BioIntuition Bench

Measuring frontier AI intuition across biological scales

Explore the task examples
Mock data · Design preview

Biological intuition rankings

Compare capability and resource use.

Mock biological intuition leaderboard, with Pass@1 and per-attempt resource use
ModelPass@1 0 — 100%CostOut tokensSteps

Design preview · All scores and resource metrics are illustrative.

01 The idea

Can scientific intuition
be learned?

Intuition is one of the most consequential human capabilities. It gives experts a feel for what to do, an eye for what others overlook, and a sense of what comes next.

In science, these abilities inspire stories of almost superhuman insight, as when a structural biologist predicts how a protein might fold just by looking at its sequence, or when a scientist recognizes healthy and unhealthy cells at a glance. These extraordinary judgments keep experiments working and make discovery possible.

From the outside, these reflexes can seem indistinguishable from magic, but it is our belief that scientific intuition is a learned capability that can be measured.

BioIntuition Bench measures intuition capability with frontier AI doing, seeing, and predicting the life sciences.

The unwritten parts of a protocolGiven the same starting protocol and state, human experts often achieve better outcomes than agents. In a feeding protocol, tacit decisions about plate tilt, pipette position, aspiration speed, and precise spatial movement can determine whether agents match or exceed an expert’s results.Illustrative confluency comparison.
02 The dimensions

Capability, scale, and resources.

Across biological scales

Performance & resource use

Mock
Hover or focus a model to inspect its mock metrics.
Six task examples · Four scales

From molecules
to living cultures.

Explore doing, seeing, and predicting across molecular, cellular, and organoid tasks.

Benchmark overview Draft copy

Benchmark overview

Placeholder copy

Lorem ipsum dolor sit amet: six task examples span wet-lab execution, microscopy interpretation, and molecular prediction. Sed do eiusmod tempor, recurring failures include missed cell boundaries, imprecise pipette placement, and overlooked protocol steps. Ut enim ad minim veniam, Model A leads in seeing while Model B trails in doing; predicting reveals surprising reversals. Duis aute irure dolor highlights task composition, resource tradeoffs, and findings worth investigating.

03 The capabilities

Three ways to understand biology.

The physical Trillo robot operating in the wet lab, filmed on a phone

Doing

We gave agents control of our 10,000 sq ft BSL-2 wet lab in San Francisco and assessed their ability to carry out real protocols and experiments with human biological material. Agents can move robotic arms with 2.5 mm precision, pipette, and read from or control fridges, incubators, centrifuges, autonomous microscopes, and more.

Animated electron microscopy volume showing marked synapses

Seeing

Scientists develop an extraordinary eye for subtle biological signals, from recognizing an unhealthy culture to identifying a synapse in a dense microscopy image. We assess whether agents can make the same judgments by interpreting images, identifying cells and structures, and distinguishing healthy from unhealthy biological states.

Rotating space-filling SOD1 dimer, with its two protein chains colored teal and gold

Predicting

We evaluate whether agents can predict experimental outcomes before they happen. Tasks range from predicting the effects of protein mutations without tool calls to anticipating how a small-molecule perturbation changes cellular stress.

04 Explore the benchmark

Task examples

Open model playground
Simulation

Making Organoids

Follow a simulated batch from one stem-cell culture to 96 wells, then inspect the medium exchanges through day 18.

Simulation

Feeding & Passaging

Follow the Trillo workcell as it exchanges medium and transfers a source culture into a six-well daughter plate.

Imaging · Preview

Synapse Reconstruction

Follow marked synapses through successive electron microscopy slices.

Imaging · Preview

Colony Health Check

Inspect colony shape, density, and gaps in the culture.Illustrative schematic

Imaging

Cell Segmentation

Trace individual cell bodies in a fluorescence image and compare each predicted outline with the original pixels.

Prediction

Protein Ranking

Explore the SOD1 dimer and compare model rankings of 453 variants with the measured dimerization scores.

Launch article · Preview

Introducing BioIntuition Bench

Read the full blog

Chapter breakdown

  1. What makes a neuroscience task?
  2. Inside a Feeding & Passaging rollout
  3. How we score the evidence