SYSONE-MED / biomedia

System One · direct medical decisions

Same question. Five ways to decide.

How medical vision-language models reach a multiple-choice answer — five decision paths, from slow autoregressive writing to a single non-autoregressive scoring pass. Don't just read about it: play with each one and watch every decision it makes.

▶ Open the playground See the five paths
5 decision variants 7 medical datasets 5 interactive playgrounds saved measurements · no live GPU

Playground

Five ways to feel a decision model

Each one replays the same five variants — but from a different angle. Every action shows the decision it made, how confident it was, and how long it took.

🩻OBSERVATORY

Medical VQA / QA

Step through held-out SLAKE and MedQA cases. Watch all five variants answer the same question — each prediction, confidence, and latency, plus a live decision stream.

Open observatory →SLAKE · MedQA
🐍RACE

Snake race

Five snakes, one clock. Each moves at its variant's measured decision speed — and every move is logged in the decision stream with its confidence and latency.

Start the race →side-by-side
🚨RAPID TRIAGE

Time is Brain

A critical case lands and the treatment window is closing. All five paths race to make the right call before it shuts — each at its measured latency. Speed is the whole game.

Run the triage →5 paths · 1 deadline
🚪CLOSING DOOR

The Closing Door

Anaphylaxis: the airway is collapsing in real time. All five paths race to inject epinephrine before the door shuts — the fast paths get the shot in, the slow ones arrive to a code blue.

Watch the airway close →epinephrine vs. the door
📟ALARM STORM

ICU Alarm Storm

Ten patients alarm at once. Five triage teams race on throughput — each clears alarms at its measured latency. In a storm, the model that can't triage fast is useless, no matter how accurate.

Clear the storm →throughput · 10 beds
—
Fastest decision (mean latency)
—
Slowest decision (mean latency)
—
Speed spread across variants
—
Best held-out accuracy (SLAKE)

The comparison

Five decision paths, one question

Each path's measured mean decision latency and held-out accuracy, ranked by speed.

From the paper

The evidence

Figures from the SysOne study.

Four decision formulations

Four decision formulations

AR generation, label likelihood, and learned heads on AR and NAR backbones — same inputs, different decision interfaces.

NAR direct decision architecture

Direct decision architecture

Text and image are encoded once, fused, and scored by a shared head — batched candidate scoring, no autoregressive decoding.

Accuracy versus latency

Accuracy vs. latency

Held-out accuracy against p50 decision latency across image and text tasks. Upper-left is better.

Calibration versus latency

Calibration vs. latency

ECE and NLL against latency. Lower-left is better — direct scoring stays calibrated without the generation cost.

Post-hoc test-time adaptation

Post-hoc test-time adaptation

Routing and conformal TTA over candidate scores — no weight updates, with a marginal coverage guarantee.

Visual dependence ablation

Visual dependence

Accuracy change when the image is missing or shuffled — how much each path actually relies on the visual input.

Under the hood

How each path reaches a decision

Autoregressive

AR Gen · AR LL

Writes text — up to 32 tokens, or exactly one — then parses it into a choice.

Learned head

AR Head

A frozen backbone feeds a small head that scores every option at once.

Non-autoregressive

NAR

Text and image are encoded, fused, and scored in a single forward pass.