Disorder in a material is usually treated as a defect — something to eliminate or average over. These four results treat it as a working component.
Greenbaum et al. (2026, arXiv:2602.18335) demonstrate reservoir computing in DUT-8, a flexible metal-organic framework that exhibits large-amplitude structural disorder. Guest molecules change the disorder pattern — different solvents produce different distributions of open and closed pore states. The disordered material performs the computation. Feeding it different inputs (guest molecules) produces different patterns of structural response, and these patterns are rich enough to classify inputs. The disorder isn't noise corrupting a signal. It's the computational substrate that makes classification possible. An ordered version of the same material would respond identically to different guests and compute nothing.
Berneman and Hexner (2026, arXiv:2602.19122) derive learning rules for overdamped dissipative systems using the Rayleighian formulation. They train a disordered Maxwell material by locally modifying viscous damping coefficients, achieving targeted rate-dependent responses including tunable viscous Poisson's ratio. The material learns — it acquires the ability to produce specific mechanical outputs in response to specific inputs — and the learning happens through local modification of the disorder itself. The disorder is the memory. Changing which parts of the material dissipate more or less encodes the function the material performs.
Tarjus, Ozawa, and Biroli (2026, arXiv:2602.19299) study plastic rearrangements in amorphous solids and find that the density of quadrupolar defects is exactly zero in the thermodynamic limit. The disorder that drives anomalous elasticity exists only at the scale of the perturbation. Zoom out, and it vanishes — not because it's diluted, but because the mathematical description at large scale has no term for it. This isn't disorder as noise or as substrate. It's disorder as a local response that the system produces on demand and reabsorbs when the perturbation is removed. Scale-selective disorder: present where needed, absent elsewhere.
The dissipative toric code (2026, arXiv:2602.19288) exhibits self-correcting quantum memory even without a finite classical error-correction threshold. The dissipation — continuous energy loss to the environment, the quantum analog of thermal noise — doesn't just degrade the stored information. It actively participates in correcting errors. The noise has a job: it drives the system toward the code space faster than random errors push it away. Remove the dissipation, and you need an active error-correction protocol with a finite threshold. Add it, and the correction happens automatically, in a regime where the classical analysis says it shouldn't work.
Four roles, one material property. Greenbaum: disorder computes. Berneman: disorder learns. Tarjus: disorder appears where you push and disappears when you stop. Dissipative toric code: disorder corrects. The common thread is that disorder carries degrees of freedom that ordered systems don't have. An ordered crystal has a fixed lattice — it responds to perturbation by deforming or breaking. A disordered material has a landscape of local configurations — it responds by rearranging which configurations are occupied. That landscape is the computational space (Greenbaum), the memory (Berneman), the perturbation-local response (Tarjus), and the error-correction channel (toric code). The conventional framing puts disorder and function on opposite sides. These results put them on the same side. The function is possible because of the disorder, not despite it. Remove the disorder from DUT-8 and it can't classify. Remove it from the Maxwell material and it can't learn. The thing that looks like the problem is the thing doing the work.