Biology
Cell and molecular biology, genomics and bioinformatics, structural biology, neuroscience, microbiology, and ecology.
Tools you might use
- Fiji / ImageJ
- CellProfiler
- Biopython
- BLAST
- Snakemake
- R / Bioconductor
- AlphaFold
- PyMOL
- GROMACS
Open call · Biology & materials science
Let's build AI that actually helps your research.
We're AI researchers who teach language-model agents to use real scientific software. Now we want to learn from the people doing the science: how you work, where AI already helps, and where it falls short. If you're a graduate student in biology or materials science, come shape what we build and write the paper with us.
- No ML background needed
- Your workflows, your expertise
- Contributors can join the paper
Our team includes researchers from
- University of Maryland
- University of Georgia
- National University of Singapore
- Indiana University
- The Hong Kong Polytechnic University
and collaborators at other universities
Why we need you
Frontier AI agents still struggle with real science.
We built a test from 100 tasks based on real scientific software and ran eight frontier models on it. The best one solved one task in five. Life-science tasks were solved once in 120 attempts.
Most training data for AI agents comes from software engineering, so the workflows scientists actually run are barely covered. Closing that gap takes people who know what a correct result looks like in their field.
20%Success rate of the best of eight frontier models on our held-out scientific-software test
1 in 120Attempts that succeeded on life & molecular science tasks, across all eight models
98%Agreement between our automated checks and blinded human reviewers
Who we're looking for
Graduate students in biology and materials science.
Master's or PhD, wet lab or computational, any level of AI experience. If you script, simulate, image, or analyze data as part of your research, you have exactly the knowledge we're missing.
Materials science
Computational materials, characterization, polymers and soft matter, energy materials, and metals and ceramics.
Tools you might use
- VASP
- Quantum ESPRESSO
- LAMMPS
- ASE
- pymatgen
- Materials Project
- GSAS-II
- OVITO
- COMSOL
Working in a neighboring field, like chemistry, biophysics, or chemical engineering? We'd still love to hear from you.
Let's talk about your research
Where does AI help you today, and where does it fall short?
We want to understand research from the inside: what fills your week, which steps AI already speeds up, and which ones it gets wrong. Here is our starting picture of a typical project. Help us correct it.
Pick a stage of a typical project.
Biology
Reading & planning
- What it looks like
- Keeping up with papers and preprints, comparing protocols, and planning the next experiment.
- Where AI falls short
- Confident summaries with missing or invented citations, and little sense of which protocol details actually matter.
- What we'd build together
- Tasks that check an agent's answers against the papers and protocols you rely on.
Designing experiments
- What it looks like
- Designing primers, guide RNAs, constructs, and controls, often across several separate tools.
- Where AI falls short
- Chaining specialized tools together and respecting organism- and lab-specific constraints.
- What we'd build together
- Design tasks scored by the same tools and rules you use to sanity-check a design.
Processing data
- What it looks like
- Sequencing pipelines, microscopy segmentation, and flow-cytometry or plate-reader data.
- Where AI falls short
- Installing tools, wiring up multi-step pipelines, and catching silent errors in the output.
- What we'd build together
- Pipelines rebuilt from your real runs, with checks on the numbers, not just on whether files exist.
Analysis & modeling
- What it looks like
- Statistics, structure prediction, docking, and molecular dynamics.
- Where AI falls short
- Choosing sensible methods and parameters, and noticing when a result is biologically implausible.
- What we'd build together
- Analyses checked on hidden test cases, so an agent has to get the method right, not just the format.
Figures & writing
- What it looks like
- Figures, methods sections, and notebooks that others can rerun.
- Where AI falls short
- Methods text that drifts from what the code actually did.
- What we'd build together
- Tasks where every figure and methods detail has to match the analysis that produced it.
Materials science
Reading & planning
- What it looks like
- Finding synthesis routes, processing conditions, and property data scattered across many papers.
- Where AI falls short
- Pulling values out of tables and figures without mixing up units, samples, or conditions.
- What we'd build together
- Extraction tasks checked against values curated by people who know the field.
Setting up simulations
- What it looks like
- Building DFT or molecular-dynamics inputs, running convergence tests, and writing cluster job scripts.
- Where AI falls short
- Physically sensible parameters, pseudopotential and force-field choices, and the quirks of each cluster.
- What we'd build together
- Simulation tasks checked by rerunning the real codes on inputs the agent never saw.
Characterization
- What it looks like
- Processing XRD patterns, electron-microscopy images, and spectra.
- Where AI falls short
- Peak fits and refinements that look reasonable but aren't physically right.
- What we'd build together
- Characterization pipelines rebuilt from real instrument data, with checks that respect measurement tolerances.
Screening & data
- What it looks like
- Querying materials databases, featurizing structures, and screening candidates.
- Where AI falls short
- Keeping structures, compositions, and units consistent from one tool to the next.
- What we'd build together
- Screening workflows checked against reference results from the codes you trust.
Figures & writing
- What it looks like
- Phase diagrams, band structures, and methods that others can reproduce.
- Where AI falls short
- Plots and text that drift from the underlying calculations.
- What we'd build together
- Tasks where every figure traces back to a verified calculation.
Questions we'd love to ask you
No preparation needed. These are the kinds of things we'd talk about in a first conversation.
- Which part of your week feels most like busywork?
- Where do you already use AI, and where have you stopped trusting it?
- What software, instruments, or pipelines does your work depend on?
- How do you check that a result is right before you believe it?
- Which task would you hand to a reliable assistant tomorrow?
- What would an AI agent need to show before you'd publish its result?
Demos from our previous work
Agents working with real scientific software.
Each demo is a task from our research. An agent sees a few real runs of a scientific tool and writes a program that reproduces it. We then run that program on settings it has never seen and check every output against the real software.
Life & Molecular Sciences
Microscopy cell segmentation and counting
The same steps as counting nuclei or measuring cell shape in your own fluorescence images.
- Biology
- Microscopy
- CellProfiler
- scikit-image
- Bio-Formats
Physical & Engineering Simulation
OpenFOAM 2-D dam-break CFD
Meshing, boundary conditions, and solver settings: the setup behind any flow or transport simulation.
- Simulation
- Fluid dynamics
- OpenFOAM
Spatial, Visual & 3D
OpenSCAD / BOSL2 parametric fan duct
Parametric CAD for 3D-printed sample holders, fixtures, and flow cells.
- Lab hardware
- CAD
- OpenSCAD
- BOSL2
Earth & Space Sciences
Astronomy FITS/WCS cutout and photometry
Aperture photometry and background subtraction, the same math as quantifying spots or bands in noisy images.
- Imaging
- Signal extraction
- Astropy
- Photutils
Spatial, Visual & 3D
Color-space and tone-curve pipeline
The color-space math behind calibrated imaging, colorimetric assays, and optical characterization.
- Imaging
- Optics
- colour-science
- ImageMagick
- scikit-image
Data & Document Workflows
Tesseract table-image OCR
Reading numbers out of scanned tables, the first step in mining property data from papers.
- Literature
- Data extraction
- Pandoc
- Poppler
- Tesseract
- QPDF
Physical & Engineering Simulation
MuJoCo / URDF mass-inertia repair
Physical models with masses, inertias, and joint limits, as used to simulate lab robots before running them.
- Lab automation
- Physics
- MuJoCo
- URDF/MJCF tooling
Spatial, Visual & 3D
Blender cloth flag scene
A physics simulation whose outcome shifts with wind, constraints, and resolution.
- Simulation
- Physics
- Blender
Spatial, Visual & 3D
KiCad LED-matrix driver PCB
Design-rule checks on circuit boards, like the custom electronics inside lab instruments.
- Instrumentation
- Electronics
- KiCad
Software & Formal Systems
Container SBOM and supply-chain attestation
Recording exactly which software went into an environment, the backbone of reproducible analysis.
- Reproducibility
- Software
- Hadolint
- Syft
- Skopeo
Everyday research computing, too
Plenty of research time goes to installing tools, managing data, and fixing broken environments. Our agents practice that as well. Each replay shows commands from the task's own reference solution.
HPC & clusters
Install an HPC scheduler's Python bindings, verified
Install the Flux Python package from the tarball at /app/flux-python-0.48.0rc6.tar.gz after verifying its integrity against the provided checksum file.
Research data management
Version a dataset with DVC and reconcile its history
Set up a DVC-tracked dataset under /app/dataset/ with two items: dataset/foo and dataset/test/0.
Python environments
Diagnose and fix a Python dependency bug
In /app, diagnose and fix the environs URL parsing bug.
How we'd work together
From one conversation to a shared paper.
Start small and go as far as you like. Every step helps, and you decide how involved you want to be.
- 1
Talk
An informal conversation about your research, your tools, and where AI helps or gets in the way.
- 2
Pick a workflow
Together we choose a real piece of your work: an analysis script, a simulation setup, or a data pipeline.
- 3
Turn it into a test
We package it as a task an agent can attempt, with checks you agree are scientifically right.
- 4
Improve and publish
We measure today's agents, train better ones, and write up what we learn together.
What you'd bring
- How research in your field actually gets done
- Real workflows: scripts, simulations, and pipelines
- Judgment about what counts as a correct result
- Feedback on what agents get right and wrong
What you'd get
- Co-authorship on the resulting paper for substantial contributions
- A first look at how today's AI agents handle your own workflows
- A packaged, tested version of the workflow you contributed
- Hands-on experience with how AI agents are trained and evaluated
Ways to take part
One conversation
Tell us how you work and where AI lets you down.
Contribute a workflow
Help turn a piece of your research into a task with real checks.
Co-author
Join the analysis and the writing of the paper.
Our research so far
Results from our recent papers.
The demos above come from this work. Each paper tackles one part of the problem: building tasks that check results reliably, making tasks harder over time, and turning an agent's successes into better training data.
Self-Supervised Scaling of Terminal Environments for Scientific Domains
Builds agent tasks directly from existing scientific software. The agent sees a few real runs and must reproduce the workflow, and hidden runs check the result. The tasks span six domains, including life and molecular sciences, physical and engineering simulation, and earth and space sciences.
Why it matters to you: this is how we'd turn your workflows into tasks.
Key results
Fine-tuning on these tasks, Terminal-Bench 2
Before 47.9%
After 53.6%
Long-horizon terminal tasks
Before 20.7%
After 27.7%
Automated checks agree with blinded human reviewers
98%
Recursive Synthesis for Long-Horizon Terminal Tasks
Grows simple, verified tasks into long, multi-step ones, round after round. Every new task is checked in a fresh sandbox before it is kept, at roughly $0.05 per task.
Why it matters to you: agents learn to finish whole projects, not just single commands.
Key results
Same strong model, tasks from round 1 vs. round 15
Round 1 solved 90%
Round 15 solved 2.5%
Median reference solution length
67 lines
374 lines
Reinforcement learning on these tasks, Terminal-Bench Hard
Before 22.7%
After 32.0% (+41% relative)
Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite
Collects an agent's successful solutions from several different setups, then rewrites them into one consistent style the model can learn from. The rewritten data teaches the model far more than the raw successes do.
Why it matters to you: one good solution to your task can become many training examples.
Key results
Terminal-Bench 2, solved within three tries
Before 57.0%
After 74.2%
Terminal-Bench Hard, solved within three tries
Before 39.0%
After 63.0%
Terminal-Bench 4, solved within three tries
Before 1.5%
After 9.1%
Questions
Before you write to us.
Anything else? Ask us directly. We're happy to talk it through.
Who are you?
AI researchers from the University of Maryland, the University of Georgia, the National University of Singapore, Indiana University, The Hong Kong Polytechnic University, and other universities. We build and train AI agents that work with real software, and we publish our results on arXiv.
Do I need machine-learning experience?
No. We bring the AI side. What we need is your knowledge of how the science is done and how you know a result is right.
How much time does it take?
It's up to you. A single conversation already helps us. Contributing a workflow takes more time, and we'll plan it around your schedule.
Do I have to share unpublished data?
No. Tasks can be built from open-source software with public or synthetic inputs. We'll agree together on anything that goes into a paper.
Can I be an author on the paper?
Yes. Collaborators who contribute workflows, tasks, or analysis can be co-authors on the resulting paper, following standard authorship norms.
Who can join?
Graduate students, Master's or PhD, in biology, materials science, and related fields. Not sure you fit? Write to us anyway.
Get involved
Tell us about your research.
A few lines is plenty, and we'll reply to set up a conversation. Sending opens your email app with your answers filled in, so nothing is stored on this site.
Prefer plain email? Write to zli12321@terpmail.umd.edu and zl22754@uga.edu.