Nick

Galaxy-Halo Priors Dominate Small-Scale Clustering Constraints

This study demonstrates that cosmological constraints from small-scale galaxy clustering (5-80 Mpc/h) are almost entirely contingent on prior assumptions about the galaxy-halo connection. Using LRG-like mocks from the AbacusSummit suite, the authors find only 25% of cosmologies are ruled out under conservative HOD assumptions, versus ~81% when HOD parameters are assumed known—a difference of orders of magnitude in χ² for many models. The result implies that without informative HOD priors, small-scale clustering cannot independently deliver tight cosmological constraints.

Unsupervised Feature-Based Inference Reveals Latent Parametric Variation in Time Series Without Model Fitting

This paper demonstrates that low-dimensional parametric variation in unknown generative processes can be inferred directly from time-series data using a large library of over 7,000 interpretable features, without specifying or fitting a model. The approach reconstructs underlying degrees of freedom across 13 simulated systems (including stochastic, nonlinear, and chaotic dynamics) and successfully extracts biologically meaningful components like sex and circadian rhythmicity from fruit fly movement data. This bridges the gap between interpretable dynamical theory and data-driven analysis of complex time series.

Bernstein Rigidity for Capillary Minimal Graphs: Linearity in Low Dimensions and Connectivity of Minimizers

Edelen, Li, and Zhu establish a Bernstein-type rigidity theorem for capillary minimal graphs in a half-space, proving that entire capillary minimal graphs must be linear (i.e., affine hyperplanes) in low dimensions, or more generally when no tangent cone at infinity splits off a vertical line. They further prove that the regular set of any entire capillary-minimizing hypersurface is connected, extending classical Bernstein-type results to the capillary (boundary contact angle) setting. The work also draws explicit connections to the one-phase Bernoulli free boundary problem, situating capillary geometry within the broader PDE landscape of free boundary regularity.

Projected Tree Enumeration Breaks the Output-Linear Barrier via Fast Matrix Multiplication

Bringmann, Fischer, and Wang present the first algorithms for enumerating projected trees (and projected hypergraphs) with polynomial preprocessing time (~O(n^17.42)) and polylogarithmic delay, shattering the prior lower bound of Ω(n^Ω(k) + t) that all previous algorithms obeyed. The key technical ingredients are fast rectangular and output-sensitive matrix multiplication, which the authors show is not merely a convenience but a necessity via fine-grained lower bounds. They also establish a generic equivalence between listing and enumeration under natural complexity assumptions, proving that any listing algorithm can be transformed into an enumeration algorithm with only logarithmic overhead in preprocessing and no overhead in delay.

Bayesian Belief Tracing Outperforms Vanilla LLMs at Simulating Human Persuasion Dynamics

PERSUASIONTRACE is a multi-turn experimental framework that moves persuasion evaluation beyond pre/post endpoint measurement to process-level fidelity, capturing how beliefs shift turn-by-turn during human-LLM dialogues. The framework annotates persuader turns along rhetorical dimensions (logos/pathos/ethos) and benchmarks simulated targets against real human belief dynamics. A key finding is that vanilla-prompted LLMs fail as human simulators, while a Bayesian-network target with an explicit latent belief state achieves near-human fidelity (81 vs. 80 on a human-likeness scale), compared to 64 for baseline LLM simulators. LLMs are shown to be broadly persuasive across topic types, modalities, and multi-turn contexts, raising safety implications for persuasive system optimization.

Ultra-Low-Noise Brillouin Hybrid Synthetic Laser for Sub-Hertz Lattice Clock Spectroscopy

Symmetry Defeats Auditing

The 'Right' Extension of Type-I Error to Data-Dependent Levels

Sequential Physics-Constrained Neural Operator Forward Modeling for the $\textit{Norne}$ Reservoir System

Unsupervised Semantic Segmentation Facilitates Model Understanding

BCI-sift: A Scikit-Learn-Compatible Toolbox That Improves Neural Decoding via Automated Feature Selection on HD ECoG Data

BCI-sift is a Python/scikit-learn-compatible toolbox that systematizes feature selection for Brain-Computer Interface (BCI) pipelines by integrating multiple optimization algorithms across electrode, temporal, and frequency dimensions. Validated on high-density ECoG recordings from 8 participants (64–128 electrodes over sensorimotor cortex) performing a 12-word speech task, it consistently identifies neurophysiologically plausible features—electrode selections aligned with known sensorimotor organization, time points clustered around speech production, and the high-frequency band flagged as most informative. The toolbox demonstrates that principled feature selection outperforms using the full feature set for classification accuracy, while also improving interpretability. Although validated on HD ECoG, the authors assert broad applicability to other BCI modalities.

CMB Large-Scale Anisotropies as a Proxy for Planck-Scale String Theory and Extra-Dimensional Topology

This paper leverages CMB temperature and polarization two-point correlation functions at large angular scales as an indirect observational window into high-scale string theory — a regime inaccessible to particle colliders. The authors analyze toroidal compactification signatures on the primordial power spectrum using Planck satellite data, finding a tentative indication of spatial-parity breaking consistent with six compactified extra dimensions, compactified at the GUT epoch prior to inflation. The framework is extended to B-mode polarization, positioning it as a future discriminator for next-generation CMB experiments.

Featureless Insulators Have Hidden Topological Structure: Berry Phase Textures Reveal Phase Diagram Complexity

Trivial ("featureless") insulating phases of non-interacting fermions are not as structurally inert as their name implies. Non-trivial topological families of states — including charge pumps and their generalizations — imprint topological textures onto phase diagrams, detectable via Berry phases and higher-dimensional analogs computed from non-abelian Chern-Simons integrals over momentum-parameter space. Singularities in these textures manifest as "diabolical points" (gap-closing loci) that obstruct contraction of topological families, and bulk-boundary correspondence produces robust boundary modes terminating at these bulk singularities. The dimensionality of the system critically governs the robustness of edge modes: in 2D and above they are generically stable under finite chemical potential, while in 1D they are "estranged" — appearing at distinct parameter values on opposite edges.

On the Stability of Growth in Structural Plasticity

Filter-then-Verify: A Multiphase GNN and ModernBERT Framework for Social Engineering Detection in Email Networks

Mechanistic Interpretability of EEG Foundation Models via Sparse Autoencoders

Quantum Monte Carlo fermion spectroscopy of a non-compact CP$^1$ model

How Much Can Gravitons Be Squeezed?

Quantile-Coupled Flow Matching for Distributional Reinforcement Learning

Node-Weighted Triangles: Faster and Simpler

Little Impact of ChatGPT Availability on High School Student Test Score Performance

OpenZL: Using Graphs to Compress Smaller and Faster

Approximations with Non-Symmetric Green's Kernels and their Application to Fractional Differential Equations