Nam H. Le, Douglas Blackiston, Michael Levin and Josh Bongard have trained a system to turn short instructions such as “speed up” into a choice from previously recorded electrical-stimulation trials on xenobots. Their September 30 arXiv preprint reports that its preferred model chose an archived outcome in the intended direction for 80.0% of its most demanding held-out comparisons. That is an offline matching result. The authors did not report applying the model’s predictions to newly stimulated xenobots.

For biologists interested in directing living systems, the useful question is what AI supplied between a sentence and a physical intervention. Here it judged old movement traces, helped train a prompt-to-duration model, and returned an earlier trial that most closely matched the requested behavior. It did not establish that a fresh organism would respond the same way.

The big change

  • What changed: The researchers used an AI judge to connect plain-language motion requests with electrical stimulation durations in an existing xenobot archive. This gives them a way to train a language interface without collecting a new experiment for every instruction.
  • Why it matters: Researchers with intervention and outcome records could investigate whether those records support a useful language interface before spending laboratory time on prospective trials. In this study, the interface handles one stimulation parameter and coarse motion directions, so its practical value remains untested.
  • What to watch: The decisive next experiment is to apply a selected duration to a new xenobot and measure whether its motion follows the request. The preprint proposes that test; its reported accuracy measures archive matching.

What the model actually learned

Xenobots are small living constructs made from Xenopus laevis cells. In the experiments reused for this paper, an individual construct was tracked before and after an electrical stimulus. The stimulus duration was the single parameter the new system could choose. The archived video showed whether that xenobot moved faster or slower afterward.

The authors first extracted two-dimensional paths from those videos. Their tracking pipeline used the Segment Anything Model with several detection methods to handle bubbles, reflections and debris. They then gave a vision-language model, identified as Claude in the paper’s AI-use statement, a prompt and a plotted movement trace. Claude scored how well the recorded motion fitted the words. It did not see the stimulus duration when making that judgment.

Those scores became a fixed reward table. A separate network took an embedding of the instruction and predicted a stimulation duration. For the main training method, the researchers fitted a smooth score curve across archived durations for each prompt, derived a target duration and trained the network to predict it. When showing a result, the system matched the prediction to an already recorded trial and displayed that trial’s trajectory. Figure 1 of the paper sets out this sequence, while its Figure 10 demo returns old trajectories for three typed prompts. “Live interface” in the paper describes that interactive display, not live control of an organism.

How strong is the archive result?

The starting collection contained 149 tracked trials. The authors excluded 48 as unreliable, leaving 101 trials from distinct xenobots: 26 classified as faster and 75 as slower. They report a 61-trial training subset and a 40-trial test subset. The system used 33 training prompts; its evaluation also used prompts absent from training. The reported test counted a predicted duration as correct when the three closest trials in the relevant archive subset, by duration, voted for the instruction’s intended motion direction.

On the comparison using both new prompts and the test archive subset, the paper’s preferred training method reached 80.0% accuracy across 30 training seeds. A rule that always predicts “slower” scores 66.7% here, because the test balances three prompt types but combines “slow down” and “stop” into the same slower category. A model trained with a simpler target selected from measured velocity changes reached 59.1% on that comparison. Those are the authors’ results on their dataset, not a measured success rate on new xenobots.

The authors also checked the AI judge against motion labels derived independently from the tracked trajectories. They report 97% agreement on faster versus slower classification. That check supports the judge’s use for this coarse distinction, although its score correlated less strongly with the size of a speed change. The method could not reliably learn a separate command for a mild slowdown versus a stop; both count as “slower” in the main result.

There is a split detail for an independent reviewer to resolve. Section 3.4 describes scoring all 101 archive trials against the 33 training prompts to make the reward table, while Section 3.5 describes a disjoint archive train/test evaluation. The paper presents the latter as held out, but the main-text description does not make clear how test-trial scores were excluded from the training targets. This uncertainty limits how confidently the 80.0% can be read as generalization to unseen archive records.

The missing biological test

The intervention remains a duration retrieved from the archive. The model cannot compose a new electrical protocol, choose other stimulus settings or demonstrate that a new xenobot will respond as an archived one did. Its training and evaluation did not add a wet-lab trial. Nor does the preprint establish similar results for organoids, tissue repair or clinical use, despite those broader possibilities in its title and discussion.

The paper is an arXiv preprint, and we found no independent replication in its reported evidence. Its immediate contribution is a method for learning from records in which instructions were never paired with interventions. Whether that method can direct living cells prospectively remains an experimental question.

Sources & further reading