Lila Sciences outlines four challenges for creative scientific AI and releases four preprints addressing two of them
Lila Sciences outlines four challenges for creative scientific AI and releases four preprints addressing two of them
On November 10, Lila Sciences, which is building autonomous scientific laboratories, published a research agenda for open-endedness on its blog. Open-endedness describes a search that branches and keeps producing interesting discoveries indefinitely, as biological evolution does. The agenda identifies four challenges: generating ideas, judging them before running experiments, organizing knowledge within neural networks, and managing the entire scientific cycle from idea to conclusion. Four preprints, papers that have yet to undergo peer review, address the first and third challenges.
The blog post was written by Kenneth Stanley, who leads Lila’s open-endedness team. Test scores, the authors argue, capture little of the intelligence that matters in the world: “future generations remember no one for outstanding test scores.”
The authors argue that idea generation faces a “creativity paradox”: a model trained on successful creative responses becomes more likely to repeat them, yet an idea the model has generated “by definition ceases to be new to it.” In the first preprint, one model writes a primer prompt for a second model, which is kept fixed. The first model is rewarded when the second produces responses that are novel, high-quality, and distinct from one another. This rewards the approach to thinking itself. With a primer, the small open model produced 1,4–3 times as many such responses as the same model directly trained for creativity using two methods. Giving the same primer to four commercial models increased the number of distinct responses by a factor of 4,5–10, with quality maintained or improved.
The second preprint applies a newcomer’s “fresh perspective” to AI: a model is deprived of knowledge about one topic, either by training it on texts that exclude the topic or by fine-tuning it to forget. On three of four topics, “naive” models with 7 billion parameters generated more original ideas than a standard model and the same model instructed to pretend it lacked that knowledge. A model trained on almost no texts about dental hygiene proposed a “tongue stone” that a person moves around the mouth with their tongue to clean their teeth; a mentor model develops these rough ideas into feasible proposals.
The third challenge concerns how knowledge is organized within a neural network. Knowledge can be represented in different ways, just as a number can be written in Roman or decimal numerals, and the representation affects how easily a new conclusion can be reached. In 2025, Stanley and colleagues compared two networks that produced the same image: one evolved on the Picbreeder website and one trained conventionally. The first had an orderly internal structure: its signals followed a shared rule and could be encoded more compactly. The second was fragmented and entangled. The UFO method in the third preprint adds an auxiliary network that compresses the main network’s signals: poorer compression incurs a larger penalty for the main network. In a “grokking” task, a neural network learns division in modular arithmetic. It first memorizes examples, then reorganizes its internal structure and discovers the rule. With the fastest UFO variant, it needed roughly one-twelfth as many steps, a reduction by a factor of 12, as it did without UFO.
The fourth preprint investigates why neural networks forget: further training can erode previously acquired skills. The authors link this to an entangled internal structure in which a change to one part disrupts the others. Two networks that produced an image of a skull were repeatedly retrained to change the mouth. In the conventionally trained network, the rest of the image deteriorated; in the Picbreeder network, it deteriorated noticeably less.
On the second challenge, the blog argues that ideas will outnumber what even an automated laboratory can test, so selecting ideas before experiments will require subjective judgments about what is interesting. The blog illustrates the fourth challenge with a “tree” of runs in which swarms of AI agents write code and conduct experiments. Each branching run builds on the conclusions of its ancestors. In a circle-packing problem, one such run matched the record for 28 circles in 22 evaluations.