The same knowledge that accelerates learning can also blind the system to discovery.
At the edge of what physics can explain, a team from Princeton and the Flatiron Institute has found a way to ask deeper questions more cheaply — by teaching machines the familiar before confronting them with the unknown. Their transfer learning approach slashes the computational cost of probing physics beyond the standard cosmological model, yet in doing so surfaces an older, more human dilemma: that prior knowledge, the very thing that enables understanding, can also foreclose it. The universe may be offering new signals, but an intelligence shaped by the past risks hearing only echoes of what it already knows.
- Testing theories of new physics requires thousands of costly universe simulations — a computational burden so heavy it can outpace even well-funded science.
- Transfer learning cuts that burden by more than tenfold, letting AI absorb the standard cosmological model first before tackling stranger, more speculative physics.
- A hidden danger emerged: the AI sometimes mistook genuinely novel phenomena — like the gravitational fingerprint of massive neutrinos — for familiar patterns it had already learned to recognize.
- This 'negative transfer' is not a glitch but a structural risk, rooted in the physics itself, where different parameters can produce nearly identical observable effects.
- Upcoming sky surveys will flood researchers with unprecedented data, making the stakes of this bias problem urgent — an AI too comfortable with the known may miss the moment the universe speaks something new.
Physicists testing theories beyond the standard cosmological model face a brutal computational reality: each candidate universe must be simulated, and running enough simulations to properly evaluate a new theory can exhaust even well-resourced institutions. Researchers at Princeton and the Flatiron Institute turned to a technique borrowed from machine learning — transfer learning — to ease that burden. The idea is straightforward: train a neural network on the simpler, well-understood standard model first, then adapt it to the harder task of learning new physics. The network carries its prior knowledge forward, needing far fewer expensive simulations the second time around. In some cases, the savings exceeded a factor of ten.
But the study's most consequential finding was a warning embedded in that success. Lead author Veena Krishnaraj observed that the pretrained AI sometimes grew too fluent in the familiar. When simulations involving massive neutrinos produced effects that resembled a known standard-model parameter — the clustering amplitude σ8 — the network initially misread the new physics as a variation on old patterns. Machine learning has a name for this failure mode: negative transfer. The same prior knowledge that accelerates learning can cause a system to interpret ambiguous signals through the lens of what it already knows, rather than recognizing them as something genuinely different.
The risk is not incidental. It arises from the physics itself, where different parameters can generate nearly indistinguishable observational signatures — degeneracies that confuse human and machine alike. As vast new cosmological surveys prepare to release unprecedented volumes of precision data, the question is no longer only whether AI can help physicists work faster, but whether it can be trusted to recognize the unfamiliar when it finally appears. The researchers have so far tested their method only on simulations; the deeper reckoning awaits real data, and the possibility that the universe is already showing us something we are not yet prepared to see.
Physicists have a problem that grows worse the more ambitious their questions become. Testing whether the universe operates according to physics beyond what we currently understand requires running simulations—thousands upon thousands of them, each one a virtual cosmos built from different physical assumptions. The computational cost is staggering. A team of researchers at Princeton University and the Flatiron Institute wondered whether they could borrow a trick from machine learning to make the work faster: teach an AI system the basics first, then let it tackle the harder stuff.
The standard cosmological model, known as ΛCDM, works remarkably well. It explains the universe's expansion, the way galaxies cluster, the afterglow of the Big Bang. But physicists suspect it is incomplete. Observations hint at phenomena the model cannot account for—massive neutrinos, gravity that behaves differently than Einstein predicted, dark energy that changes over time. To test whether these alternatives might be real, researchers need to simulate universes under different physical rules and see whether those simulations match what telescopes actually observe. Each simulation is expensive. Running enough of them to properly test a new theory can demand computational resources that strain even well-funded institutions.
Transfer learning offers a shortcut. The technique, familiar in artificial intelligence, works like this: train a neural network on one task, then adapt it to a second, more difficult task. The network reuses what it learned the first time around. In this case, Adrian Bayer and his colleagues trained their AI on simulations based on the standard model—the easier, less computationally demanding work. Only after the network had absorbed those patterns did they ask it to learn from simulations that included possible new physics. The strategy worked. In some cases, transfer learning reduced the number of expensive simulations needed by more than a factor of ten.
But the researchers discovered something unsettling in their results. Sometimes, the AI became too confident in what it already knew. When new physics produced effects that resembled patterns the network had seen before, the system would misinterpret the new information through the lens of the old knowledge. The phenomenon has a name in machine learning: negative transfer. Veena Krishnaraj, the paper's lead author, observed this behavior clearly in simulations involving massive neutrinos. Certain effects produced by neutrino mass look similar to variations in a standard model parameter called σ8, which describes how tightly matter clusters across the universe. The pretrained network initially confused the two, struggling to recognize the genuinely new physics hiding beneath a familiar pattern.
The problem is not random. It emerges from the underlying physics itself. Different physical parameters can produce observable effects that are nearly identical, creating what physicists call degeneracies. An AI system trained to recognize one set of patterns may naturally interpret ambiguous new data as variations on those familiar themes rather than as signs of something genuinely different. This is the risk embedded in the transfer learning approach: the same prior knowledge that accelerates learning can also blind the system to discovery.
The work represents both promise and peril in applying modern machine learning strategies to fundamental physics. Foundation models—the conceptual cousins of large language models and generative AI—can dramatically reduce computational burden. But they carry a hidden cost: the possibility that an AI system, having learned to see the universe one way, might miss the moment when the universe reveals something new. For now, the researchers have tested their method only on simulations. Real observational data from upcoming cosmological surveys will be the true test. Those surveys will generate unprecedented volumes of high-precision measurements about the cosmos. The question facing physicists is whether their AI systems will be ready to recognize genuinely new physics when it appears, or whether they will be too comfortable with what they already know.
Citas Notables
It's basically a shortcut. Usually people train the AI directly on the most computationally expensive simulations. What we do instead is first use simpler and less expensive ΛCDM simulations to give the AI an idea of what's happening, and only afterward move to the more complex models.— Adrian Bayer, cosmologist at Flatiron Institute and Princeton University
The negative transfer is not random. It is driven by underlying physical degeneracies in the model. So this is something we need to be aware of and try to mitigate.— Veena Krishnaraj, lead author and Princeton undergraduate
La Conversación del Hearth Otra perspectiva de la historia
So the AI gets faster at its job by learning the basics first. That sounds like a win. What's the catch?
The catch is that speed and blindness can come together. When the AI learns patterns from the standard model, it gets very good at seeing the universe through that lens. When something new appears that looks even slightly like something old, the AI interprets it as a variation on the familiar rather than as something genuinely different.
But couldn't you just tell the AI to look for new things? To be skeptical of its first instinct?
That's the hard part. The problem isn't carelessness. It's that the physics itself is ambiguous. A massive neutrino produces effects that genuinely resemble a change in how matter clusters. The AI isn't making a mistake—it's following the patterns it learned. You'd have to teach it to doubt itself in ways that might also make it miss real patterns.
So you're saying the more efficient the AI becomes, the more likely it is to miss something important?
Not exactly. But there's a tension there. The efficiency comes from reusing old knowledge. The discovery comes from being open to what doesn't fit the old knowledge. You have to be intentional about managing that tension.
How do you do that?
That's what the researchers are working on now. They've identified where negative transfer happens. The next step is building safeguards—ways to make the AI aware of its own blind spots, to flag when it's interpreting something ambiguous, to ask for human judgment when the physics is genuinely degenerate.
And if they don't figure it out?
Then the next generation of cosmological surveys—which will produce more data than we've ever had—might confirm theories we already believe in while missing the ones that would actually overturn our understanding.