arXiv subjects

Ethan Smith

Publications and source records attributed to Ethan Smith.

At least 19 recordsLinked to original sources

arXiv2026

RAG Collapse: LLM Responses Collapse When Retrieved Documents Are Self-Authored

LLM responses are based on the internet (via training or RAG), and AI is now used to generate a significant amount of content online (Paredes et al., 2026), creating the potential for a self-reinforcing feedback loop. Prior work has shown that when LLMs are recursively trained on their own output, they experience model collapse (Shumailov et al., 2024): responses become less diverse, and eventually no longer resemble the original training data. In this paper, we show that a similar collapse occurs if LLM-based AI systems retrieve references they authored using a search tool. We call this RAG collapse. We conduct extensive experiments with three types of simulations of AI systems retrieving references they generated, using three model families, and 1,019 information-seeking prompts, totaling 1,528 simulations and over one million LLM API calls, and find that 79.6% (1,216/1,528) of simulations end in collapse. Surprisingly, even a single self-authored reference can trigger collapse because the LLM disproportionately cites its own content. This self-bias persists even after controlling for reference quality.

cs.CL↗

arXiv2026

Learning in Curved Weight Space:Exponential-Linear Weight Reparameterization for Improved Optimization

Many neural networks operations have a multiplicative nature rather than additive: halving or doubling a norm are analogous relatively but require unequal optimization distances when taking linear steps. Adaptive optimizers such as Adam normalize updates per coordinate, but update steps remain additive; weights with very different magnitudes receive similarly sized absolute changes, producing very different relative perturbations. We introduce \textbf{\method} (\textbf{\methodshort}), a weight reparameterization for neural networks that combines a sign-aware symmetric-exponential pathway with an identity-like linear pathway. The symmetric-exponential pathway is near-linear for small raw weights but increasingly curved at larger magnitudes. Additive updates in logarithmic space map to magnitude-proportional changes in effective weight space. The linear pathway provides a direct route through the transform that we hypothesize stabilizes optimization, while learnable scale, curvature, and offset parameters control balance between pathways and the curvature of the exponential pathway. These components create a curved parameter-space geometry that empirically improves speed of loss descent over standard linear parameterization. We also identify a useful \emph{mismatched initialization}: raw weights are chosen so a symmetric version of the transform matches Xavier statistics, but training uses an asymmetric forward transform that leaves positive weights at full strength while making negative weights smaller in magnitude; in small-model ablations, this improves early optimization and may act as a form of symmetry breaking. We train transformers on OpenWebText over nine width$\times$depth configurations, \methodshort reaches matched validation loss in 1.32--1.49$\times$ fewer training steps, with the largest widths seeing the biggest gains.

cs.LG↗

arXiv2026

DiffusionBench: On Holistic Evaluation of Diffusion Transformers

Diffusion transformer (DiT) research on image generation has converged to a single evaluation setup: class-conditional generation on ImageNet. While methods improve the FID and related metrics, it is increasingly unclear whether they reflect real progress in generative modeling. The natural alternative, i.e., text-to-image (T2I) generation, is perceived as too costly or inconvenient to train and evaluate and is often skipped. We argue that this perception no longer holds. We introduce NanoGen, a unified DiT training and evaluation framework. NanoGen matches state-of-the-art DiT baselines on ImageNet and, with 12 lines of configuration change, also trains competitive text-to-image models. It currently supports RAE, VAE, pixel-space, and MeanFlow diffusion methods under both ImageNet and T2I setups. Under NanoGen, training T2I requires comparable compute to ImageNet. After training 21 latent diffusion models with NanoGen, we observe that method ranking shows no strong correlation between ImageNet and T2I generation: Pearson correlation is between -0.377 and -0.580 across three metrics. This suggests that a method which improves class-conditional ImageNet FID may show no corresponding improvement on T2I, clearly indicating the necessity of evaluating DiTs on both tasks. To this end, we summarize ImageNet and text-to-image results, which yields DiffusionBench, a holistic benchmark for DiT research. We recommend reporting DiffusionBench in place of ImageNet alone: methods that improve DiffusionBench are more likely to reflect broader progress.

cs.CV↗

arXiv2026

MRT: Masked Region Transformer for Layered Image Generation and Editing at Scale

Layered image generation and editing is a fundamental capability that enables layer-wise reuse, editing, and composition of generated visual content, analogous to word-level editing in natural language. Despite its importance, this remains an underexplored area at scale. To address this gap, we present MRT, a 20B-parameter masked region diffusion model tailored for multi-layer transparent image generation and editing, trained on over 10M multilingual design samples spanning diverse aspect ratios and textual prompts. To fully leverage this scale, we make two key technical contributions. First, we unify three complementary tasks including text-to-layers, image-to-layers, and layers-to-layers within a shared masked region diffusion framework, where selective token masking enables flexible layer-wise generation and editing. Second, to enable overflow layer generation, we introduce an overflow-aware canvas layer that handles boundary inconsistencies and supports semi-transparent background synthesis, enabling complete editable layers extending beyond visible canvas boundaries. Additionally, we apply diffusion distillation to achieve 8-step, real-time multi-layer generation with minimal quality degradation. Extensive experiments demonstrate that our framework substantially outperforms prior state-of-the-art approaches, including various commercial systems, across all three tasks, establishing a new benchmark for multi-layer transparent image generation. Notably, our model significantly outperforms the concurrent Qwen-Image-Layered model in image-to-layers quality according to user-study results, while achieving 10-100\times faster inference and reducing activation GPU memory consumption by 50-90\% during image-to-layer inference.

cs.CV↗

arXiv2026

Weak-DMD: A Galerkin approach to the problem of noise in the Dynamic Mode Decomposition algorithm

Dynamic Mode Decomposition (DMD) is a data-driven method for approximating the spatiotemporal modes of a system. The eigenvectors and eigenvalues of the system are approximated from a series of time-snapshots of the state variables. The standard formulation of DMD is subject to strict assumptions concerning the time-spacing of the snapshots and is biased by measurement noise. Variations on the method have been developed to address these shortcomings, but the problem is still open. Motivated by the effectiveness of Galerkin methods in the field of model discovery, a weak formulation of DMD is presented, weak-DMD. Weak-DMD precludes timestep considerations and also filters noise. Results for two nuclear engineering applications and the flow of fluid past a cylinder are given and compared with a state of the art DMD algorithm.

cs.CE↗

arXiv2026

NOBLE: Accelerating Transformers with Nonlinear Low-Rank Branches

We introduce NOBLE (Nonlinear lOw-rank Branch for Linear Enhancement), an architectural augmentation that adds nonlinear low-rank branches to transformer linear layers. Unlike LoRA and other parameter-efficient fine-tuning (PEFT) methods, NOBLE is designed for pretraining from scratch. The branch is a permanent part of the architecture as opposed to an adapter for finetuning on top of frozen weights. The branch computes {\sigma}(xWdown)Wup where {\sigma} is a learnable nonlinearity. We evaluate several activation functions and find that CosNet, a two-layer cosine nonlinearity with learnable frequency and phase with a linear projection in between them in the bottleneck space, performs best. NOBLE achieves substantial improvements with minimal overhead: up to 1.47x step speedup to reach baseline eval loss (up to 32% fewer training steps), with as low as 4% additional parameters and 7% step time overhead, resulting in up to 1.22x net wallclock speedup. Experiments on LLMs (250M and 1.5B parameters), BERT, VQGAN, and ViT consistently show improved training efficiency. We identify one caveat: Mixup/CutMix augmentation interferes with NOBLE's benefits in Imagenet classification along with other stochastic augmentations, but when disabled, ViT also improves. This discrepancy is possibly explained by regularization techniques that encourage smoother fits to the target function while NOBLE may specialize more in sharper aspects of the target function.

cs.LG↗

arXiv2026

From Tokens to Numbers: Continuous Number Modeling for SVG Generation

For certain image generation tasks, vector graphics such as Scalable Vector Graphics (SVGs) offer clear benefits such as increased flexibility, size efficiency, and editing ease, but remain less explored than raster-based approaches. A core challenge is that the numerical, geometric parameters, which make up a large proportion of SVGs, are inefficiently encoded as long sequences of tokens. This slows training, reduces accuracy, and hurts generalization. To address these problems, we propose Continuous Number Modeling (CNM), an approach that directly models numbers as first-class, continuous values rather than discrete tokens. This formulation restores the mathematical elegance of the representation by aligning the model's inputs with the data's continuous nature, removing discretization artifacts introduced by token-based encoding. We then train a multimodal transformer on 2 million raster-to-SVG samples, followed by fine-tuning via reinforcement learning using perceptual feedback to further improve visual quality. Our approach improves training speed by over 30% while maintaining higher perceptual fidelity compared to alternative approaches. This work establishes CNM as a practical and efficient approach for high-quality vector generation, with potential for broader applications. We make our code available http://github.com/mikeogezi/CNM.

cs.LG↗

arXiv2025

A Momentum-Resolved X-ray Thomson Scattering Benchmark of Electronic-Response Models in Warm Dense Aluminium

The robust diagnosis of conditions generated in warm dense matter (WDM) experiments remains a persistent challenge. Here we describe the measurement of shock-compressed aluminium at 50 GPa with angle-resolved femtosecond x-ray Thomson scattering (XRTS) over a wide range of scattering wavevectors at the European XFEL. The measured plasmon dispersion and line shape show that the de facto standard approach for analysing XRTS spectra, based on uniform-electron-gas models, systematically overestimates the resonance energy by up to 8 eV. We present an ab initio approach that agrees within the experimental uncertainty and demonstrates that accounting for shock-induced disorder is critical for interpreting shock-compressed systems, providing evidence that ab initio treatments are required for reliable XRTS inference in warm dense aluminium.

physics.plasm-ph↗

arXiv2025

Consistent Solutions of the Radiation Diffusion Equation in Spherical and Cylindrical Geometries

We have extended the radiation diffusion model of Hammer and Rosen to diverging spherical and cylindrical geometries. The effect of curvilinear geometry on the supersonic, expanding wavefront increases as the internal radius of a spherical or cylindrical shell approaches zero. Small spherical geometries are important for modeling systems at the size scale of ICF capsules, at these scales existing quasi-analytic models for planar geometry significantly disagree with the results of simulation. With this method, the benefits of rapid iteration can be applied to common spherical systems at much smaller length scales. We present comparisons between numerical diffusion solutions and the analytic model to give ranges of applicability for the model.

physics.comp-ph↗

arXiv2024

LoRA Diffusion: Zero-Shot LoRA Synthesis for Diffusion Model Personalization

Low-Rank Adaptation (LoRA) and other parameter-efficient fine-tuning (PEFT) methods provide low-memory, storage-efficient solutions for personalizing text-to-image models. However, these methods offer little to no improvement in wall-clock training time or the number of steps needed for convergence compared to full model fine-tuning. While PEFT methods assume that shifts in generated distributions (from base to fine-tuned models) can be effectively modeled through weight changes in a low-rank subspace, they fail to leverage knowledge of common use cases, which typically focus on capturing specific styles or identities. Observing that desired outputs often comprise only a small subset of the possible domain covered by LoRA training, we propose reducing the search space by incorporating a prior over regions of interest. We demonstrate that training a hypernetwork model to generate LoRA weights can achieve competitive quality for specific domains while enabling near-instantaneous conditioning on user input, in contrast to traditional training methods that require thousands of steps.

cs.LG↗

arXiv2024

Transit drivers' reflections on the benefits and harms of eye tracking technology

Eye tracking technology offers great potential for improving road safety. It is already being built into vehicles, namely cars and trucks. When this technology is integrated into transit service vehicles, employees, i.e., bus drivers, will be subject to being eye tracked on their job. Although there is much research effort advancing algorithms for eye tracking in transportation, less is known about how end users perceive this technology, especially when interacting with it in an employer-mandated context. In this first study of its kind, we investigated transit bus operators' perceptions of eye tracking technology. From a methodological perspective, we introduce a mixed methods approach where participants experience the technology first-hand and then reflect on their experience while viewing a playback of the recorded data. Thematic analysis of the interview transcripts reveals interesting potential uses of eye tracking in this work context and surfaces transit operators' fears and concerns about this technology.

cs.HC↗

arXiv2024

EZIGen: Enhancing zero-shot personalized image generation with precise subject encoding and decoupled guidance

Zero-shot personalized image generation models aim to produce images that align with both a given text prompt and subject image, requiring the model to incorporate both sources of guidance. Existing methods often struggle to capture fine-grained subject details and frequently prioritize one form of guidance over the other, resulting in suboptimal subject encoding and imbalanced generation. In this study, we uncover key insights into overcoming such drawbacks, notably that 1) the choice of the subject image encoder critically influences subject identity preservation and training efficiency, and 2) the text and subject guidance should take effect at different denoising stages. Building on these insights, we introduce a new approach, EZIGen, that employs two main components: leveraging a fixed pre-trained Diffusion UNet itself as subject encoder, following a process that balances the two guidances by separating their dominance stage and revisiting certain time steps to bootstrap subject transfer quality. Through these two components, EZIGen, initially built upon SD2.1-base, achieved state-of-the-art performances on multiple personalized generation benchmarks with a unified model, while using 100 times less training data. Moreover, by further migrating our design to SDXL, EZIGen is proven to be a versatile model-agnostic solution for personalized generation. Demo Page: zichengduan.github.io/pages/EZIGen/index.html

cs.CV↗

arXiv2024

Shock-driven amorphization and melt in Fe$_2$O$_3$

We present measurements on Fe$_2$O$_3$ amorphization and melt under laser-driven shock compression up to 209(10) GPa via time-resolved in situ x-ray diffraction. At 122(3) GPa, a diffuse signal is observed indicating the presence of a non-crystalline phase. Structure factors have been extracted up to 182(6) GPa showing the presence of two well-defined peaks. A rapid change in the intensity ratio of the two peaks is identified between 145(10) and 151(10) GPa, indicative of a phase change. Present DFT+$U$ calculations of temperatures along Fe$_2$O$_3$ Hugoniot are in agreement with SESAME 7440 and indicate relatively low temperatures, below 2000 K, up to 150 GPa. The non-crystalline diffuse scattering is thus consistent with the - as yet unreported - shock amorphization of Fe$_2$O$_3$ between 122(3) and 145(10) GPa, followed by an amorphous-to-liquid transition above 151(10) GPa. Upon release, a non-crystalline phase is observed alongside crystalline $\alpha$-Fe$_2$O$_3$. The extracted structure factor and pair distribution function of this release phase resemble those reported for Fe$_2$O$_3$ melt at ambient pressure.

cond-mat.mtrl-sci↗

arXiv2024

High Energy Density Radiative Transfer in the Diffusion Regime with Fourier Neural Operators

Radiative heat transfer is a fundamental process in high energy density physics and inertial fusion. Accurately predicting the behavior of Marshak waves across a wide range of material properties and drive conditions is crucial for design and analysis of these systems. Conventional numerical solvers and analytical approximations often face challenges in terms of accuracy and computational efficiency. In this work, we propose a novel approach to model Marshak waves using Fourier Neural Operators (FNO). We develop two FNO-based models: (1) a base model that learns the mapping between the drive condition and material properties to a solution approximation based on the widely used analytic model by Hammer & Rosen (2003), and (2) a model that corrects the inaccuracies of the analytic approximation by learning the mapping to a more accurate numerical solution. Our results demonstrate the strong generalization capabilities of the FNOs and show significant improvements in prediction accuracy compared to the base analytic model.

physics.comp-ph↗

arXiv2024

ToDo: Token Downsampling for Efficient Generation of High-Resolution Images

Attention mechanism has been crucial for image diffusion models, however, their quadratic computational complexity limits the sizes of images we can process within reasonable time and memory constraints. This paper investigates the importance of dense attention in generative image models, which often contain redundant features, making them suitable for sparser attention mechanisms. We propose a novel training-free method ToDo that relies on token downsampling of key and value tokens to accelerate Stable Diffusion inference by up to 2x for common sizes and up to 4.5x or more for high resolutions like 2048x2048. We demonstrate that our approach outperforms previous methods in balancing efficient throughput and fidelity.

cs.CV↗

arXiv2024

Resonant inelastic x-ray scattering in warm-dense Fe compounds beyond the SASE FEL resolution limit

Resonant inelastic x-ray scattering (RIXS) is a widely used spectroscopic technique, providing access to the electronic structure and dynamics of atoms, molecules, and solids. However, RIXS requires a narrow bandwidth x-ray probe to achieve high spectral resolution. The challenges in delivering an energetic monochromated beam from an x-ray free electron laser (XFEL) thus limit its use in few-shot experiments, including for the study of high energy density systems. Here we demonstrate that by correlating the measurements of the self-amplified spontaneous emission (SASE) spectrum of an XFEL with the RIXS signal, using a dynamic kernel deconvolution with a neural surrogate, we can achieve electronic structure resolutions substantially higher than those normally afforded by the bandwidth of the incoming x-ray beam. We further show how this technique allows us to discriminate between the valence structures of Fe and Fe$_2$O$_3$, and provides access to temperature measurements as well as M-shell binding energies estimates in warm-dense Fe compounds.

cond-mat.mtrl-sci↗

arXiv2022

Variable Dynamic Mode Decomposition for Estimating Time Eigenvalues in Nuclear Systems

We present a new approach to calculating time eigenvalues of the neutron transport operator (also known as $\alpha$ eigenvalues) by extending the dynamic mode decomposition (DMD) to allow for non-uniform time steps. The new method, called variable dynamic mode decomposition (VDMD), is shown to be accurate when computing eigenvalues for systems that were infeasible with DMD due to a large separation in time scales (such as those that occur in delayed supercritical systems). The $\alpha$ eigenvalues of an infinite medium neutron transport problem with delayed neutrons and consequently having multiple, very different relevant time scales are computed. Furthermore, VDMD is shown to be of similar accuracy to the original DMD approach when computing eigenvalues in other systems where the previously studied DMD approach can be used.

physics.comp-ph↗

arXiv2022

I Learn to Diffuse, or Data Alchemy 101: a Mnemonic Manifesto

In this manifesto, we put forward the idea of data alchemy as a narrative device to discuss storytelling and transdisciplinarity in visualization. If data is the prima materia of modern science, how does one perform the Great Work? We use text-to-image diffusion-based generative art to develop the concept, and structure our argument in ten propositions, as if they were ten issues of a comic novel on data alchemy: Ad Disco Diffusionem. To follow the argument, the reader must immerse themselves in our miro board, and navigate a multimedia semiotic topology that includes comics, videos, code demos, and ergotic literature in a true alchemic sense. By accessing this paradigm one might find new sources of inspiration for scientific inquiry in familiar places, or get lost in the creative exploration of the unknown. Our colorful, sometimes poetic, exposition should not distract the reader from the seriousness of the ideas discussed, but ultimately it is about the journey.

cs.HC↗