (d) Ohmic power density; and (e) stored radiation energy density. Field maps are longitudinal cuts at y=0 , interpolated between the two bracketing cell-center planes. Temperature is shown where ρ≥0.01kg/m3 .
Figure 5: Energy partition of the aluminum load and fields over 0 – 32 ns in 3D-128-M4/4 (IV) . Bands show magnetic energy, load kinetic and thermal energies, stored radiation, cumulative emitted radiation, and remaining inferred net input (gray). Thermal energy includes electrons and ions; magnetic and radiation-field energies cover the whole domain. The horizontal top is the fixed budget through 32 ns; the lower dark curve traces stored load-and-field energy.
Figure 6: Time-resolved radiation spectrograms, power histories, and integrated spectra for the unperturbed and density-seeded references (a)–(c), the fine-mesh deformation set from Stages II–III (d)–(f), and the Stage IV fine-mesh trailing-width pair (g),(h). Each set uses its own matched reference. Shaded power regions and bold spectral segments mark emission above 1 keV; dashed curves show the indicated reference.
Figure 7: Response of the matched fine-grid trailing-width pair, 3D-128-M4/4 and 3D-128-M4/7 . (a) Signed axial-current fraction within 2 mm of the slice centroid. (b) Tagged load mass above 300 eV. (c) Cumulative radiation difference, wider minus narrower profile, for all photons and for >1 keV photons. Dotted lines mark 32 ns.
Role
Input
Cached input
Output
Worker
1,042.693
1,005.254
8.933
Director
169.151
129.747
1.365
Reviewer
119.934
102.835
0.916
Memory
0.347
0.197
0.016
Total
1,332.125
1,238.032
11.230
Table 1: Recorded language-model token usage in millions across all four Z-pinch stages. Cached input is included in total input; the reviewer row includes report assessment.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Study
Worker
Director and reviewers
Supplied starting material
Reconnection scaling
GPT-6.1 Sol, medium
GPT-6.1 Sol, high
FLASH resistive-MHD application, simulation driver, diagnostics, and validation examples.
Z-pinch, Stages I–III
DeepSeek Flash
GPT-6.1 Sol, high
Experimental load and question, FLASH radiation-MHD runtime, aluminum material tables, and commissioning examples.
Z-pinch, Stage IV
GPT-6.1 Terra, high
Director: GPT-6 Astra, high; reviewers: GPT-6.1 Sol, high
Existing FLASH workspace, Stage I–III trajectories, checkpoints, diagnostics and physical synthesis, and the Conjecture 1 coprocessor.
Appendix
Table 2: Agent configurations and supplied starting material across the two investigations, with the Z-pinch configuration shown by stage.
Case
Seed
D
σt
Mesh
Eγ
E>1keV
(mm)
(mm)
(kJ)
(J)
3D-64-L5
Density, 5%
0.0
0.4
642×32
3.834
34.8
3D-96-L0
Density, 0%
0.0
0.4
962×32
3.783
85.1
3D-96-L5
Density, 5%
0.0
0.4
962×32
3.786
82.7
3D-96-L10
Density, 10%
0.0
0.4
962×32
3.781
83.8
3D-128-S4/4
Helix
0.4
0.4
1282×64
3.678
449.4
Appendix
Table 3: Principal three-dimensional Z-pinch trajectories over 0 – 32 ns, where D is the initial centerline displacement amplitude and σt is the trailing-annulus width.
Case
E0
EB
K
Ug
Uγ
Eγ
B
3D-128-S4/4
4.298
0.791
0.607
0.593
0.002
3.678
1.374
3D-128-S8/4
4.293
0.887
0.581
0.706
0.002
3.482
1.364
3D-128-M4/4 (III)
4.298
0.786
0.555
0.593
0.002
3.718
1.357
3D-64-S4/4
4.290
0.665
0.550
0.420
0.003
3.933
1.281
3D-64-M4/4
4.290
0.654
0.543
0.419
0.003
3.926
1.256
3D-64-M4/10
4.286
0.609
0.608
0.393
0.004
3.768
1.096
Appendix
Table 4: Whole-domain initial stored energy E0 , 32 ns stored components EB , K , Ug=Ue+Ui , and Uγ , cumulative escaping radiation Eγ , and net boundary-and-residual balance B=EB+K+Ug+Uγ+Eγ−E0 , in kJ.
Interval
Eγ (kJ)
fhard (%)
(ns)
Narrow
Broad
Narrow
Broad
Stage III: 0.4/1.0 mm, 642×32
0–8
1.012
0.985
17.81
17.71
8–16
0.983
0.946
7.50
3.88
16–24
1.203
0.957
6.31
6.47
24–32
0.728
0.880
4.97
6.79
Appendix
Table 5: Interval radiation yield Eγ and hard-band fraction fhard=E>1keV/Eγ for the Stage III and Stage IV trailing-width pairs.
Figure 8: Energy accounting for 3D-128-M4/4 (IV) . (a) Whole-domain budget over 0 – 32 ns, including the numerical background. (b) Separate kinetic and thermal energies of the aluminum load (solid) and numerical background (dashed) through 44 ns. Thermal energy includes electrons and ions.
Autonomous research agents use language models and tools to search literature, design and run experiments, analyze results, and support scientific discovery.
Autonomous-research agents have demonstrated end-to-end LLM automation in machine-learning sandboxes where execution provides calibration. Frontier physical science differs categorically: physical reasoning underlies every methodology choice, toolchains are often underdocumented, and calibration must come from external literature anchors - which unscaffolded agents cite but do not confront, hallucinating plausible, unverifiable results from internal priors. We present a pipeline that runs end-to-end from a corpus of 11,083 recent condensed-matter physics arXiv papers to a publication-grade manuscript with three substantive physics findings (here on altermagnetic piezomagnetism): the agent autonomously conceives a research direction by mapping the corpus, calibrates methodology by reproducing published references, conducts novel first-principles computations, and writes the manuscript - grounded in literature throughout, across 47 fresh-context sessions in six phases sharing only on-disk state, with 2,162 literature-consultation events. Fault tolerance emerges from redundancy: fresh-context isolation, distributed grounding, and adversarial review catch what any single session misses; pre- and post-pilot stages are fully autonomous, and pilot requires bounded human intervention only at reproduction failures - operational knowledge curation, not scientific direction. Two paired failure modes - a pre-architecture baseline and a no-pilot ablation - isolate structurally enforced numerical confrontation at calibration checkpoints as the operative grounding mechanism. The primitives, characterized failure modes, and quantified intervention pattern lay a foundation for autonomous research in high-stakes scientific domains beyond computational physics.
Haonan Huang
Department of Physics, Princeton University, Princeton, NJ, USA.
AI agents are playing an increasingly important role in physics research. However, existing physics benchmarks only evaluate fixed questions, isolated sub-tasks, or paper reproduction with given solutions, and rarely check research quality alongside correctness. We introduce FrontierPhysics, an end-to-end benchmark with 56 tasks made by 42 physicists based on their own research, covering nine domains across theoretical, computational, and experimental physics. Each task requires agents to review the literature, develop a research plan, conduct research and experiments, and produce structured outputs and scientific manuscripts. Combining deterministic verifiers with expert-authored rubrics, FrontierPhysics evaluates whether agents' research is correct and publishable. Across 22 models in 26 configurations, no configuration passes more than 46 of its 168 rollouts. Manual review of agent trajectories shows that frontier agents' main weaknesses are insufficiently rigorous reasoning and limited expert intuition, rather than gaps in physics knowledge or execution errors.
AI agents powered by large language models exhibit strong reasoning and problem-solving capabilities, enabling them to assist scientific research tasks such as formula derivation and code generation. However, whether these agents can reliably perform end-to-end reproduction from real scientific papers remains an open question. We introduce PRBench, a benchmark of 30 expert-curated tasks spanning 11 subfields of physics. Each task requires an agent to comprehend the methodology of a published paper, implement the corresponding algorithms from scratch, and produce quantitative results matching the original publication. Agents are provided only with the task instruction and paper content, and operate in a sandboxed execution environment. All tasks are contributed by domain experts from over 20 research groups at the School of Physics, Peking University, each grounded in a real published paper and validated through end-to-end reproduction with verified ground-truth results and detailed scoring rubrics. Using an agentified assessment pipeline, we evaluate a set of coding agents on PRBench and analyze their capabilities across key dimensions of scientific reasoning and execution. The best-performing agent, OpenAI Codex powered by GPT-5.3-Codex, achieves a mean overall score of 34%. All agents exhibit a zero end-to-end callback success rate, with particularly poor performance in data accuracy and code correctness. We further identify systematic failure modes, including errors in formula implementation, inability to debug numerical simulations, and fabrication of output data. Overall, PRBench provides a rigorous benchmark for evaluating progress toward autonomous scientific research.