Benchmark

BioResearchBench: Long-Horizon Biological Research Tasks

Samar Abedrabbo, Alex Pryhoda, Keenan Goodman, Abbey Albright, Jesse Taavola, Cosette Burrese, Justin Tubby, Juan Villagomez, Evrhet Milam, Joe Kim

View leaderboard →Request sample dataRead preprint →
50
Expert tasks
10
Biology subdomains
18%
Top pass rate

Models evaluated

  • Claude Opus 5.5
  • Gemini 4 Argon
  • Muse Wonder
  • GPT-6 Astra
  • Grok 4.7
Leaderboard

Results by model

Five models answered the same 50 tasks. A task passes only if the answer meets every rubric criterion. The best model passed fewer than one task in five, yet every model met most rubric criteria on average, missing about 18 to 32% per task.

Task success rate
RankModelPass rateGraded answersPassing answersCriteria missed
1Claude Opus 5.5
18% 10–28
1963618.3%
2Gemini 4 Argon
13% 7–20
1702318.4%
3Muse Wonder
11% 5–19
2002223.3%
4GPT-6 Astra
4% 1–8
1961121.4%
5Grok 4.7
4% 1–8
133731.7%
Task-level pass rates
Claude Opus 5.5Gemini 4 ArgonMuse WonderGPT-6 AstraGrok 4.7
Coupled-assay enzyme kinetics100%100%100%40%25%
Neonatal resting potentials100%25%100%0%75%
Baboon daily rhythms75%67%50%0%0%
Proton-linked binding thermodynamics75%0%75%40%0%
CRISPR dependency screen100%0%0%0%67%
Cytostatic arrest versus apoptosis50%25%75%0%0%
Transgenerational methylation in zebrafish100%25%0%0%0%
ORE15 targets in Arabidopsis roots100%0%0%0%0%
Skin barrier0%75%25%0%0%
16S microbiome analysis50%50%0%0%0%
Cortical protein half-lives50%50%0%0%0%
Marker baseline before an unequal run-in0%50%25%0%25%
Yeast copy number and TMT ratio compression0%0%75%0%0%
Low-biomass metagenomics25%50%0%0%0%
Isolate or contaminant: antibiotic resistance25%50%0%0%0%
Proline-linked SILAC ratio bias25%0%25%0%0%
REST ChIP-seq and RNA-seq0%0%0%40%0%
THRB and the daily liver transcriptome0%0%0%40%0%
Nocturnal activity under light at night0%0%0%40%0%
Plasma NfL after football practice0%25%0%0%0%
N170 EEG pre-processing0%25%0%0%0%
Repeat-array knock-in in mouse cortex0%25%0%0%0%
Elbow mechanics after pitching0%25%0%0%0%
Sequencing depth versus coverage25%0%0%0%0%
Mitochondrial respiratory failure0%0%0%20%0%
Pass rate on the 25 tasks at least one model passed. No model passed the other 25.
Rubric criteria missed
  • Rubric criteria missed % (Lower is better)
groupRubric criteria missed % (Lower is better)
Claude Opus 5.518.3
Gemini 4 Argon18.4
Muse Wonder23.3
GPT-6 Astra21.4
Grok 4.731.7
Five-model average22.6
Average share of rubric criteria each model's answer missed, with each task weighted equally. Most of the work gets done; the task still fails.
By category

What the tasks cover

Task subdomains
  • Genetics, genomics & RNA11
  • Physiology & metabolism9
  • Cancer biology6
  • Neuroscience6
  • Biochemistry & biophysics5
  • Immunology5
  • Microbiology4
  • Other: cell & developmental, clinical, pharmacology4
groupTasks
Genetics, genomics & RNA11
Physiology & metabolism9
Cancer biology6
Neuroscience6
Biochemistry & biophysics5
Immunology5
Microbiology4
Other: cell & developmental, clinical, pharmacology4
Scientific skills tested
  • Share of tasks (%)
groupShare of tasks (%)
Computation & data handling98
Uncertainty & identification88
Biological interpretation84
Controls & causal claims72
Reporting & completeness60
Next experiment or decision40
Share of tasks whose rubric demands each skill.
Failures

Where the models fail

Failures by scientific requirement
  • Claude Opus 5.5
  • Gemini 4 Argon
  • Muse Wonder
  • GPT-6 Astra
  • Grok 4.7
  • Five-model average
groupClaude Opus 5.5Gemini 4 ArgonMuse WonderGPT-6 AstraGrok 4.7Five-model average
Computation59.255.169.467.383.766.9
Interpretation35.73147.633.347.639
Controls27.833.338.933.352.837.2
Uncertainty48.855.853.544.274.455.3
Next step253030304031
Completeness55.151494969.454.7
Lower is better. Share of eligible tasks where a model's answer missed at least one rubric criterion in the category.

Failed tasks

Six tasks from the paper, with what the models got right and the step that changed the conclusion.

  • Genetics & RNA biology
    Why are the roots shorter?
    An Arabidopsis line carrying a disruption of ORE15 has shorter roots. Models explain the observations and choose 15 genes for a binding test.
    Got right
    Four of the five noticed that many genes with increased RNA lie beside one another near ORE15, and excluded that region from the binding-test list.
    Missed
    Grok 4.7 missed the neighboring expression pattern and included genes from that region among its binding candidates, which can direct binding tests toward a linked change rather than ORE15 regulation.
  • Biochemistry & biophysics
    How fast is cortical protein replaced?
    Labeled-lysine data from mouse cortex at days 4, 8, 16 and 32, in ordinary and 5xFAD mice. Models estimate protein half-lives.
    Got right
    Four of the five corrected for newly made protein built from unlabeled lysine, which would otherwise make replacement appear slower than it is.
    Missed
    Claude Opus 5.5 interpreted time courses that do not fit the cortical precursor as evidence the proteins were made outside the cortex. A mismatch has other explanations.
  • Cancer biology
    Which genes do the cancer cells need?
    A pooled CRISPR screen with guide counts, reference gene sets and DNA copy number. Models identify candidate growth dependencies.
    Got right
    All five excluded a replicate culture that disagreed with the others, and recognized the problem posed by high-copy regions.
    Missed
    Guides aimed at the same amplified region can share the same DNA-cutting damage, so their agreement does not distinguish that damage from a requirement for the gene.
  • Physiology & metabolism
    What does methyl pyruvate rescue?
    Oxygen consumption, imaging, ATP and protein measurements from six donors. Models explain why NRX13 reduces respiration and methyl pyruvate restores some of it.
    Got right
    Gemini 4 Argon, GPT-6 Astra, Claude Opus 5.5 and Muse Wonder corrected sample identities and an image-area unit error, and compared treatments within donors.
    Missed
    Gemini 4 Argon described near-complete respiratory recovery, then broadened it to complete functional rescue in its final table.
  • Genetics & RNA biology
    Which proteins does miR-185 regulate?
    Activity-associated protein signals, 3′ untranslated region sequences and the miR-185 sequence from Huh7.5 cells.
    Got right
    Four of the five identified PREPL and ACOT1 as candidates supported by both the protein signal and a matching seed sequence.
    Missed
    Gemini 4 Argon labeled candidates as validated direct targets, although no validation experiment had been performed.
  • Neuroscience
    Is the early face response real?
    Averaged EEG from 24 participants, processed with 0.01 Hz and 1.0 Hz high-pass filters, and a report of a 17 ms display delay.
    Got right
    All five identified the early positive signal as an effect of the higher filter setting and applied the display-delay correction.
    Missed
    Muse Wonder and Grok 4.7 omitted requested details about a later response and the effect of filtering on comparisons across electrodes. These were partial successes rather than whole-task passes.
Methodology

Task creation methodology

  1. 01

    Expert-authored from real scientific workflows

    PhD-level domain experts design realistic research problems based on work they or their peers would actually perform, rather than synthetic textbook questions.

  2. 02

    Grounded in scientific data

    Tasks incorporate realistic research materials, including sequencing files, microscopy images, molecular profiles, experimental measurements, and other domain-specific data.

  3. 03

    Verifiable by construction

    Each task includes a golden reference answer and supporting verification materials, enabling independent validation of the expected scientific results.

  4. 04

    Multi-stage expert quality assurance

    Every task undergoes independent PhD-level review followed by internal domain expert validation, with iterative revisions to ensure scientific accuracy, clarity, and verifiability.

  5. 05

    Frontier model testing and difficulty calibration

    Tasks are evaluated against leading AI models through repeated runs. During development, a task was accepted only if GPT-6 Astra failed it. We chose GPT-6 Astra because it was one of the strongest biology models at the time, which set a high bar for difficulty. As newer models answered some questions correctly, those questions were revised or removed.

Example tasks

See exactly what the models saw

Two of the 50 tasks as the models received them, packaged for Harbor. Open a task to browse its files.

  • Cancer biologyTask 1

    Which genes are needed for cancer-cell growth in a 21-day CRISPR screen?

  • Biochemistry & biophysicsTask 2

    How quickly are brain proteins replaced in control and amyloid-model mice?

Want more than two? Request sample data and we'll send a larger set.

Horizon
Officehours Expert Data Lab

Leave your email and our team will send you sample data.