ulamai/STEM-RL-Tasks

STEM RL Tasks

40 advanced STEM tasks · 8 domains · executable verifiers · 55 recorded model attempts

STEM RL Tasks is a public sample of Ulam's scientific reasoning tasks for reinforcement learning with verifiable rewards (RLVR) and model evaluation. The problems ask models to connect scientific concepts, work through mathematical models, and produce structured answers that can be checked programmatically. Each task includes a problem statement, a machine-readable instance, a reference answer, and an offline Python verifier.

The collection spans astronomy, biology, chemistry, Earth science, engineering, mathematics, medicine, and physics. Problems range from black-hole superradiance and enzyme kinetics to nonlinear fault dynamics, quantum transport, algebraic curves, and cardiovascular hemodynamics. Many combine several calculations, making consistency across assumptions, units, and numerical methods part of the challenge.

Explore Ulam's STEMBench

STEMBench is Ulam's correctness-first evaluation of AI research across biology, chemistry, physics, astronomy, Earth science, and engineering. It measures how far a model advances a well-scoped scientific problem, with credit for supported partial progress, reproducibility, and a clear account of what remains unresolved. Explore the domain comparisons, evaluation methodology, and public research-task atlas to see how today's AI systems approach demanding scientific work. STEMBench uses a research-progress rubric; this repository provides a separate collection of fixed-answer tasks with executable verifiers.

Domains and task coverage

Domain	Tasks	Examples
Astronomy	5	Cosmological recombination, black-hole superradiance, stellar structure, exoplanet irradiation
Biology	5	Enzyme kinetics, molecular thermodynamics, ion transport, population genetics, nutrient cycling
Chemistry	5	Impedance spectroscopy, hydrogel swelling, molecular orbitals, catalytic kinetics, reaction equilibria
Earth science	5	Fault dynamics, geothermal convection, ice-shelf hydrofracture, ocean iron cycling, ozone chemistry
Engineering	5	Magnetohydrodynamics, geotechnical mechanics, information theory, control systems, adsorption
Mathematics	5	Arithmetic geometry, nonlinear elasticity, random matrices, modular forms, knot invariants
Medicine	5	Blood flow, renal physiology, cancer genetics, MRI physics, clinical-trial statistics
Physics	5	Quantum oscillations, open quantum systems, plasma damping, fluid stability, relativistic spin dynamics

Browse the sample split in the Hugging Face dataset viewer for the task statements and metadata. The complete task catalog, attempt inventory, and technical guide are in docs/tasks.json and docs/README.md.

What's included

tasks/<TASK_ID>/
  instruction.md          Problem statement and submission requirements
  instance.json           Task parameters and answer schema
  task.json               Task metadata and file references
  verify.py               Executable answer verifier
  solution/solution.json  Reference answer
trajectory/<TASK_ID>/attempt_<N>/
  trajectory.json         Available captured model and tool record
  submitted_action.json  Complete submitted answer
  result.json            Recorded outcome and available diagnostics
environment/             Shared grading CLI and one-step STEMEnv interface
data/tasks.jsonl          Flat task index for the dataset viewer
docs/                    Package guide, inventory, and SHA-256 checksums

Some mathematics tasks also include an optional reference solver. Task IDs and the original task and trajectory files are retained from the supplied package. The viewer index provides a consistent browsing layer; the original files remain the source of truth.

Download and grade

Download the repository using the Hugging Face CLI:

hf download ulamai/STEM-RL-Tasks --repo-type dataset --local-dir STEM-RL-Tasks
cd STEM-RL-Tasks

Grading requires Python 3.10 or newer and uses only the standard library. It runs offline, with no additional datasets or private container images.

# Validate every packaged reference answer.
python environment/grade.py --all

# Grade your own JSON answer against a task's verifier.
python environment/grade.py --task CHEM1 --candidate candidate.json

# Regrade saved submissions and compare with their recorded outcomes.
python environment/grade.py --check-trajectories

Follow each task's exact answer schema, required keys, units, types, and tolerances. Schemas appear as submission_schema or action_schema in instance.json. Answers can include numerical fields, arrays, integers, or rational strings. A full-task pass requires the entire answer to satisfy the verifier; any diagnostic partial reward is separate from that outcome.

For programmatic use, environment.STEMEnv exposes reset(seed=...) and step(answer). Each episode accepts one answer, making it convenient to connect a policy to task observations and verifier rewards. See the technical guide for an example.

Recorded attempts

The package includes 55 historical attempts: 19 passes and 36 failures, a pooled attempt pass rate of 34.55%. Chemistry includes two attempts per task, mathematics three, and the other domains one. These deliberately selected examples support inspection of successful and unsuccessful submissions; their pooled rate is not a pass@1 or pass@k estimate or a measure of general model performance.

The source records label the models as gemini-4-argon and, for astronomy, gemini-4-argon-high. These are preserved source labels. The release does not add new model runs. Captured trajectory formats vary, and some records contain excerpts rather than complete conversations. Medical records include truncated tool-code and output snippets. Complete saved submitted answers are included for independent regrading.

Validation reproduced all 40 reference-answer passes and all 55 recorded outcomes, including failures. Regrading checks saved answers against the supplied verifiers; it does not replay historical model or tool execution.

Intended use

Use the tasks to prototype RLVR pipelines, evaluate numerical and scientific reasoning, study tool-assisted problem solving, and inspect failure modes against explicit answer contracts. For a blind evaluation, give the solving model only the problem instructions and instance. Keep reference answers, verifiers, saved trajectories, and results on the evaluator side. STEMEnv is a convenience interface and does not provide sandbox isolation.

These are scientific benchmark problems. Verifier acceptance establishes agreement with the packaged reference values and tolerances. Medicine tasks are for research and evaluation, not clinical decision-making.

About Ulam

Ulam develops reasoning data and evaluations for advanced AI systems. Explore STEMBench for scientific research evaluations, methodology, and domain-level comparisons, or visit our website to discuss custom scientific reasoning tasks and evaluation work.