arXiv subjects

Eli Shlizerman

Publications and source records attributed to Eli Shlizerman.

At least 19 recordsLinked to original sources

arXiv2026

HALO: A Physics-Aware LLM Agent Framework for Nanophotonic Design

Language models have recently been applied to nanophotonic design, but it remains unclear whether they can reliably translate optical objectives into simulation-ready designs, execute electromagnetic analysis, and revise decisions from numerical feedback. We introduce HALO, a physics-aware framework that couples language-model planners with typed design specifications, electromagnetic simulation, diagnostic evaluation, and optional reuse of prior failure trajectories in an iterative design loop. We further introduce HALO-Bench, a 52-task benchmark spanning lab-derived, paper-derived, and open-ended nanophotonic design tasks under a shared evaluation protocol. We compare three planner configurations: a Fixed Structured Workflow, an Autonomous Structured Agent using the same simulation interface, and an Autonomous Coding Agent that directly writes and executes simulation code. The Fixed Structured Workflow is the most token-efficient and exhibits no observed code- or path-level failures, while autonomous coding can achieve higher task success with stronger models at the cost of additional operational failures. We also study reuse of prior failed trajectories. On targeted multi-round tasks, retrieved failure feedback reduces both iterations to first success and total token use. These results clarify the tradeoffs between explicit interfaces, autonomous execution, and reusable design experience in scientific agents.

physics.optics↗

arXiv2026

SceneBind: Binding What and Where Across Vision, Audio and Language

We present SceneBind, an omni-modal representation of realistic scenes with joint semantic and 3D spatial understanding across vision, audio and language. Existing omni-modal encoders excel at instance-level semantics (i.e., what is present), but often lack explicit spatial structure (i.e., where it is). SceneBind addresses this gap by representing each scene as a semantic-spatial entity, combining a global semantic embedding with object-centric semantic-spatial slots. This representation explicitly captures object-level semantics, spatial attributes, and uncertainty. We further propose SceneBind Matching, a semantic-spatial matching scheme that integrates global scene similarity with object alignment, supporting cross-modal scene retrieval and object grounding. To train and evaluate SceneBind, we curate a novel real-world binaural audio-visual dataset with structured semantic and spatial annotations, and propose a training protocol for aligning semantic and spatial signals across modalities. SceneBind is compatible with large-scale pretrained semantic encoders, adds lightweight spatial modeling with only a few additional tokens. It achieves state-of-the-art scene and spatial retrieval while enabling strong zero-shot transfer to downstream tasks such as audio-visual localization.

cs.CV↗

arXiv2026

The Global Empirical NTK: Self-Referential Bias and Dimensionality of Gradient Descent Learning

In training a neural network with gradient descent (GD), each iteration induces a linear operator that governs first-order updates to a model's internal state variables. We define this operator as the Global Empirical Neural Tangent Kernel (NTK). In finite-width networks, the NTK is typically intractable to form, leading prior work to focus on restrictive settings such as tracking outputs only or taking infinite-width limits. Here, we study the structure of the NTK for a range of models. Formulating the model state as the solution to a single global implicit constraint, we derive the NTK as a product of two operators: K, accounting for immediate parameter-to-state interactions, and P, describing internal state-to-state dependencies. For a broad class of weight-based models, including RNNs and transformers, we prove a universal Kronecker-core theorem showing that K admits an exact, computable form given by the Gram matrix of weight-site variables. This core structure reveals that the NTK is structurally bottlenecked, constraining its effective rank and giving rise to a self-referential bias whereby GD preferentially learns within dominant modes of joint hidden and input activity. For recurrent models, we examine the spectrum of the NTK and show when it is biased and low-rank in space or time under the proposed decomposition. We further demonstrate that model dynamics at initialization bias the NTK, restricting learning and preventing task components from being learned effectively. Finally, we show that the NTK associated with a self-attention transformer is likewise structurally constrained to be low-rank. Overall, we show that the NTK possesses tractable structure that explains GD bias toward task solutions and the emergence of low-rank representations. To enable use of the NTK as a practical metric, we build kpflow, a library relying on randomized matrix-free numerical linear algebra.

cs.LG↗

arXiv2026

DiffuMask: Diffusion Language Model for Token-level Prompt Pruning

In-Context Learning and Chain-of-Thought prompting improve reasoning in large language models (LLMs). These typically come at the cost of longer, more expensive prompts that may contain redundant information. Prompt compression based on pruning offers a practical solution, yet existing methods rely on sequential token removal which is computationally intensive. We present DiffuMask, a diffusion-based framework integrating hierarchical shot-level and token-level pruning signals, that enables rapid and parallel prompt pruning via iterative mask prediction. DiffuMask substantially accelerates the compression process via masking multiple tokens in each denoising step. It offers tunable control over retained content, preserving essential reasoning context and achieving up to 80\% prompt length reduction. Meanwhile, it maintains or improves accuracy across in-domain, out-of-domain, and cross-model settings. Our results show that DiffuMask provides a generalizable and controllable framework for prompt compression, facilitating faster and more reliable in-context reasoning in LLMs.

cs.CL↗

arXiv2026

RPNT: Robust Pre-trained Neural Transformer -- A Pathway for Generalized Motor Decoding

Brain motor decoding aims to interpret and translate neural activity into behaviors. Decoding models should generalize across variations, such as recordings from different brain sites, experimental sessions, behavior types, and subjects, will be critical for real-world applications. Current decoding models only partially address these challenges. In this work, we develop a pretrained neural transformer model, RPNT - Robust Pretrained Neural Transformer, designed to achieve robust generalization through pretraining, which in turn enables effective finetuning for downstream motor decoding tasks. We achieved the proposed RPNT architecture by systematically investigating which transformer building blocks could be suitable for neural spike activity modeling, since components from models developed for other modalities, such as text and images, do not transfer directly to neural data. The final RPNT architecture incorporates three unique enabling components: 1) Multidimensional rotary positional embedding to aggregate experimental metadata such as site coordinates, session ids and behavior types; 2) Context-based attention mechanism via convolution kernels operating on global attention to learn local temporal structures for handling non-stationarity of neural population activity; 3) Robust self-supervised learning objective with stochastic causal masking strategies and contrastive representations. We pretrained two versions of RPNT on distinct datasets that present significant generalization challenges: a) Multi-session, multi-task, and multi-subject microelectrode benchmark; b) Multi-site recordings using high-density Neuropixel 1.0 probes from many cortical locations. After pretraining, we evaluated RPNT generalization on cross-session, cross-type, cross-subject, and cross-site downstream behavior decoding tasks. Our RPNT consistently outperforms the existing decoding models on these tasks.

cs.LG↗

arXiv2025

Flowing from Reasoning to Motion: Learning 3D Hand Trajectory Prediction from Egocentric Human Interaction Videos

Prior works on 3D hand trajectory prediction are constrained by datasets that decouple motion from semantic supervision and by models that weakly link reasoning and action. To address these, we first present the EgoMAN dataset, a large-scale egocentric dataset for interaction stage-aware 3D hand trajectory prediction with 219K 6DoF trajectories and 3M structured QA pairs for semantic, spatial, and motion reasoning. We then introduce the EgoMAN model, a reasoning-to-motion framework that links vision-language reasoning and motion generation via a trajectory-token interface. Trained progressively to align reasoning with motion dynamics, our approach yields accurate and stage-aware trajectories with generalization across real-world scenes.

cs.CV↗

arXiv2025

Advantages of Broadband Metalenses for Generalizable Image Classification

Optical neural networks (ONNs) are gaining increasing attention to accelerate machine learning tasks. In particular, static meta-optical encoders designed for task-specific pre-processing have demonstrated orders of magnitude smaller energy consumption over purely digital counterparts, albeit at the cost of a slight degradation in classification accuracy. However, a lack of generalizability poses serious challenges for wide deployment of static meta-optical front-ends. Here, we investigate the utility of a single-layer metalens as a meta-optical encoder in ONNs for generalizable image classification. Specifically, we show that a visible-spectrum broadband metalens can achieve image classification accuracy comparable to high-end, sensor-limited optics and consistently outperforms the corresponding hyperboloid baseline across a wide range of sensor pixel sizes and digital backends. We further design an end-to-end optimized single-aperture metasurface for ImageNet classification and observe that the optimization tends to balance the modulation transfer function (MTF) across wavelengths within the sensor-detectable passband. Together, these observations suggest that the preservation of spatial-frequency information is an important factor influencing the performance of ONNs. Our results provide physical insight into the process of task-driven optical optimization and offer practical guidance for the design of high-performance ONNs and meta-optical encoders for generalizable computer vision tasks.

physics.optics↗

arXiv2025

Neural Tangent Knowledge Distillation for Optical Convolutional Networks

Hybrid Optical Neural Networks (ONNs, typically consisting of an optical frontend and a digital backend) offer an energy-efficient alternative to fully digital deep networks for real-time, power-constrained systems. However, their adoption is limited by two main challenges: the accuracy gap compared to large-scale networks during training, and discrepancies between simulated and fabricated systems that further degrade accuracy. While previous work has proposed end-to-end optimizations for specific datasets (e.g., MNIST) and optical systems, these approaches typically lack generalization across tasks and hardware designs. To address these limitations, we propose a task-agnostic and hardware-agnostic pipeline that supports image classification and segmentation across diverse optical systems. To assist optical system design before training, we estimate achievable model accuracy based on user-specified constraints such as physical size and the dataset. For training, we introduce Neural Tangent Knowledge Distillation (NTKD), which aligns optical models with electronic teacher networks, thereby narrowing the accuracy gap. After fabrication, NTKD also guides fine-tuning of the digital backend to compensate for implementation errors. Experiments on multiple datasets (e.g., MNIST, CIFAR, Carvana Masking) and hardware configurations show that our pipeline consistently improves ONN performance and enables practical deployment in both pre-fabrication simulations and physical implementations.

cs.CV↗

arXiv2025

SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding

Intracortical Brain-Computer Interfaces (iBCI) aim to decode behavior from neural population activity, enabling individuals with motor impairments to regain motor functions and communication abilities. A key challenge in long-term iBCI is the nonstationarity of neural recordings, where the composition and tuning profiles of the recorded populations are unstable across recording sessions. Existing methods attempt to address this issue by explicit alignment techniques; however, they rely on fixed neural identities and require test-time labels or parameter updates, limiting their generalization across sessions and imposing additional computational burden during deployment. In this work, we introduce SPINT - a Spatial Permutation-Invariant Neural Transformer framework for behavioral decoding that operates directly on unordered sets of neural units. Central to our approach is a novel context-dependent positional embedding scheme that dynamically infers unit-specific identities, enabling flexible generalization across recording sessions. SPINT supports inference on variable-size populations and allows few-shot, gradient-free adaptation using a small amount of unlabeled data from the test session. To further promote model robustness to population variability, we introduce dynamic channel dropout, a regularization method for iBCI that simulates shifts in population composition during training. We evaluate SPINT on three multi-session datasets from the FALCON Benchmark, covering continuous motor decoding tasks in human and non-human primates. SPINT demonstrates robust cross-session generalization, outperforming existing zero-shot and few-shot unsupervised baselines while eliminating the need for test-time alignment and fine-tuning. Our work contributes an initial step toward a robust and scalable neural decoding framework for long-term iBCI applications.

q-bio.NC↗

arXiv2025

KPFlow: An Operator Perspective on Dynamic Collapse Under Gradient Descent Training of Recurrent Networks

Gradient Descent (GD) and its variants are the primary tool for enabling efficient training of recurrent dynamical systems such as Recurrent Neural Networks (RNNs), Neural ODEs and Gated Recurrent units (GRUs). The dynamics that are formed in these models exhibit features such as neural collapse and emergence of latent representations that may support the remarkable generalization properties of networks. In neuroscience, qualitative features of these representations are used to compare learning in biological and artificial systems. Despite recent progress, there remains a need for theoretical tools to rigorously understand the mechanisms shaping learned representations, especially in finite, non-linear models. Here, we show that the gradient flow, which describes how the model's dynamics evolve over GD, can be decomposed into a product that involves two operators: a Parameter Operator, K, and a Linearized Flow Propagator, P. K mirrors the Neural Tangent Kernel in feed-forward neural networks, while P appears in Lyapunov stability and optimal control theory. We demonstrate two applications of our decomposition. First, we show how their interplay gives rise to low-dimensional latent dynamics under GD, and, specifically, how the collapse is a result of the network structure, over and above the nature of the underlying task. Second, for multi-task training, we show that the operators can be used to measure how objectives relevant to individual sub-tasks align. We experimentally and theoretically validate these findings, providing an efficient Pytorch package, \emph{KPFlow}, implementing robust analysis tools for general recurrent architectures. Taken together, our work moves towards building a next stage of understanding of GD learning in non-linear recurrent models.

cs.LG↗

arXiv2025

Hyperpruning: Efficient Search through Pruned Variants of Recurrent Neural Networks Leveraging Lyapunov Spectrum

A variety of pruning methods have been introduced for over-parameterized Recurrent Neural Networks to improve efficiency in terms of power consumption and storage utilization. These advances motivate a new paradigm, termed `hyperpruning', which seeks to identify the most suitable pruning strategy for a given network architecture and application. Unlike conventional hyperparameter search, where the optimal configuration's accuracy remains uncertain, in the context of network pruning, the accuracy of the dense model sets the target for the accuracy of the pruned one. The goal, therefore, is to discover pruned variants that match or even surpass this established accuracy. However, exhaustive search over pruning configurations is computationally expensive and lacks early performance guarantees. To address this challenge, we propose a novel Lyapunov Spectrum (LS)-based distance metric that enables early comparison between pruned and dense networks, allowing accurate prediction of post-training performance. By integrating this LS-based distance with standard hyperparameter optimization algorithms, we introduce an efficient hyperpruning framework, termed LS-based Hyperpruning (LSH). LSH reduces search time by an order of magnitude compared to conventional approaches relying on full training. Experiments on stacked LSTM and RHN architectures using the Penn Treebank dataset, and on AWD-LSTM-MoS using WikiText-2, demonstrate that under fixed training budgets and target pruning ratios, LSH consistently identifies superior pruned models. Remarkably, these pruned variants not only outperform those selected by loss-based baseline but also exceed the performance of their dense counterpart.

cs.LG↗

arXiv2025

SAVVY: Spatial Awareness via Audio-Visual LLMs through Seeing and Hearing

3D spatial reasoning in dynamic, audio-visual environments is a cornerstone of human cognition yet remains largely unexplored by existing Audio-Visual Large Language Models (AV-LLMs) and benchmarks, which predominantly focus on static or 2D scenes. We introduce SAVVY-Bench, the first benchmark for 3D spatial reasoning in dynamic scenes with synchronized spatial audio. SAVVY-Bench is comprised of thousands of relationships involving static and moving objects, and requires fine-grained temporal grounding, consistent 3D localization, and multi-modal annotation. To tackle this challenge, we propose SAVVY, a novel training-free reasoning pipeline that consists of two stages: (i) Egocentric Spatial Tracks Estimation, which leverages AV-LLMs as well as other audio-visual methods to track the trajectories of key objects related to the query using both visual and spatial audio cues, and (ii) Dynamic Global Map Construction, which aggregates multi-modal queried object trajectories and converts them into a unified global dynamic map. Using the constructed map, a final QA answer is obtained through a coordinate transformation that aligns the global map with the queried viewpoint. Empirical evaluation demonstrates that SAVVY substantially enhances performance of state-of-the-art AV-LLMs, setting a new standard and stage for approaching dynamic 3D spatial reasoning in AV-LLMs.

cs.CV↗

arXiv2025

2ndMatch: Finetuning Pruned Diffusion Models via Second-Order Jacobian Matching

Diffusion models achieve remarkable performance across diverse generative tasks in computer vision, but their high computational cost remains a major barrier to deployment. Model pruning offers a promising way to reduce inference cost and enable lightweight models. However, pruning leads to quality drop due to reduced capacity. A key limitation of existing pruning approaches is that pruned models are finetuned using the same objective as the dense model (denoising score matching). Since the dense model is accessible during finetuning, it warrants a more effective approach for knowledge transfer from the dense to the pruned model. Motivated by this, we propose \textbf{2ndMatch} (\textbf{2ndM}), a general-purpose finetuning framework that introduces a \textbf{2nd}-order Jacobian ($J^{\top} J$) \textbf{M}atching loss inspired by Finite-Time Lyapunov Exponents. \textbf{2ndM} teaches the pruned model to mimic the sensitivity of the dense teacher, i.e., how to respond to small perturbations over time, through scalable random projections. The framework is architecture-agnostic and applies to both U-Net- and Transformer-based diffusion models. Experiments on CIFAR-10, CelebA, LSUN, ImageNet, and MSCOCO demonstrate that \textbf{2ndM} reduces the performance gap between pruned and dense models, substantially improving output quality.

cs.GR↗

arXiv2025

Modular integration of neural connectomics, dynamics and biomechanics for identification of behavioral sensorimotor pathways in Caenorhabditis elegans

Computational approaches which emulate in-vivo nervous system are needed to investigate mechanisms of the brain to orchestrate behavior. Such approaches must integrate a series of biophysical models encompassing the nervous system, muscles, biomechanics to allow observing the system in its entirety while supporting model variations. Here we develop modWorm: a modular modeling framework for the nematode C. elegans. modWorm allows for construction of a model as an integrated series of configurable, exchangeable modules each describing specific biophysical processes across different modalities. Utilizing modWorm, we propose a base neuro-mechanical model for C. elegans built upon the complete connectome. The model integrates a series of 7 modules: i) intra-cellular dynamics, ii) electrical and iii) chemical extra-cellular neural dynamics, iv) translation of neural activity to muscle calcium dynamics, v) muscle calcium dynamics to muscle forces, vi) muscle forces to body postures and vii) proprioceptive feedback. We validate the base model by in-silico injection of constant currents into neurons known to be associated with locomotion behaviors and by applying external forces to the body. Applications of in-silico neural stimuli experimentally known to modulate locomotion show that the model can recapitulate natural behavioral responses such as forward and backward locomotion as well as mid-locomotion responses such as avoidance and turns. Furthermore, through in-silico ablation surveys, the model can infer novel neural circuits involved in sensorimotor behaviors. To further dissect mechanisms of locomotion, we utilize modWorm to introduce empirical based model variations and model optimizations to elucidate their effects on simulated locomotion. Our results show that modWorm can be utilized to identify neural circuits which control, mediate and generate natural behavior.

q-bio.QM↗

arXiv2025

Hearing Anywhere in Any Environment

In mixed reality applications, a realistic acoustic experience in spatial environments is as crucial as the visual experience for achieving true immersion. Despite recent advances in neural approaches for Room Impulse Response (RIR) estimation, most existing methods are limited to the single environment on which they are trained, lacking the ability to generalize to new rooms with different geometries and surface materials. We aim to develop a unified model capable of reconstructing the spatial acoustic experience of any environment with minimum additional measurements. To this end, we present xRIR, a framework for cross-room RIR prediction. The core of our generalizable approach lies in combining a geometric feature extractor, which captures spatial context from panorama depth images, with a RIR encoder that extracts detailed acoustic features from only a few reference RIR samples. To evaluate our method, we introduce ACOUSTICROOMS, a new dataset featuring high-fidelity simulation of over 300,000 RIRs from 260 rooms. Experiments show that our method strongly outperforms a series of baselines. Furthermore, we successfully perform sim-to-real transfer by evaluating our model on four real-world environments, demonstrating the generalizability of our approach and the realism of our dataset.

cs.CV↗

arXiv2025

SoundVista: Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding

We introduce SoundVista, a method to generate the ambient sound of an arbitrary scene at novel viewpoints. Given a pre-acquired recording of the scene from sparsely distributed microphones, SoundVista can synthesize the sound of that scene from an unseen target viewpoint. The method learns the underlying acoustic transfer function that relates the signals acquired at the distributed microphones to the signal at the target viewpoint, using a limited number of known recordings. Unlike existing works, our method does not require constraints or prior knowledge of sound source details. Moreover, our method efficiently adapts to diverse room layouts, reference microphone configurations and unseen environments. To enable this, we introduce a visual-acoustic binding module that learns visual embeddings linked with local acoustic properties from panoramic RGB and depth data. We first leverage these embeddings to optimize the placement of reference microphones in any given scene. During synthesis, we leverage multiple embeddings extracted from reference locations to get adaptive weights for their contribution, conditioned on target viewpoint. We benchmark the task on both publicly available data and real-world settings. We demonstrate significant improvements over existing methods.

cs.SD↗

arXiv2024

Brain-to-Text Benchmark '24: Lessons Learned

Speech brain-computer interfaces aim to decipher what a person is trying to say from neural activity alone, restoring communication to people with paralysis who have lost the ability to speak intelligibly. The Brain-to-Text Benchmark '24 and associated competition was created to foster the advancement of decoding algorithms that convert neural activity to text. Here, we summarize the lessons learned from the competition ending on June 1, 2024 (the top 4 entrants also presented their experiences in a recorded webinar). The largest improvements in accuracy were achieved using an ensembling approach, where the output of multiple independent decoders was merged using a fine-tuned large language model (an approach used by all 3 top entrants). Performance gains were also found by improving how the baseline recurrent neural network (RNN) model was trained, including by optimizing learning rate scheduling and by using a diphone training objective. Improving upon the model architecture itself proved more difficult, however, with attempts to use deep state space models or transformers not yet appearing to offer a benefit over the RNN baseline. The benchmark will remain open indefinitely to support further work towards increasing the accuracy of brain-to-text algorithms.

cs.CL↗

arXiv2024

Brain-to-Text Decoding with Context-Aware Neural Representations and Large Language Models

Decoding attempted speech from neural activity offers a promising avenue for restoring communication abilities in individuals with speech impairments. Previous studies have focused on mapping neural activity to text using phonemes as the intermediate target. While successful, decoding neural activity directly to phonemes ignores the context dependent nature of the neural activity-to-phoneme mapping in the brain, leading to suboptimal decoding performance. In this work, we propose the use of diphone - an acoustic representation that captures the transitions between two phonemes - as the context-aware modeling target. We integrate diphones into existing phoneme decoding frameworks through a novel divide-and-conquer strategy in which we model the phoneme distribution by marginalizing over the diphone distribution. Our approach effectively leverages the enhanced context-aware representation of diphones while preserving the manageable class size of phonemes, a key factor in simplifying the subsequent phoneme-to-text conversion task. We demonstrate the effectiveness of our approach on the Brain-to-Text 2024 benchmark, where it achieves state-of-the-art Phoneme Error Rate (PER) of 15.34% compared to 16.62% PER of monophone-based decoding. When coupled with finetuned Large Language Models (LLMs), our method yields a Word Error Rate (WER) of 5.77%, significantly outperforming the 8.93% WER of the leading method in the benchmark.

eess.SP↗