arXiv subjects

Louie Hong Yao

Publications and source records attributed to Louie Hong Yao.

13 recordsLinked to original sources

arXiv2026

Cross-Lingual Steering for Figurative Language Generation

Cross-lingual activation steering has been studied mainly for language-agnostic properties such as factual knowledge and reasoning. Its transferability remains unclear for figurative language, whose expressions are often tied to a particular language. In this work, we study cross-lingual steering of figurative language: whether steering vectors learned in one language can affect generation in another. We first show that within a single language, steering effectively increases the ratio of the intended figure of speech, especially for metaphor and idiom. Then under the cross-lingual setting, across 408 comparisons spanning four multilingual models, six languages, and four figures of speech, steering increases the figurative ratio from the unsteered 12.1% to 20.4%. To understand the mechanism, we decompose each language's own steering vector into the mean of steering vectors from other languages and a residual, and show that this mean direction matches or outperforms the language's own vector while the residual steers worse. This suggests that, despite the highly language-dependent nature of figurative expressions, a partially shared direction exists in the representation space that enables figurative knowledge to transfer across languages.

cs.CL↗

arXiv2026

Rhetorical Questions in LLM Representations: A Linear Probing Study

Rhetorical questions are asked not to seek information but to persuade or signal stance. How large language models internally represent them remains unclear. We analyze rhetorical questions in LLM representations using linear probes on two social-media datasets with different discourse contexts, and find that rhetorical signals emerge early and are most stably captured by last-token representations. Rhetorical questions are linearly separable from information-seeking questions within datasets, and remain detectable under cross-dataset transfer, reaching AUROC around 0.7-0.8. However, we demonstrate that transferability does not simply imply a shared representation. Probes trained on different datasets produce different rankings when applied to the same target corpus, with overlap among the top-ranked instances often below 0.2. Qualitative analysis shows that these divergences correspond to distinct rhetorical phenomena: some probes capture discourse-level rhetorical stance embedded in extended argumentation, while others emphasize localized, syntax-driven interrogative acts. Together, these findings suggest that rhetorical questions in LLM representations are encoded by multiple linear directions emphasizing different cues, rather than a single shared direction.

cs.CL↗

arXiv2026

MinkowskiPE: Minkowski Positional Encoding for Spatiotemporal Perception

Modeling spatiotemporal coupling is a key challenge in building physical intelligence across scales, from microscopic to macroscopic. Existing models capture such structure broadly through physics-motivated dynamical formulations or learning-motivated architectures. The former provide stronger priors but may constrain flexibility, whereas the latter are more flexible but leave the spatiotemporal coupling largely implicit. We therefore seek an approach that combines flexible learning with an explicit geometric bias for jointly modeling time and space. To this end, we propose Minkowski Positional Encoding (MinkowskiPE), which uses joint temporal and spatial coordinates to parameterize Lorentz transformations applied to query and key features. With MinkowskiPE, the query-key attention score depends on position only through the relative spacetime displacement between the two tokens and is therefore invariant to global translation of the coordinates. This paradigm retains the standard dot-product attention interface and remains compatible with efficient attention implementations. We evaluate MinkowskiPE on microscopic molecular dynamics and macroscopic video prediction tasks, achieving the best results on all nine multi-trajectory molecular evaluations and reducing KTH video-prediction MSE by 9.9% relative to the best baseline while using roughly one-tenth as many parameters.

cs.LG↗

arXiv2026

The Linear Representation Hypothesis Needs a Group Action

To make claims about representations that generalize beyond a particular trained model, we need to specify when two representations should count as equivalent. The Linear Representation Hypothesis is often discussed without making this equivalence explicit. Different notions of equivalence preserve different structures, so metrics, probes, and interventions that appear to study the same representation may in fact correspond to different hypotheses. We therefore argue that the Linear Representation Hypothesis is not one hypothesis but a family of claims distinguished by representation equivalence. We formalize this idea using group actions, specifying the representation object, the procedure that produces it, and the property ultimately asserted, while accounting for equivalences imposed by the model architecture. This framework clarifies how assumptions can change across metrics, reading points, and analysis stages, and we use it to audit common representation quantities and recent interpretability analyses.

cs.LG↗

arXiv2026

JE-IRT: A Geometric Lens on LLM Abilities through Joint Embedding Item Response Theory

Standard LLM evaluation practices compress diverse abilities into single scores, obscuring their inherently multidimensional nature. We present JE-IRT, a geometric item-response framework that embeds both LLMs and questions in a shared space. For question embeddings, the direction encodes semantics and the norm encodes difficulty, while correctness on each question is determined by the geometric interaction between the model and question embeddings. This geometry replaces a global ranking of LLMs with topical specialization and enables smooth variation across related questions. Building on this framework, our experimental results reveal that out-of-distribution behavior can be explained through directional alignment, and that larger norms consistently indicate harder questions. Moreover, JE-IRT naturally supports generalization: once the space is learned, new LLMs are added by fitting a single embedding. The learned space further reveals an LLM-internal taxonomy that only partially aligns with human-defined subject categories. We also show that simple linear probes of the embedding space recover cross-subject ability directions, such as an arithmetic axis that highlights quantitatively demanding questions in seemingly distant subjects like virology and global facts. JE-IRT thus establishes a unified and interpretable geometric lens that connects LLM abilities with the structure of questions, offering a distinctive perspective on model evaluation and generalization.

cs.AI↗

arXiv2026

A Minimal Model of Representation Collapse: Frustration, Stop-Gradient, and Dynamics

Self-supervised representation learning is central to modern machine learning because it extracts structured latent features from unlabeled data and enables robust transfer across tasks and domains. However, it can suffer from representation collapse, a widely observed failure mode in which embeddings lose discriminative structure and distinct inputs become indistinguishable. To understand the mechanisms that drive collapse and the ingredients that prevent it, we introduce a minimal embedding-only model whose gradient-flow dynamics and fixed points can be analyzed in closed form, using a classification-representation setting as a concrete playground where collapse is directly quantified through the contraction of label-embedding geometry. We illustrate that the model does not collapse when the data are perfectly classifiable, while a small fraction of frustrated samples that cannot be classified consistently induces collapse through an additional slow time scale that follows the early performance gain. Within the same framework, we examine collapse prevention by adding a shared projection head and applying stop-gradient at the level of the training dynamics. We analyze the resulting fixed points and develop a dynamical mean-field style self-consistency description, showing that stop-gradient enables non-collapsed solutions and stabilizes finite class separation under frustration. We further verify empirically that the same qualitative dynamics and collapse-prevention effects appear in a linear teacher-student model, indicating that the minimal theory captures features that persist beyond the pure embedding setting.

cond-mat.dis-nn↗

arXiv2026

Towards Robust Evaluation of Visual Activity Recognition: Resolving Verb Ambiguity with Sense Clustering

Evaluating visual activity recognition systems is challenging due to inherent ambiguities in verb semantics and image interpretation. When describing actions in images, synonymous verbs can refer to the same event (e.g., brushing vs. grooming), while different perspectives can lead to equally valid but distinct verb choices (e.g., piloting vs. operating). Standard exact-match evaluation, which relies on a single gold answer, fails to capture these ambiguities, resulting in an incomplete assessment of model performance. To address this, we propose a vision-language clustering framework that constructs verb sense clusters, providing a more robust evaluation. Our analysis of the imSitu dataset shows that each image maps to around four sense clusters, with each cluster representing a distinct perspective of the image. We evaluate multiple activity recognition models and compare our cluster-based evaluation with standard evaluation methods. Additionally, our human alignment analysis suggests that the cluster-based evaluation better aligns with human judgments, offering a more nuanced assessment of model performance.

cs.CL↗

arXiv2025

Stochastic resetting in discrete-time quantum dynamics: steady states and correlations in few-qubit systems

Time evolution in several classes of quantum devices is generated through the application of quantum gates. Resetting is a critical technological feature in these systems allowing for mid-circuit measurement and complete or partial qubit reset. The possibility of realizing discrete-time reset dynamics on quantum computers makes it important to investigate the steady-state properties of such dynamics. Here, we explore the behavior of generic discrete-time unitary dynamics interspersed by random reset events. For Poissonian resets, we compute the stationary state of the process and demonstrate, by taking a weak-reset limit, the existence of "resonances" in the quantum gates, allowing for the emergence of steady state density matrices which are not diagonal in the eigenbasis of the generator of the unitary gate. Such resonances are a genuine discrete-time feature and impact on quantum and classical correlations even beyond the weak-reset limit. Furthermore, we consider non-Poissonian reset processes and explore conditions for the existence of a steady state. We show that, when the reset probability vanishes sufficiently rapidly with time, the system does not approach a steady state. Our results highlight key differences between continuous-time and discrete-time stochastic resetting and may be useful to engineer states with controllable correlations on existing devices.

quant-ph↗

arXiv2024

Digital Quantum Simulation of Reaction-Diffusion Systems on Lattice

The quantum computer offers significant advantages in simulating physical systems, particularly those with exponentially large state spaces, such as quantum systems. Stochastic reaction-diffusion systems, characterized by their stochastic nature, also exhibit exponential growth in the dimension of the state space, posing challenges for simulation at a probability distribution level. We explore the quantum simulation of stochastic reaction-diffusion systems on a digital quantum computer, directly simulating the system at the master equation level. Leveraging a spin representation of the system, we employ Trotterization and probabilistic imaginary time evolution (PITE) to simulate the probability distribution directly. We illustrate this approach through four diverse examples, ranging from simple single-lattice site generation-annihilation processes to a system featuring active-absorbing phase transition.

quant-ph↗

arXiv2023

Coined quantum walks on the line: Disorder, entanglement, and localization

Disorder in coined quantum walks generally leads to localization. We investigate the influence of the localization on the entanglement properties of coined quantum walks. Specifically, we consider quantum walks on the line and explore the effects of quenched disorder in the coin operations. After confirming that our choice of disorder localizes the walker, we study how the localization affects the properties of the coined quantum walk. We find that the mixing properties of the walk are altered nontrivially with mixing being improved at short time scales. Special focus is given to the influence of coin disorder on the properties of the quantum state and the coin-walker entanglement. We find that disorder alters the quantum state significantly even when the walker probability distribution is still close to the nondisordered case. We observe that, generically, coin disorder decreases the coin-walker entanglement and that the localization leaves distinct traces in the entanglement entropy and the entanglement negativity of the coined quantum walk.

quant-ph↗

arXiv2023

Perturbative Field-Theoretical Analysis of Three-Species Cyclic Predator-Prey Models

We apply a perturbative Doi--Peliti field-theoretical analysis to the stochastic spatially extended symmetric Rock-Paper-Scissors (RPS) and May--Leonard (ML) models, in which three species compete cyclically. Compared to the two-species Lotka--Volterra predator-prey (LV) model, according to numerical simulations, these cyclical models appear to be less affected by intrinsic stochastic fluctuations. Indeed, we demonstrate that the qualitative features of the ML model are insensitive to intrinsic reaction noise. In contrast, and although not yet observed in numerical simulations, we find that the RPS model acquires significant fluctuation-induced renormalizations in the perturbative regime, similar to the LV model. We also study the formation of spatio-temporal structures in the framework of stability analysis and provide a clearcut explanation for the absence of spatial patterns in the RPS system, whereas the spontaneous emergence of spatio-temporal structures features prominently in the LV and the ML models.

cond-mat.stat-mech↗

arXiv2022

Critical Dynamics of the Antiferromagnetic $O(3)$ Nonlinear Sigma Model with Conserved Magnetization

We study the near-equilibrium critical dynamics of the $O(3)$ nonlinear sigma model describing isotropic antiferromagnets with non-conserved order parameter reversibly coupled to the conserved total magnetization. To calculate response and correlation functions, we set up a description in terms of Langevin stochastic equations of motion, and their corresponding Janssen--De~Dominicis response functional. We find that in equilibrium, the dynamics is well-separated from the statics, at least to one-loop order in a perturbative treatment with respect to the static and dynamical nonlinearities. Since the static nonlinear sigma model must be analyzed in a dimensional $d = 2 + \varepsilon$ expansion about its lower critical dimension $d_\textrm{lc} = 2$, whereas the dynamical mode-coupling terms are governed by the upper critical dimension $d_c = 4$, a simultaneous perturbative dimensional expansion is not feasible, and the reversible critical dynamics for this model cannot be accessed at the static critical renormalization group fixed point. However, in the coexistence limit addressing the long-wavelength properties of the low-temperature ordered phase, we can perform an $ε= 4 - d$ expansion near $d_c$. This yields anomalous scaling features induced by the massless Goldstone modes, namely sub-diffusive relaxation for the conserved magnetization density with asymptotic scaling exponent $z_Γ= d - 2$ which may be observable in neutron scattering experiments. Intriguingly, if initialized near the critical point, the renormalization group flow for the effective dynamical exponents recovers their universal critical values $z_c = d / 2$ in an intermediate crossover region.

cond-mat.stat-mech↗

arXiv2020

Social distancing and epidemic resurgence in agent-based Susceptible-Infectious-Recovered models

Once an epidemic outbreak has been effectively contained through non-pharmaceutical interventions, a safe protocol is required for the subsequent release of social distancing restrictions to prevent a disastrous resurgence of the infection. We report individual-based numerical simulations of stochastic susceptible-infectious-recovered model variants on four distinct spatially organized lattice and network architectures wherein contact and mobility constraints are implemented. We robustly find that the intensity and spatial spread of the epidemic recurrence wave can be limited to a manageable extent provided release of these restrictions is delayed sufficiently (for a duration of at least thrice the time until the peak of the unmitigated outbreak) and long-distance connections are maintained on a low level (limited to less than five percent of the overall connectivity).

physics.soc-ph↗