Who Moved My Boson? Revealing Resonance Features Inside a Neural Network
This paper demonstrates that a generative neural network trained on unlabeled collider data to model particle kinematics has internally learned human-interpretable features representing physical resonances like the boson, which can be manipulated via mechanistic interpretability techniques to alter specific physical quantities such as mass without affecting the rest of the spectrum.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer
In the vast, high-energy collisions that occur inside the Large Hadron Collider, particles smash together and break apart, creating a chaotic spray of new matter. To make sense of this chaos, physicists rely on computer models that simulate these collisions, predicting what particles should appear and how they should move. For decades, these simulations have been built by physicists writing down the known laws of nature, piece by piece. But in recent years, a different kind of model has emerged: artificial intelligence systems that learn the patterns of particle collisions directly from the raw data recorded by the detectors. These systems, known as generative models, do not need to be told the rules of physics; they simply observe millions of real events and learn to recreate them. The big question for scientists is whether these AI models are merely memorizing the data or if they have actually learned the underlying concepts of physics, such as the existence of specific particles or the way they decay. If an AI has truly learned a concept, it should be possible to find a specific "knob" inside its internal workings that controls that concept, allowing researchers to turn it up or down and change the outcome in a predictable way.
A team of researchers has now demonstrated that they can find and turn such a knob inside a neural network trained on real collision data. The model they studied, called ShellFlow, was trained on millions of events from the ATLAS experiment at the Large Hadron Collider. Its job was to generate the kinematics—the speeds, directions, and energies—of the particles that emerge after a collision. Crucially, the researchers never told the model which particles came from a specific source, nor did they label the data with the names of the particles involved. The model was simply asked to recreate the statistical patterns of the real world. Despite this lack of instruction, the researchers discovered that the network had independently learned to represent the Z boson, a heavy particle that decays into a pair of muons. They found a specific direction within the network's internal data that corresponded to the mass of this particle. By adding a small mathematical adjustment to this direction, they could shift the mass of the Z boson in the generated events, moving it from its natural value of about 91 GeV to as low as 70 GeV or as high as 115 GeV.
The experiment worked with a precision that surprised the team. When they adjusted this internal direction, the peak representing the Z boson in the data moved exactly as predicted, while the rest of the particle spectrum remained largely untouched. Other particles, such as the J/psi and the Upsilon, which appear as distinct peaks at much lower masses, stayed in their original positions. This selectivity proved that the network had not just learned a general rule about particle masses, but had formed a specific, isolated concept for the Z boson. The researchers also tested whether this concept was essential to the network's operation. By removing the direction associated with the Z boson from the model's internal stream of information, they found that the model could no longer produce a sharp, well-defined peak for the particle; instead, the peak became significantly broader and less distinct. This confirmed that the network was actively using this specific internal feature to construct the Z boson in its simulations.
One of the most striking aspects of the discovery is that the model learned this concept without any human guidance. The training data was unlabeled, meaning the computer did not know that a Z boson existed or that it had a specific mass. It only saw the final outcomes of collisions. Yet, through the process of learning to generate these outcomes, it organized its internal calculations in a way that separated the Z boson from other particles. The researchers used a method based on how the model's output changes in response to small tweaks in its internal state. They identified a path through the network's layers that, when nudged, caused the mass of the muon pairs to shift. This path was found by analyzing the gradients, which are essentially the slopes that tell the model how to adjust its internal numbers to get a better result. By finding the direction where these slopes were strongest for the Z boson and weakest for everything else, they isolated the specific "steering vector" needed to control the mass.
The ability to steer a physical quantity inside a generative model opens a new way to test what these systems have learned. In the past, if a physicist wanted to see how a change in the mass of a particle would affect a collision, they would have to re-run the entire simulation with new parameters, a process that is slow and computationally expensive. With this new method, the change can be made instantly by adjusting the internal direction of the model. However, the researchers noted that while they could move the mass of the Z boson, the peak also became wider as it moved, suggesting that the model's internal representation of the particle is not yet a perfect, isolated copy of reality. The width of the peak, which relates to how quickly the particle decays, changed in a way that was harder to control independently. This indicates that while the model has learned the concept of the Z boson, the details of its internal representation are still intertwined with other physical properties.
This work represents a significant step toward understanding the "black box" of artificial intelligence in science. It shows that deep learning models trained on complex physical data can develop internal structures that align with human concepts, even when those concepts are never explicitly taught. The researchers did not just find a correlation; they demonstrated a causal link by showing that manipulating the internal direction directly altered the physical output. This suggests that these models are not just pattern-matching machines but are building internal representations of the physical world that can be probed and modified. The findings offer a new tool for physicists to explore the behavior of generative models and potentially use them to generate new data scenarios for testing theories, all without the need to retrain the model from scratch. The study confirms that when a neural network is trained on the fundamental laws of nature, it can learn to organize those laws in a way that is accessible and understandable to the humans who built it.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.