Datasets:

prometheus-eval
/

into-the-unknown

paper string	group string	title string	domain string	venue string	citations int64	lab string	original_paper string	follow_up_direction string	task_spec string	paper_pdf string	paper_md string
P1	A	Hierarchical modeling of molecular energies using a deep neural network	Computational chemistry	J. Chem. Phys.	428	LANL	https://arxiv.org/abs/1710.00017	Use HIP-NN's hierarchy-breakdown signal to identify failure motifs and calibrate uncertainty estimates that transfer across conformations.	benchmark/paper1/task_spec.md	benchmark/paper1/paper.pdf	benchmark/paper1/paper.md
P2	A	Extending the accuracy of the SNAP interatomic potential form	Materials science	J. Chem. Phys.	311	SNL	https://arxiv.org/abs/1711.11131	Replace SNAP bispectrum components with a learned descriptor and test its accuracy, data efficiency, and cross-material transfer.	benchmark/paper2/task_spec.md	benchmark/paper2/paper.pdf	benchmark/paper2/paper.md
P3	A	Convolutional Dictionary Learning: A Comparative Review and New Algorithms	Computational imaging	IEEE TCI	282	LANL	https://arxiv.org/abs/1709.02893	Unify CDL dictionary-update algorithms within one multichannel derivation framework and use it to reveal a new method and equivalence.	benchmark/paper3/task_spec.md	benchmark/paper3/paper.pdf	benchmark/paper3/paper.md
P4	A	A Kriging-based approach to autonomous experimentation with applications to X-ray scattering	Scientific ML	Sci. Rep.	136	LBNL	https://doi.org/10.1038/s41598-019-48114-3	Develop an acquisition strategy that reconstructs synthetic functions faster than SMART and gSMART.	benchmark/paper4/task_spec.md	benchmark/paper4/paper.pdf	benchmark/paper4/paper.md
P5	A	Adaptive variational quantum dynamics simulations	Quantum	PRX Quantum	203	Ames	https://arxiv.org/abs/2011.00622	Make AVQDS compatible with fixed nearest-neighbor qubit topology while measuring the resulting accuracy and resource overhead.	benchmark/paper5/task_spec.md	benchmark/paper5/paper.pdf	benchmark/paper5/paper.md
P6	A	Simulations of relativistic-quantum plasmas using real-time lattice scalar QED	Plasma physics	Phys. Rev. E	32	PPPL	https://arxiv.org/abs/1802.00524	Stress-test variational lattice scalar QED in new valid and fundamentally invalid regimes, then diagnose and repair its failures.	benchmark/paper6/task_spec.md	benchmark/paper6/paper.pdf	benchmark/paper6/paper.md
P7	A	The non-Riemannian nature of perceptual color space	Color science	PNAS	51	LANL	https://doi.org/10.1073/pnas.2119753119	Test whether perceptual diminishing returns are organized by cone channels rather than global scaling or generic anisotropy.	benchmark/paper7/task_spec.md	benchmark/paper7/paper.pdf	benchmark/paper7/paper.md
P8	A	Parameterized Physics-informed Neural Networks for Parameterized PDEs	Scientific ML	ICML	113	null	https://arxiv.org/abs/2408.09446	Extend P²INN into a problem-specification encoder that generalizes jointly across coefficients, initial conditions, and boundary conditions.	benchmark/paper8/task_spec.md	benchmark/paper8/paper.pdf	benchmark/paper8/paper.md
P9	B	An efficient explicit implementation of a near-optimal quantum algorithm for simulating linear dissipative differential equations	Quantum	J. Comput. Phys.	6	LLNL	https://arxiv.org/abs/2501.11146	Improve the mathematical LCHS algorithm by substituting a kernel with better asymptotic resource scaling.	benchmark/paper9/task_spec.md	benchmark/paper9/paper.pdf	benchmark/paper9/paper.md
P10	B	Efficient Berry Phase Calculation via Adaptive Variational Quantum Computing Approach	Quantum	APL Quantum	5	Ames	https://arxiv.org/abs/2506.19150	Determine how much realistic hardware noise the AVQDS Berry-phase method can tolerate before its topological output fails.	benchmark/paper10/task_spec.md	benchmark/paper10/paper.pdf	benchmark/paper10/paper.md
P11	B	Projected Hessian Learning: Fast Curvature Supervision for Accurate Machine-Learning Interatomic Potentials	Computational chemistry	Unpublished	3	LANL	https://arxiv.org/abs/2603.04523	Reduce the stochastic-probe requirement of Projected Hessian Learning without sacrificing estimator or downstream model quality.	benchmark/paper11/task_spec.md	benchmark/paper11/paper.pdf	benchmark/paper11/paper.md
P12	B	gp2Scale: Compactly-Supported Non-Stationary Kernels and Distributed Computing for Exact Gaussian Processes on 10 Million Data Points	Scientific ML	ICML	2	LBNL	https://arxiv.org/abs/2512.06143	Apply exact large-scale Gaussian processes to a new domain using a domain-specific kernel or sparsity construction.	benchmark/paper12/task_spec.md	benchmark/paper12/paper.pdf	benchmark/paper12/paper.md
P13	B	Particle-in-cell simulations of laser crossbeam energy transfer via magnetized ion-acoustic wave	Plasma physics	Physics	2	LLNL	https://arxiv.org/abs/2511.03090	Map the boundary where Vlasov theory ceases to predict MagCBET PIC behavior and identify the first invalid assumption.	benchmark/paper13/task_spec.md	benchmark/paper13/paper.pdf	benchmark/paper13/paper.md
P14	B	Simulating plasma wave propagation on a superconducting quantum chip	Quantum	Phys. Rev. App.	1	LLNL	https://arxiv.org/abs/2507.09479	Add ZZ interactions to the spin-chain model and test whether it quantitatively reproduces nonlinear plasma dynamics.	benchmark/paper14/task_spec.md	benchmark/paper14/paper.pdf	benchmark/paper14/paper.md
P15	B	Beyond Classical Molecular Dynamics with Layered Interatomic Potentials	Materials science	Unpublished	0	SNL	null	Replace the Sommerfeld potential's heuristic sigmoid with a physically principled switching function.	benchmark/paper15/task_spec.md	benchmark/paper15/paper.pdf	benchmark/paper15/paper.md

Shoulders of Giants

Can AI agents build on a scientist's work and write a follow-up paper? Each of the 15 tasks gives an agent a published (or recently submitted) scientific paper and a follow-up research direction proposed by the paper's own author. The agent has to carry out the research and write the follow-up paper. Code, rubrics and graders: GitHub. Leaderboard and every judge verdict: website.

Subsets

Subset	Content
`benchmark/`	The 15 tasks: `paper<i>/task_spec.md`, the original paper (`paper.pdf`, its text `paper.md`, figures `images/`) and `index.csv` (domain, venue, lab, link to the original paper, the follow-up direction). The code and data a task uses are downloaded from their original sources by `prepare_tasks.py` in the GitHub repository.
`outcomes/`	The follow-up papers the agents wrote: `<agent>/paper<i>.pdf` is the `proposal/report.pdf` of that submission, the PDF the judge read. `index.csv` lists every submission; four GLM-5.3 runs produced no PDF.
`file_systems/`	For every submission, the files the judge read: `proposal/` plus the rest of the agent's working directory (as `workdir_outside_proposal/`). Python/conda environments, caches, agent configuration files and the final transcript `ai_scientist_trajectory.json` were not shown to the judge and are not included; for Astra-6 and Opus-5.5 the view does contain the runner's streamed transcript `ai_scientist_trajectory.jsonl`. Each submission is an uncompressed tar split into parts, with `manifest.json` listing every file's path, size and SHA-256. For nine runs, `supplement.tar.part-*` and `supplement_manifest.json` add files of the run's working directory that the copy the judge read lacked, recovered from the original run folders (the judge did not see them): Opus-5 P1, P2, P9, P11 and P13, whose archived copies kept only `proposal/` and dropped checkpoints and files over 100 MB (50.7 GiB recovered), and a few files whose names differ only in case for Astra-6 P2 and P13 and Opus-5.5 P2 and P12. Extract `view.tar` and `supplement.tar` into the same folder.
`trajectories/`	Each run's `ai_scientist_trajectory.json` (`<agent>/paper<i>.json`): run metadata (model, effort, attempts, run time) and every recorded agent event. DeepSeek-V4-Pro's harness kept its transcript in its own session logs, so its files hold only a few marker events. `<agent>/paper<i>.harness_logs.tar` hold each harness's own logs: DeepSeek-V4-Pro's session logs (`.dsh/`, its full conversation), GLM-5.3's OpenHands event log (`.oh_conversations/`) and Fable-5's runner logs (`_runlogs/`); `harness_logs.csv` lists them. `<agent>/paper<i>.runner_logs.tar` hold the console logs of the runner that drove the run (a one-line summary of each agent step, attempts, completion checks and final status; PBS job output for the ABCI runs), where they were kept; `runner_logs.csv` lists them. `index.csv` summarizes each run.

Agents (<agent>): opus-4.8, opus-5, opus-5.5, fable-5, sol-5.6, astra-6, deepseek-v4-pro, kimi-k3, glm-5.3.

Unpacking a file system

huggingface-cli download prometheus-eval/shoulders-of-giants --repo-type dataset \
    --include "file_systems/astra-6/paper3/*" --local-dir .
cd file_systems/astra-6/paper3
cat view.tar.part-* > view.tar && tar -xf view.tar      # proposal/ and workdir_outside_proposal/

manifest.json gives each file's SHA-256 (as recorded by the judge when it graded the submission) and each part's SHA-256, so a download can be checked file by file.

Notes

outcomes/fable-5/paper8.pdf: the agent put a real person's name, affiliation and e-mail address on its title page; they are blacked out here. All other PDFs are byte-identical to the submissions.
Files written on the compute cluster can contain absolute paths of the cluster's file system.
file_systems/opus-5.5/paper12: two third-party R help pages (proposal/tools/Rlib/fields/help/qsreg.html, proposal/tools/Rlib/spam/help/spam.html) have names that differ only in case from other help pages, so the judge's copy on a case-insensitive file system held another page's content under these names. The archive has the original files; manifest.json lists both under checked_against_judge_record.sha256_mismatch, and the other pages are in supplement.tar.

License

Task specifications, rubrics and this dataset's own documentation are released under the Apache-2.0 license of the GitHub repository. The original papers in benchmark/ remain under their authors' and publishers' terms and are included for research use. Agent outputs (outcomes/, file_systems/) may contain third-party code and data the agents downloaded, which keep their own licenses.

Downloads last month: 575