FrontierChallenge: A scientific workflow benchmark

FrontierChallenge

A benchmark for whether agent systems can complete specified scientific workflows and deliver a checkable artifact bundle.

GitHub Paper Hugging Face
Rank Model Agent Scaffold

13 model–scaffold configurations · 97 tasks · Pass Rate = native Score ≥ 99.9.

Selected cases

See the workflow, not just the score.

Three raw X-ray diffraction patterns from different processing conditions 01
Materials characterization

Duplex phase quantification

Nine XRD scans become corrected, sample-level phase and lattice summaries.

Explore case →
Bright-field wound-healing images at zero, twelve, and twenty-four hours 02
Life science

Cell migration imaging

Segment, pair, and compare a complete time-point × group × replicate experiment.

Explore case →
Raw electrochemical impedance spectra at two states of charge 03
Electrochemistry

Equivalent-circuit analysis

Move from raw spectra to circuit comparison, residual evidence, and provenance.

Explore case →
…GACCTGACTGJ1CTGAGCCTGA…
Primer3junction spanning33 off-targetspaired specificity
target.fasta→primers.csv→qc_summary.json
04
Molecular biology

RT-qPCR primer design

Coordinate sequence constraints, thermodynamic QC, offline specificity, and a defensible recommendation.

Explore case →
Gaselectrostaticvan der Waals
Waterelectrostaticvan der Waals
9 λ windows × 4 pathsAMBER-native evidence
05
Molecular simulation

Hydration free energy by TI

Carry a fixed alchemical protocol through four paths and preserve every computational artifact.

Explore case →
Calibrate→Blank correct→Integrate heat→Screen safety
ΔTadMTSRaccumulationheat recovery
raw time series→protocol decision
06
Process chemistry

Reaction-calorimetry safety

Trace a scale-up decision back through calibration, corrections, integration, and replicate summaries.

Explore case →
Four raw quantitative proton NMR spectra from two batches 07
Analytical chemistry

qNMR purity and spectral QC

Turn replicate spectra into a traceable purity summary while rejecting poor-quality inputs from signal evidence.

Explore case →
Raw laser-flash detector curves for three boron nitride loadings 08
Materials characterization

Laser-flash thermal analysis

Sort shuffled temperature-rise traces, identify anomalous curves, and preserve run-to-sample provenance.

Explore case →
Raw voltage sequence from a hard-carbon sodium-ion battery GITT experiment 09
Electrochemistry

GITT diffusion analysis

Resolve pulse, rest, and relaxation stages across six cells into auditable SOC-dependent transport metrics.

Explore case →
Raw blue LED spectra for three photoredox experiment conditions 10
Photochemistry

Photoredox photon balance

Combine actinometry, lamp spectra, reaction absorbance, assays, controls, and held-out validation.

Explore case →