Machine learning models that fit data beautifully while quietly violating the laws of physics are not a niche annoyance. They are a structural risk. A model that predicts fluid flow but creates mass from nothing, or a climate simulation that drifts into unphysical temperature regimes, can look excellent on validation curves and still be useless for real decisions. The CL4SM session at the ELLIs UnConference, ahead of NeurIPS 2026 in Paris, is asking the right question: how do we make physical correctness a feature of the model itself, not an afterthought?
This is the same tension we have explored elsewhere in our coverage of AI systems that optimize for one metric at the expense of another. When Smarter resource use: how batching transforms LLM performance showed that careful batching can cut costs without degrading output, the lesson was about respecting operational constraints. CL4SM pushes that idea into the scientific domain: accuracy without constraint awareness is not really accuracy. Similarly, Explore how hierarchical routing cuts memory while preserving long-context accuracy demonstrates that memory efficiency and accuracy can coexist when the architecture is designed around the constraint. The same principle applies here, except the constraint is a conservation law rather than a memory budget.
The call for contributions is deliberately broad, covering soft penalties, hard architectural guarantees, projection layers, and probabilistic constraints. That breadth is a strength. It acknowledges that there is no single right way to enforce physics. Some problems will tolerate a penalty term that gently discourages violations. Others, like safety-critical control systems, need hard guarantees built into the architecture. The session also invites negative results and early-stage ideas, which is refreshing. Too many venues only want polished success stories, but the field needs to know what does not work and why.
For practitioners, the practical takeaway is straightforward: before you trust a model's predictions, ask what it is allowed to violate. If the answer is "nothing, because it was trained on enough data," you are likely to be disappointed. Data alone does not teach conservation of energy. The CL4SM session is not just an academic exercise. It is a signal that the community is moving toward models that are trustworthy because they are constrained, not because they are confident. We also note that recently published papers from top venues are welcome, which lowers the barrier for sharing mature work without re-packaging it.
The specific detail to watch is the deadline: October 30, 2026. That gives researchers a clear runway to shape the conversation before the December session in Paris. If you have a model that respects symmetries or a benchmark that exposes physical violations in popular architectures, this is the venue to bring it. The field does not need more models that fit curves. It needs models that respect the world they are trying to represent.