Scientific AI models can start from very different kinds of input and still appear to move toward a similar internal view of matter. A new study from researchers at the Massachusetts Institute of Technology examined this pattern across models built for molecules, materials, and proteins.
The finding is important because it suggests that strong models may not simply memorize separate shortcuts for separate tasks. Instead, when they perform well, they may learn representations that line up with those learned by other high-performing systems.
What the MIT team compared
The team, led by Sathya Edamadaka and Soojung Yang, studied 59 different models. The group included specialized scientific systems for molecules, materials, and proteins, along with large language models such as DeepSeek and Qwen.
These models do not all see the scientific world in the same format. Some work with molecules as coded strings. Others use 3D atomic coordinates. Others process protein sequences. That variation makes the comparison useful, because it tests whether the models are learning something that depends only on format, or something deeper about the structures they represent.
To make the comparison, the researchers extracted each model's internal representations and measured how closely those representations aligned. The source article reports that they used several metrics, rather than relying on a single test.
Strong models converged more clearly
The study found significant alignment across learned representations. Models using 3D coordinates showed strong agreement with one another, and text-based models also aligned with each other. More unexpectedly, similarities also appeared between these groups, even though their inputs were different.
The pattern became clearer when performance was considered. The better a model did on its training task, the closer its internal representation came to that of the best-performing model. According to the researchers, that points toward a shared representation of physical reality among high-performing models.
The study also found that the complexity of these internal representations stayed within a similarly narrow range across the models. That detail matters because it suggests the alignment is not only about isolated similarities. It may reflect a broader structure that different scientific AI models move toward as they improve.
Why representation alignment matters
Representation alignment gives researchers another way to judge whether a model is learning a useful scientific abstraction. A model can perform well on a task, but still rely on a brittle internal solution that does not transfer well. If its representation also aligns with other strong models, that may be evidence that it has captured something more general.
The researchers propose representation alignment as a benchmark for deciding whether a model deserves to be considered foundational. Under that idea, high performance alone is not enough. A model would also need to show strong alignment with other top-performing models.
This is especially relevant for current materials models. The researchers argue that these models have not yet reached foundation model status because their representations are too strongly shaped by limited training data. In their view, broader generality requires far more diverse datasets.
- Performance shows whether a model handles its assigned task well.
- Alignment shows whether its internal representation resembles those of other strong models.
- Generality depends on whether the representation holds up beyond familiar structures.
Novel structures remain the weak point
The same analysis also exposed a major limitation. For known structures that resemble training data, strong models produced matching representations. Weaker models tended to form their own less transferable solutions.
But when the models faced completely novel structures that differed significantly from their training data, almost all of them failed. Their representations became shallow and lost important chemical information. That result narrows the meaning of the convergence finding: the models may agree strongly in familiar regions, but that does not mean they can reliably generalize everywhere.
The source article connects this to a wider problem in current AI systems. Transformer architectures, for example, have been shown to systematically fail at composition tasks when facing out-of-distribution scenarios. In that case, the problem involves combining known facts into new derived facts outside the training distribution.
A broader pattern in AI research
The MIT work extends an idea from earlier research. Back in May 2024, a study from the same institute found that different AI models converge toward shared representations as performance increases. The researchers called this "Platonic representation," referring to Plato's Allegory of the Cave.
The new study applies that idea to scientific models. It provides evidence that specialized AI systems for chemistry and biology may also converge toward a universal representation of matter.
Other research mentioned in the source article shows that convergence can have a downside. The recently published SDE benchmark for scientific research found that models often arrive at the same wrong answers on the hardest questions, with the best models showing the greatest agreement. An earlier study found a similar pattern in AI systems supervising other AI processes, where similarity in assessments can create blind spots and new failure modes.
Taken together, the message is measured rather than simple. Scientific AI models may be learning overlapping internal pictures of molecules, materials, and proteins. But the same field still has a hard generalization problem: when structures move far enough away from training data, agreement among models does not guarantee understanding.