A language model was asked to perform the part of a surgical procedure that experienced researchers describe as pure practice, and it improved across three attempts by keeping its own records.
The experiment
Gabriel Ocana-Santero, co-founder and chief executive of the biotechnology startup Exin Therapeutics, shared results in which GPT-6.1 Sol took part in three consecutive rounds of localisation for mouse brain surgery.
After each round, a professional surgeon corrected the model’s localisation. The model retained its own images, its measurements and the record of the expert corrections, and used them as reference for the following round.
By the third round, the time taken to find anatomical landmarks had fallen from 169 seconds to 86. The number of corrective moves required from the surgeon fell from four to two. Localisation error fell from 0.6 millimetres to 0.3.
Ocana-Santero described the process as in-context learning based on real-world experience.
The sample is three operations. That is far from evidence that a model has acquired stable surgical capability, and the company confirmed in a public exchange that the procedures were real rather than simulated.
Why the landmark matters
In mouse brain experiments, the difficulty comes before the needle enters. Two anatomical landmarks on the skull surface, Bregma and Lambda, are the reference points for stereotaxic coordinates.
Bregma sits at the junction of the sagittal and coronal sutures. Lambda sits at the junction of the sagittal and lambdoid sutures. Researchers use Bregma as the origin of the coordinate system, then use Lambda and other landmarks to calibrate the position and tilt of the skull.
Many mouse brain experiments require a needle, electrode or instrument to be delivered to a specific region using a stereotaxic frame. The operator must find the reference points, convert the target region into anterior-posterior, lateral and depth coordinates using a brain atlas, and hand those coordinates to the instrument.
The step looks like finding an intersection. It actually combines visual identification, spatial reasoning and coordinate calibration. Suture lines are not always clearly visible, anatomy varies between animals, and the angle at which the head is fixed changes the result.
For a researcher new to the technique, locating the landmarks and completing the head calibration is basic training that requires repetition.
What the model was asked to do
GPT-6.1 Sol was asked to identify Bregma and Lambda from procedure images and give its own localisation. A specialist surgeon then checked those positions and corrected the inaccurate predictions.
At the end of each operation the model saved its estimated images and measurements alongside the corrected data, which became the context available for the next attempt.
The improvement is therefore not a change in the model’s weights. It is a change in what the model was shown before it started.
The part worth taking seriously
Surgical and laboratory skill has historically been stored in people. It takes a researcher months to become reliable at landmark identification, and the knowledge leaves when the person leaves.
If a record of corrections can be carried into the next attempt and reduce the error rate, the cost of the second procedure falls, even when the first is no better than before.
That has a practical implication for laboratories that run large numbers of animal experiments, where consistency between operators is a persistent source of variance.
It also has an obvious limit. Three attempts with one model on one task, checked by one surgeon, establish a direction rather than a result. The published numbers are a prompt for replication, not a capability claim.
Editor’s note: This is an adapted translation of the original 36Kr report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://36kr.com/p/4060588343825282.