Position: Many generalization measures for deep learning are fragile

(2026)

Authors:

Shuofeng Zhang, Ard Louis

Predicting the topography of fitness landscapes from the structure of genotype鈥損henotype maps

Genetics Oxford University Press 232:4 (2026) iyag026

Authors:

Malvika Srivastava, Ard A Louis, Nora S Martin

Abstract:

Ruggedness鈥攖he prevalence of fitness peaks鈥攁nd navigability鈥攖he existence of fitness-increasing paths to a target鈥攁re key factors affecting evolution on fitness landscapes. Here, we analyze these properties in landscapes that inherit biophysically grounded genotype鈥損henotype (GP) maps. By assuming a random phenotype-fitness assignment as a baseline, the structure of the GP maps is included without imposing further fitness correlations. We show analytically that the expected ruggedness can be predicted from two quantities: the sizes of neutral components (NCs)鈥攎utationally connected genotype sets with the same phenotype鈥攁nd their evolvabilities, defined as the number of distinct phenotypes among the NC鈥檚 mutational neighbors. Other features鈥攕uch as robustness鈥攊nfluence ruggedness only indirectly via correlations with evolvability. Numerical results across diverse GP maps confirm that NC size and evolvability alone suffice to predict both the mean prevalence and heights of peaks. These calculations also provide new insights: Under random phenotype-fitness assignment, peaks arising from high-evolvability NCs have higher expected fitness than those from low-evolvability NCs. Thus, when evolvability correlates positively with NC size, the formation of large low-fitness peaks is impeded. We further derive an approximate scaling law for the minimal average evolvability required for navigability. Our framework applies broadly across GP maps, providing general insight into when and why fitness landscapes are expected to be rugged or navigable.

$q$-state Potts ice

ArXiv 2601.22909 (2026)

Authors:

Mark Potts, SA Parameswaran, Roderich Moessner

Bridging Elastic and Active Turbulence

(2026)

Authors:

Vedad Dzanic, Sumesh P Thampi, Julia M Yeomans

Emergent interacting phases in the strong-coupling limit of twisted M-valley moir茅 systems: Application to SnSe 2

Physical Review B 114:5 (2026) 1-12

Authors:

MR Li, D Calugaru, Y Jiang, H Pi, A Fischer, H Schl枚mer, L Klebl, Z Xia, MG Vergniory, DM Kennes, KF Mak, J Shan, SA Parameswaran, H Yao, BA Bernevig, H Hu

Abstract:

We establish twisted SnSe 2 as a tunable platform for simulating dimension-dependent correlated physics, distinct from conventional K-valley moir茅 systems. By constructing interacting Wannier models, we show that the stacking configuration dictates the effective lattice geometry. In AA-stacked bilayers, a momentumspace nonsymmorphic symmetry constrains the single-particle hopping within each valley to be effectively one-dimensional while still allowing fully two-dimensional interactions, thereby giving rise to an effective quasi-one-dimensional system. This dimensional reduction stabilizes exotic phases including dimerized states with finite residual entropy, valence bond solids, and quantum paramagnetism. Conversely, AB stacking maps to a frustrated Kagome lattice; here, strong interactions drive the emergence of a classical spin liquid. The high tunability of this moir茅 system, which allows control over both the filling and interaction strength (via twist angle), renders twisted SnSe 2 a versatile platform for realizing a wide range of exotic correlated quantum phases.