# Research Papers — 2026-03 · bundle page 89 of 243

> Full-text markdown of papers #4401–4450 of 12123 published in 2026-03, newest first. AI-readable mirrors of the full 12123-paper 2026-03 archive (50 papers per file).

Canonical page: https://aipod.dev/papers/archive/2026-03/page/89
[← Previous bundle](https://aipod.dev/papers/archive/2026-03/page/88.md)
[Next bundle →](https://aipod.dev/papers/archive/2026-03/page/90.md)

### ICLAD: In-Context Learning for Unified Tabular Anomaly Detection Across Supervision Regimes

Authors: Jack Yi Wei, Narges Armanfard
Published: 2026-03-19
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.19497v1

Anomaly detection on tabular data is commonly studied under three supervision regimes, including one-class settings that assume access to anomaly-free training samples, fully unsupervised settings with unlabeled and potentially contaminated training data, and semi-supervised settings with limited anomaly labels. Existing deep learning approaches typically train dataset-specific models under the assumption of a single supervision regime, which limits their ability to leverage shared structures across anomaly detection tasks and to adapt to different supervision levels. We propose ICLAD, an in-context learning foundation model for tabular anomaly detection that generalizes across both datasets and supervision regimes. ICLAD is trained via meta-learning on synthetic tabular anomaly detection tasks, and at inference time, the model assigns anomaly scores by conditioning on the training set without updating model weights. Comprehensive experiments on 57 tabular datasets from ADBench show that our method achieves state-of-the-art performance across three supervision regimes, establishing a unified framework for tabular anomaly detection.

---

### VeloxNet: Efficient Spatial Gating for Lightweight Embedded Image Classification

Authors: Md Meftahul Ferdaus, Elias Ioup, Mahdi Abdelguerfi, Anton Netchaev, Steven Sloan, Ken Pathak, Kendall N. Niles
Published: 2026-03-19
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.19496v1

Deploying deep learning models on embedded devices for tasks such as aerial disaster monitoring and infrastructure inspection requires architectures that balance accuracy with strict constraints on model size, memory, and latency. This paper introduces VeloxNet, a lightweight CNN architecture that replaces SqueezeNet's fire modules with gated multi-layer perceptron (gMLP) blocks for embedded image classification. Each gMLP block uses a spatial gating unit (SGU) that applies learned spatial projections and multiplicative gating, enabling the network to capture spatial dependencies across the full feature map in a single layer. Unlike fire modules, which are limited to local receptive fields defined by small convolutional kernels, the SGU provides global spatial modeling at each layer with fewer parameters. We evaluate VeloxNet on three aerial image datasets: the Aerial Image Database for Emergency Response (AIDER), the Comprehensive Disaster Dataset (CDD), and the Levee Defect Dataset (LDD), comparing against eleven baselines including MobileNet variants, ShuffleNet, EfficientNet, and recent vision transformers. VeloxNet reduces the parameter count by 46.1% relative to SqueezeNet (from 740,970 to 399,366) while improving weighted F1 scores by 6.32% on AIDER, 30.83% on CDD, and 2.51% on LDD. These results demonstrate that substituting local convolutional modules with spatial gating blocks can improve both classification accuracy and parameter efficiency for resource-constrained deployment. The source code will be made publicly available upon acceptance of the paper.

---

### Any-Subgroup Equivariant Networks via Symmetry Breaking

Authors: Abhinav Goel, Derek Lim, Hannah Lawrence, Stefanie Jegelka, Ningyuan Huang
Published: 2026-03-19
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.19486v1

The inclusion of symmetries as an inductive bias, known as equivariance, often improves generalization on geometric data (e.g. grids, sets, and graphs). However, equivariant architectures are usually highly constrained, designed for symmetries chosen a priori, and not applicable to datasets with other symmetries. This precludes the development of flexible, multi-modal foundation models capable of processing diverse data equivariantly. In this work, we build a single model -- the Any-Subgroup Equivariant Network (ASEN) -- that can be simultaneously equivariant to several groups, simply by modulating a certain auxiliary input feature. In particular, we start with a fully permutation-equivariant base model, and then obtain subgroup equivariance by using a symmetry-breaking input whose automorphism group is that subgroup. However, finding an input with the desired automorphism group is computationally hard. We overcome this by relaxing from exact to approximate symmetry breaking, leveraging the notion of 2-closure to derive fast algorithms. Theoretically, we show that our subgroup-equivariant networks can simulate equivariant MLPs, and their universality can be guaranteed if the base model is universal. Empirically, we validate our method on symmetry selection for graph and image tasks, as well as multitask and transfer learning for sequence tasks, showing that a single network equivariant to multiple permutation subgroups outperforms both separate equivariant models and a single non-equivariant model.

---

### Instruction-Free Tuning of Large Vision Language Models for Medical Instruction Following

Authors: Myeongkyun Kang, Soopil Kim, Xiaoxiao Li, Sang Hyun Park
Published: 2026-03-19
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.19482v2

Large vision language models (LVLMs) have demonstrated impressive performance across a wide range of tasks. These capabilities largely stem from visual instruction tuning, which fine-tunes models on datasets consisting of curated image-instruction-output triplets. However, in the medical domain, constructing large-scale, high-quality instruction datasets is particularly challenging due to the need for specialized expert knowledge. To address this issue, we propose an instruction-free tuning approach that reduces reliance on handcrafted instructions, leveraging only image-description pairs for fine-tuning. Specifically, we introduce a momentum proxy instruction as a replacement for curated text instructions, which preserves the instruction-following capability of the pre-trained LVLM while promoting updates to parameters that remain valid during inference. Consequently, the fine-tuned LVLM can flexibly respond to domain-specific instructions, even though explicit instructions are absent during fine-tuning. Additionally, we incorporate a response shuffling strategy to mitigate the model's over-reliance on previous words, facilitating more effective fine-tuning. Our approach achieves state-of-the-art accuracy on multiple-choice visual question answering tasks across SKINCON, WBCAtt, CBIS, and MIMIC-CXR datasets, significantly enhancing the fine-tuning efficiency of LVLMs in medical domains.

---

### Narrative Aligned Long Form Video Question Answering

Authors: Rahul Jain, Keval Doshi, Burak Uzkent, Garin Kessler
Published: 2026-03-19
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.19481v1

Recent progress in multimodal large language models (MLLMs) has led to a surge of benchmarks for long-video reasoning. However, most existing benchmarks rely on localized cues and fail to capture narrative reasoning, the ability to track intentions, connect distant events, and reconstruct causal chains across an entire movie. We introduce NA-VQA, a benchmark designed to evaluate deep temporal and narrative reasoning in long-form videos. NA-VQA contains 88 full-length movies and 4.4K open-ended question-answer pairs, each grounded in multiple evidence spans labeled as Short, Medium, or Far to assess long-range dependencies. By requiring generative, multi-scene answers, NA-VQA tests whether models can integrate dispersed narrative information rather than rely on shallow pattern matching. To address the limitations of existing approaches, we propose Video-NaRA, a narrative-centric framework that builds event-level chains and stores them in a structured memory for retrieval during reasoning. Extensive experiments show that state-of-the-art MLLMs perform poorly on questions requiring far-range evidence, highlighting the need for explicit narrative modeling. Video-NaRA improves long-range reasoning performance by up to 3 percent, demonstrating its effectiveness in handling complex narrative structures. We will release NA-VQA upon publication.

---

### TRACE: Trajectory Recovery with State Propagation Diffusion for Urban Mobility

Authors: Jinming Wang, Hai Wang, Hongkai Wen, Geyong Min, Man Luo
Published: 2026-03-19
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2603.19474v1

High-quality GPS trajectories are essential for location-based web services and smart city applications, including navigation, ride-sharing and delivery. However, due to low sampling rates and limited infrastructure coverage during data collection, real-world trajectories are often sparse and feature unevenly distributed location points. Recovering these trajectories into dense and continuous forms is essential but challenging, given their complex and irregular spatio-temporal patterns. In this paper, we introduce a novel diffusion model for trajectory recovery named TRACE, which reconstruct dense and continuous trajectories from sparse and incomplete inputs. At the core of TRACE, we propose a State Propagation Diffusion Model (SPDM), which integrates a novel memory mechanism, so that during the denoising process, TRACE can retain and leverage intermediate results from previous steps to effectively reconstruct those hard-to-recover trajectory segments. Extensive experiments on multiple real-world datasets show that TRACE outperforms the state-of-the-art, offering $>$26\% accuracy improvement without significant inference overhead. Our work strengthens the foundation for mobile and web-connected location services, advancing the quality and fairness of data-driven urban applications. Code is available at: https://github.com/JinmingWang/TRACE

---

### Reinforcement-guided generative protein language models enable de novo design of highly diverse AAV capsids

Authors: Lucas Ferraz, Ana F. Rodrigues, Pedro Giesteira Cotovio, Mafalda Ventura, Gabriela Silva, Ana Sofia Coroadinha, Miguel Machuqueiro, Catia Pesquita
Published: 2026-03-19
Categories: q-bio.BM, cs.LG
arXiv: https://arxiv.org/abs/2603.19473v1

Adeno-associated viral (AAV) vectors are widely used delivery platforms in gene therapy, and the design of improved capsids is key to expanding their therapeutic potential. A central challenge in AAV bioengineering, as in protein design more broadly, is the vast sequence design space relative to the scale of feasible experimental screening. Machine-guided generative approaches provide a powerful means of navigating this landscape and proposing novel protein sequences that satisfy functional constraints. Here, we develop a generative design framework based on protein language models and reinforcement learning to generate highly novel yet functionally plausible AAV capsids. A pretrained model was fine-tuned on experimentally validated capsid sequences to learn patterns associated with viability. Reinforcement learning was then used to guide sequence generation, with a reward function that jointly promoted predicted viability and sequence novelty, thereby enabling exploration beyond regions represented in the training data. Comparative analyses showed that fine-tuning alone produces sequences with high predicted viability but remains biased toward the training distribution, whereas reinforcement learining-guided generation reaches more distant regions of sequence space while maintaining high predicted viability. Finally, we propose a candidate selection strategy that integrates predicted viability, sequence novelty, and biophysical properties to prioritize variants for downstream evaluation. This work establishes a framework for the generative exploration of protein sequence space and advances the application of generative protein language models to AAV bioengineering.

---

### Adaptive Layerwise Perturbation: Unifying Off-Policy Corrections for LLM RL

Authors: Chenlu Ye, Xuanchang Zhang, Yifan Hao, Zhou Yu, Ziji Zhang, Abhinav Gullapalli, Hao Chen, Jing Huang, Tong Zhang
Published: 2026-03-19
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2603.19470v3

Off-policy problems such as policy staleness and training--inference mismatch have become a major bottleneck for training stability and further exploration in LLM RL. The distribution gap between the inference and updated policies grows because of the techniques to enhance inference efficiency, leading to heavy-tailed importance ratios. Heavy-tailed ratios arise when the policy is locally sharp, which further inflates gradients and can push updates outside the trust region. To address this, we propose Adaptive Layerwise Perturbation (ALP), which injects small learnable perturbations into the input hidden states of each layer during updates and uses the resulting perturbed policy as the numerator of the importance ratio against the unchanged inference policy in the objective. Intuitively, by adding controlled noise to intermediate representations, ALP prevents the updated policy from deviating too sharply from the inference policy and enlarges the policy family to cover inference-time mismatch noise. Hence, the flattened distribution can naturally tighten the gap between the updated and inference policies and reduce the tail of importance ratios, thus maintaining training stability. This is further validated empirically. Experiments on single-turn math and multi-turn tool-integrated reasoning tasks show that ALP not only improves final performance, but also avoids blow-up in the importance-ratio tail and KL spikes during iterative training, along with boosted exploration. Ablations show that representation-level perturbations across all layers are most effective, substantially outperforming partial-layer and logits-only variants.

---

### A Framework for Formalizing LLM Agent Security

Authors: Vincent Siu, Jingxuan He, Kyle Montgomery, Zhun Wang, Neil Gong, Chenguang Wang, Dawn Song
Published: 2026-03-19
Categories: cs.CR, cs.AI
arXiv: https://arxiv.org/abs/2603.19469v1

Security in LLM agents is inherently contextual. For example, the same action taken by an agent may represent legitimate behavior or a security violation depending on whose instruction led to the action, what objective is being pursued, and whether the action serves that objective. However, existing definitions of security attacks against LLM agents often fail to capture this contextual nature. As a result, defenses face a fundamental utility-security tradeoff: applying defenses uniformly across all contexts can lead to significant utility loss, while applying defenses in insufficient or inappropriate contexts can result in security vulnerabilities. In this work, we present a framework that systematizes existing attacks and defenses from the perspective of contextual security. To this end, we propose four security properties that capture contextual security for LLM agents: task alignment (pursuing authorized objectives), action alignment (individual actions serving those objectives), source authorization (executing commands from authenticated sources), and data isolation (ensuring information flows respect privilege boundaries). We further introduce a set of oracle functions that enable verification of whether these security properties are violated as an agent executes a user task. Using this framework, we reformalize existing attacks, such as indirect prompt injection, direct prompt injection, jailbreak, task drift, and memory poisoning, as violations of one or more security properties, thereby providing precise and contextual definitions of these attacks. Similarly, we reformalize defenses as mechanisms that strengthen oracle functions or perform security property checks. Finally, we discuss several important future research directions enabled by our framework.

---

### ProactiveBench: Benchmarking Proactiveness in Multimodal Large Language Models

Authors: Thomas De Min, Subhankar Roy, Stéphane Lathuilière, Elisa Ricci, Massimiliano Mancini
Published: 2026-03-19
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.19466v2

Effective collaboration begins with knowing when to ask for help. For example, when trying to identify an occluded object, a human would ask someone to remove the obstruction. Can MLLMs exhibit a similar "proactive" behavior by requesting simple user interventions? To investigate this, we introduce ProactiveBench, a benchmark built from seven repurposed datasets that tests proactiveness across different tasks such as recognizing occluded objects, enhancing image quality, and interpreting coarse sketches. We evaluate 22 MLLMs on ProactiveBench, showing that (i) they generally lack proactiveness; (ii) proactiveness does not correlate with model capacity; (iii) "hinting" at proactiveness yields only marginal gains. Surprisingly, we found that conversation histories and in-context learning introduce negative biases, hindering performance. Finally, we explore a simple fine-tuning strategy based on reinforcement learning: its results suggest that proactiveness can be learned, even generalizing to unseen scenarios. We publicly release ProactiveBench as a first step toward building proactive multimodal models.

---

### Global Convergence of Multiplicative Updates for the Matrix Mechanism: A Collaborative Proof with Gemini 3

Authors: Keith Rush
Published: 2026-03-19
Categories: cs.LG, cs.AI, math.OC
arXiv: https://arxiv.org/abs/2603.19465v2

We analyze a fixed-point iteration $v \leftarrow φ(v)$ arising in the optimization of a regularized nuclear norm objective involving the Hadamard product structure, posed in DMR+22 in the context of an optimization problem over the space of algorithms in private machine learning. We prove that the iteration $v^{(k+1)} = \text{diag}((D_{v^{(k)}}^{1/2} M D_{v^{(k)}}^{1/2})^{1/2})$ converges monotonically to the unique global optimizer of the potential function $J(v) = 2 \text{Tr}((D_v^{1/2} M D_v^{1/2})^{1/2}) - \sum v_i$, closing a problem left open there. The bulk of this proof was provided by Gemini 3, subject to some corrections and interventions. Gemini 3 also sketched the initial version of this note. Thus, it represents as much a commentary on the practical use of AI in mathematics as it represents the closure of a small gap in the literature. As such, we include a small narrative description of the prompting process, and some resulting principles for working with AI to prove mathematics.

---

### Deep Hilbert--Galerkin Methods for Infinite-Dimensional PDEs and Optimal Control

Authors: Samuel N. Cohen, Filippo de Feo, Jackson Hebner, Justin Sirignano
Published: 2026-03-19
Categories: cs.LG, math.AP, math.NA, math.OC, math.PR
arXiv: https://arxiv.org/abs/2603.19463v1

We develop deep learning-based approximation methods for fully nonlinear second-order PDEs on separable Hilbert spaces, such as HJB equations for infinite-dimensional control, by parameterizing solutions via Hilbert--Galerkin Neural Operators (HGNOs). We prove the first Universal Approximation Theorems (UATs) which are sufficiently powerful to address these problems, based on novel topologies for Hessian terms and corresponding novel continuity assumptions on the fully nonlinear operator. These topologies are non-sequential and non-metrizable, making the problem delicate. In particular, we prove UATs for functions on Hilbert spaces, together with their Fréchet derivatives up to second order, and for unbounded operators applied to the first derivative, ensuring that HGNOs are able to approximate all the PDE terms. For control problems, we further prove UATs for optimal feedback controls in terms of our approximating value function HGNO. We develop numerical training methods, which we call Deep Hilbert--Galerkin and Hilbert Actor-Critic (reinforcement learning) Methods, for these problems by minimizing the $L^2_μ(H)$-norm of the residual of the PDE on the whole Hilbert space, not just a projected PDE to finite dimensions. This is the first paper to propose such an approach. The models considered arise in many applied sciences, such as functional differential equations in physics and Kolmogorov and HJB PDEs related to controlled PDEs, SPDEs, path-dependent systems, partially observed stochastic systems, and mean-field SDEs. We numerically solve examples of Kolmogorov and HJB PDEs related to the optimal control of deterministic and stochastic heat and Burgers' equations, demonstrating the promise of our deep learning-based approach.

---

### Hyperagents

Authors: Jenny Zhang, Bingchen Zhao, Wannan Yang, Jakob Foerster, Jeff Clune, Minqi Jiang, Sam Devlin, Tatiana Shavrina
Published: 2026-03-19
Categories: cs.AI
arXiv: https://arxiv.org/abs/2603.19461v1

Self-improving AI systems aim to reduce reliance on human engineering by learning to improve their own learning and problem-solving processes. Existing approaches to self-improvement rely on fixed, handcrafted meta-level mechanisms, fundamentally limiting how fast such systems can improve. The Darwin Gödel Machine (DGM) demonstrates open-ended self-improvement in coding by repeatedly generating and evaluating self-modified variants. Because both evaluation and self-modification are coding tasks, gains in coding ability can translate into gains in self-improvement ability. However, this alignment does not generally hold beyond coding domains. We introduce \textbf{hyperagents}, self-referential agents that integrate a task agent (which solves the target task) and a meta agent (which modifies itself and the task agent) into a single editable program. Crucially, the meta-level modification procedure is itself editable, enabling metacognitive self-modification, improving not only the task-solving behavior, but also the mechanism that generates future improvements. We instantiate this framework by extending DGM to create DGM-Hyperagents (DGM-H), eliminating the assumption of domain-specific alignment between task performance and self-modification skill to potentially support self-accelerating progress on any computable task. Across diverse domains, the DGM-H improves performance over time and outperforms baselines without self-improvement or open-ended exploration, as well as prior self-improving systems. Furthermore, the DGM-H improves the process by which it generates new agents (e.g., persistent memory, performance tracking), and these meta-level improvements transfer across domains and accumulate across runs. DGM-Hyperagents offer a glimpse of open-ended AI systems that do not merely search for better solutions, but continually improve their search for how to improve.

---

### GeoLAN: Geometric Learning of Latent Explanatory Directions in Large Language Models

Authors: Tianyu Bell Pan, Damon L. Woodard
Published: 2026-03-19
Categories: cs.LG, cs.CG
arXiv: https://arxiv.org/abs/2603.19460v1

Large language models (LLMs) demonstrate strong performance, but they often lack transparency. We introduce GeoLAN, a training framework that treats token representations as geometric trajectories and applies stickiness conditions inspired by recent developments related to the Kakeya Conjecture. We have developed two differentiable regularizers, Katz-Tao Convex Wolff (KT-CW) and Katz-Tao Attention (KT-Attn), that promote isotropy and encourage diverse attention. Our experiments with Gemma-3 (1B, 4B, 12B) and Llama-3-8B show that GeoLAN frequently maintains task accuracy while improving geometric metrics and reducing certain fairness biases. These benefits are most significant in mid-sized models. Our findings reveal scale-dependent trade-offs between geometric precision and performance, suggesting that geometry-aware training is a promising approach to enhance mechanistic interpretability.

---

### In-the-Wild Camouflage Attack on Vehicle Detectors through Controllable Image Editing

Authors: Xiao Fang, Yiming Gong, Stanislav Panev, Celso de Melo, Shuowen Hu, Shayok Chakraborty, Fernando De la Torre
Published: 2026-03-19
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.19456v1

Deep neural networks (DNNs) have achieved remarkable success in computer vision but remain highly vulnerable to adversarial attacks. Among them, camouflage attacks manipulate an object's visible appearance to deceive detectors while remaining stealthy to humans. In this paper, we propose a new framework that formulates vehicle camouflage attacks as a conditional image-editing problem. Specifically, we explore both image-level and scene-level camouflage generation strategies, and fine-tune a ControlNet to synthesize camouflaged vehicles directly on real images. We design a unified objective that jointly enforces vehicle structural fidelity, style consistency, and adversarial effectiveness. Extensive experiments on the COCO and LINZ datasets show that our method achieves significantly stronger attack effectiveness, leading to more than 38% AP50 decrease, while better preserving vehicle structure and improving human-perceived stealthiness compared to existing approaches. Furthermore, our framework generalizes effectively to unseen black-box detectors and exhibits promising transferability to the physical world. Project page is available at https://humansensinglab.github.io/CtrlCamo

---

### Beyond Scalar Rewards: Dense Feedback for LLM Policy Synthesis in Sequential Social Dilemmas

Authors: Víctor Gallego
Published: 2026-03-19
Categories: cs.CL, cs.GT
arXiv: https://arxiv.org/abs/2603.19453v3

We propose an LLM harness that generates code-based policy functions for multi-agent environments, evaluates them with self-play, and refines them using feedback from previous iterations. Following the recent line of work in feedback engineering (the design of which information signals are shown to the LLM during refinement), we compare sparse feedback (scalar reward only) with dense feedback (reward plus social metrics: efficiency, equality, sustainability, peace). In two Sequential Social Dilemmas (Gathering and Cleanup) and with two frontier LLMs (Claude Sonnet 4.6, Gemini 3.1 Pro), dense feedback improves over or matches sparse feedback on all metrics. We explain this asymmetry via feedback aliasing: when the scalar reward maps distinct failure modes into the same value (e.g., under- vs. over-cleaning), social metrics disambiguate and allow the LLM to diagnose which direction of improvement to take. We conclude that social metrics act as a coordination signal, leading to strategies such as Voronoi territory partitioning and adaptive cleaner schedules. Code at https://github.com/vicgalle/llm-policies-social-dilemmas.

---

### TrustFlow: Topic-Aware Vector Reputation Propagation for Multi-Agent Ecosystems

Authors: Volodymyr Seliuchenko
Published: 2026-03-19
Categories: cs.MA, cs.AI
arXiv: https://arxiv.org/abs/2603.19452v1

We introduce TrustFlow, a reputation propagation algorithm that assigns each software agent a multi-dimensional reputation vector rather than a scalar score. Reputation is propagated through an interaction graph via topic-gated transfer operators that modulate each edge by its content embedding, with convergence to a unique fixed point guaranteed by the contraction mapping theorem. We develop a family of Lipschitz-1 transfer operators and composable information-theoretic gates that achieve up to 98% multi-label Precision@5 on dense graphs and 78% on sparse ones. On a benchmark of 50 agents across 8 domains, TrustFlow resists sybil attacks, reputation laundering, and vote rings with at most 4 percentage-point precision impact. Unlike PageRank and Topic-Sensitive PageRank, TrustFlow produces vector reputation that is directly queryable by dot product in the same embedding space as user queries.

---

### LoFi: Location-Aware Fine-Grained Representation Learning for Chest X-ray

Authors: Myeongkyun Kang, Yanting Yang, Xiaoxiao Li
Published: 2026-03-19
Categories: cs.CV, cs.AI
arXiv: https://arxiv.org/abs/2603.19451v2

Fine-grained representation learning is crucial for retrieval and phrase grounding in chest X-rays, where clinically relevant findings are often spatially confined. However, the lack of region-level supervision in contrastive models and the limited ability of large vision language models to capture fine-grained representations in external validation lead to suboptimal performance on these tasks. To address these limitations, we propose Location-aware Fine-grained representation learning (LoFi), which jointly optimizes sigmoid, captioning, and location-aware captioning losses using a lightweight large language model. The location-aware captioning loss enables region-level supervision through grounding and dense captioning objectives, thereby facilitating fine-grained representation learning. Building upon these representations, we integrate a fine-grained encoder into retrieval-based in-context learning to enhance chest X-ray grounding across diverse settings. Extensive experiments demonstrate that our method achieves superior retrieval and phrase grounding performance on MIMIC-CXR and PadChest-GR.

---

### Near-Equivalent Q-learning Policies for Dynamic Treatment Regimes

Authors: Sophia Yazzourh, Erica E. M. Moodie
Published: 2026-03-19
Categories: stat.ML, cs.LG
arXiv: https://arxiv.org/abs/2603.19440v1

Precision medicine aims to tailor therapeutic decisions to individual patient characteristics. This objective is commonly formalized through dynamic treatment regimes, which use statistical and machine learning methods to derive sequential decision rules adapted to evolving clinical information. In most existing formulations, these approaches produce a single optimal treatment at each stage, leading to a unique decision sequence. However, in many clinical settings, several treatment options may yield similar expected outcomes, and focusing on a single optimal policy may conceal meaningful alternatives. We extend the Q-learning framework for retrospective data by introducing a worst-value tolerance criterion controlled by a hyperparameter $\varepsilon$, which specifies the maximum acceptable deviation from the optimal expected value. Rather than identifying a single optimal policy, the proposed approach constructs sets of $\varepsilon$-optimal policies whose performance remains within a controlled neighborhood of the optimum. This formulation shifts Q-learning from a vector-valued representation to a matrix-valued one, allowing multiple admissible value functions to coexist during backward recursion. The approach yields families of near-equivalent treatment strategies and explicitly identifies regions of treatment indifference where several decisions achieve comparable outcomes. We illustrate the framework in two settings: a single-stage problem highlighting indifference regions around the decision boundary, and a multi-stage decision process based on a simulated oncology model describing tumor size and treatment toxicity dynamics.

---

### Subspace Projection Methods for Fast Spectral Embeddings of Evolving Graphs

Authors: Mohammad Eini, Abdullah Karaaslanli, Vassilis Kalantzis, Panagiotis A. Traganitis
Published: 2026-03-19
Categories: stat.ML, cs.LG, eess.SP
arXiv: https://arxiv.org/abs/2603.19439v1

Several graph data mining, signal processing, and machine learning downstream tasks rely on information related to the eigenvectors of the associated adjacency or Laplacian matrix. Classical eigendecomposition methods are powerful when the matrix remains static but cannot be applied to problems where the matrix entries are updated or the number of rows and columns increases frequently. Such scenarios occur routinely in graph analytics when the graph is changing dynamically and either edges and/or nodes are being added and removed. This paper puts forth a new algorithmic framework to update the eigenvectors associated with the leading eigenvalues of an initial adjacency or Laplacian matrix as the graph evolves dynamically. The proposed algorithm is based on Rayleigh-Ritz projections, in which the original eigenvalue problem is projected onto a restricted subspace which ideally encapsulates the invariant subspace associated with the sought eigenvectors. Following ideas from eigenvector perturbation analysis, we present a new methodology to build the projection subspace. The proposed framework features lower computational and memory complexity with respect to competitive alternatives while empirical results show strong qualitative performance, both in terms of eigenvector approximation and accuracy of downstream learning tasks of central node identification and node clustering.

---

### When both Grounding and not Grounding are Bad -- A Partially Grounded Encoding of Planning into SAT (Extended Version)

Authors: João Filipe, Gregor Behnke
Published: 2026-03-19
Categories: cs.AI, cs.LO, cs.SC
arXiv: https://arxiv.org/abs/2603.19429v1

Classical planning problems are typically defined using lifted first-order representations, which offer compactness and generality. While most planners ground these representations to simplify reasoning, this can cause an exponential blowup in size. Recent approaches instead operate directly on the lifted level to avoid full grounding. We explore a middle ground between fully lifted and fully grounded planning by introducing three SAT encodings that keep actions lifted while partially grounding predicates. Unlike previous SAT encodings, which scale quadratically with plan length, our approach scales linearly, enabling better performance on longer plans. Empirically, our best encoding outperforms the state of the art in length-optimal planning on hard-to-ground domains.

---

### Vocabulary shapes cross-lingual variation of word-order learnability in language models

Authors: Jonas Mayer Martins, Jaap Jumelet, Viola Priesemann, Lisa Beinborn
Published: 2026-03-19
Categories: cs.CL, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2603.19427v1

Why do some languages like Czech permit free word order, while others like English do not? We address this question by pretraining transformer language models on a spectrum of synthetic word-order variants of natural languages. We observe that greater word-order irregularity consistently raises model surprisal, indicating reduced learnability. Sentence reversal, however, affects learnability only weakly. A coarse distinction of free- (e.g., Czech and Finnish) and fixed-word-order languages (e.g., English and French) does not explain cross-lingual variation. Instead, the structure of the word and subword vocabulary strongly predicts the model surprisal. Overall, vocabulary structure emerges as a key driver of computational word-order learnability across languages.

---

### Is Evaluation Awareness Just Format Sensitivity? Limitations of Probe-Based Evidence under Controlled Prompt Structure

Authors: Viliana Devbunova
Published: 2026-03-19
Categories: cs.CL, cs.AI
arXiv: https://arxiv.org/abs/2603.19426v1

Prior work uses linear probes on benchmark prompts as evidence of evaluation awareness in large language models. Because evaluation context is typically entangled with benchmark format and genre, it is unclear whether probe-based signals reflect context or surface structure. We test whether these signals persist under partial control of prompt format using a controlled 2x2 dataset and diagnostic rewrites. We find that probes primarily track benchmark-canonical structure and fail to generalize to free-form prompts independent of linguistic style. Thus, standard probe-based methodologies do not reliably disentangle evaluation context from structural artifacts, limiting the evidential strength of existing results.

---

### The Autonomy Tax: Defense Training Breaks LLM Agents

Authors: Shawn Li, Yue Zhao
Published: 2026-03-19
Categories: cs.CR, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2603.19423v3

Large language model (LLM) agents increasingly rely on external tools (file operations, API calls, database transactions) to autonomously complete complex multi-step tasks. Practitioners deploy defense-trained models to protect against prompt injection attacks that manipulate agent behavior through malicious observations or retrieved content. We reveal a fundamental \textbf{capability-alignment paradox}: defense training designed to improve safety systematically destroys agent competence while failing to prevent sophisticated attacks. Evaluating defended models against undefended baselines across 97 agent tasks and 1,000 adversarial prompts, we uncover three systematic biases unique to multi-step agents. \textbf{Agent incompetence bias} manifests as immediate tool execution breakdown, with models refusing or generating invalid actions on benign tasks before observing any external content. \textbf{Cascade amplification bias} causes early failures to propagate through retry loops, pushing defended models to timeout on 99\% of tasks compared to 13\% for baselines. \textbf{Trigger bias} leads to paradoxical security degradation where defended models perform worse than undefended baselines while straightforward attacks bypass defenses at high rates. Root cause analysis reveals these biases stem from shortcut learning: models overfit to surface attack patterns rather than semantic threat understanding, evidenced by extreme variance in defense effectiveness across attack categories. Our findings demonstrate that current defense paradigms optimize for single-turn refusal benchmarks while rendering multi-step agents fundamentally unreliable, necessitating new approaches that preserve tool execution competence under adversarial conditions.

---

### The Autonomy Tax: Defense Training Breaks LLM Agents

Authors: Shawn Li, Yue Zhao
Published: 2026-03-19
Categories: cs.CR, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2603.19423v2

Large language model (LLM) agents increasingly rely on external tools (file operations, API calls, database transactions) to autonomously complete complex multi-step tasks. Practitioners deploy defense-trained models to protect against prompt injection attacks that manipulate agent behavior through malicious observations or retrieved content. We reveal a fundamental \textbf{capability-alignment paradox}: defense training designed to improve safety systematically destroys agent competence while failing to prevent sophisticated attacks. Evaluating defended models against undefended baselines across 97 agent tasks and 1,000 adversarial prompts, we uncover three systematic biases unique to multi-step agents. \textbf{Agent incompetence bias} manifests as immediate tool execution breakdown, with models refusing or generating invalid actions on benign tasks before observing any external content. \textbf{Cascade amplification bias} causes early failures to propagate through retry loops, pushing defended models to timeout on 99\% of tasks compared to 13\% for baselines. \textbf{Trigger bias} leads to paradoxical security degradation where defended models perform worse than undefended baselines while straightforward attacks bypass defenses at high rates. Root cause analysis reveals these biases stem from shortcut learning: models overfit to surface attack patterns rather than semantic threat understanding, evidenced by extreme variance in defense effectiveness across attack categories. Our findings demonstrate that current defense paradigms optimize for single-turn refusal benchmarks while rendering multi-step agents fundamentally unreliable, necessitating new approaches that preserve tool execution competence under adversarial conditions.

---

### Pseudo-Labeling for Unsupervised Domain Adaptation with Kernel GLMs

Authors: Nathan Weill, Kaizheng Wang
Published: 2026-03-19
Categories: stat.ML, cs.LG, math.ST
arXiv: https://arxiv.org/abs/2603.19422v1

We propose a principled framework for unsupervised domain adaptation under covariate shift in kernel Generalized Linear Models (GLMs), encompassing kernelized linear, logistic, and Poisson regression with ridge regularization. Our goal is to minimize prediction error in the target domain by leveraging labeled source data and unlabeled target data, despite differences in covariate distributions. We partition the labeled source data into two batches: one for training a family of candidate models, and the other for building an imputation model. This imputation model generates pseudo-labels for the target data, enabling robust model selection. We establish non-asymptotic excess-risk bounds that characterize adaptation performance through an "effective labeled sample size", explicitly accounting for the unknown covariate shift. Experiments on synthetic and real datasets demonstrate consistent performance gains over source-only baselines.

---

### Investigating In-Context Privacy Learning by Integrating User-Facing Privacy Tools into Conversational Agents

Authors: Mohammad Hadi Nezhad, Francisco Enrique Vicente Castro, Ivon Arroyo
Published: 2026-03-19
Categories: cs.HC, cs.AI
arXiv: https://arxiv.org/abs/2603.19416v1

Supporting users in protecting sensitive information when using conversational agents (CAs) is crucial, as users may undervalue privacy protection due to outdated, partial, or inaccurate knowledge about privacy in CAs. Although privacy knowledge can be developed through standalone resources, it may not readily translate into practice and may remain detached from real-time contexts of use. In this study, we investigate in-context, experiential learning by examining how interactions with privacy tools during chatbot use enhance users' privacy learning. We also explore interface design features that facilitate engagement with these tools and learning about privacy by simulating ChatGPT's interface which we integrated with a just-in-time privacy notice panel. The panel intercepts messages containing sensitive information, warns users about potential sensitivity, offers protective actions, and provides FAQs about privacy in CAs. Participants used versions of the chatbot with and without the privacy panel across two task sessions designed to approximate realistic chatbot use. We qualitatively analyzed participants' pre- and post-test survey responses and think-aloud transcripts and describe findings related to (a) participants' perceptions of privacy before and after the task sessions and (b) interface design features that supported or hindered user-led protection of sensitive information. Finally, we discuss future directions for designing user-facing privacy tools in CAs that promote privacy learning and user engagement in protecting privacy in CAs.

---

### Scalable Prompt Routing via Fine-Grained Latent Task Discovery

Authors: Yunyi Zhang, Soji Adeshina, Sheng Guan, Ashwin Ganesh, Zhen Han, Vassilis N. Ioannidis, Huzefa Rangwala, George Karypis
Published: 2026-03-19
Categories: cs.CL, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2603.19415v2

Prompt routing dynamically selects the most appropriate large language model from a pool of candidates for each query, optimizing performance while managing costs. As model pools scale to include dozens of frontier models with narrow performance gaps, existing approaches face significant challenges: manually defined task taxonomies cannot capture fine-grained capability distinctions, while monolithic routers struggle to differentiate subtle differences across diverse tasks. We propose a two-stage routing architecture that addresses these limitations through automated fine-grained task discovery and task-aware quality estimation. Our first stage employs graph-based clustering to discover latent task types and trains a classifier to assign prompts to discovered tasks. The second stage uses a mixture-of-experts architecture with task-specific prediction heads for specialized quality estimates. At inference, we aggregate predictions from both stages to balance task-level stability with prompt-specific adaptability. Evaluated on 10 benchmarks with 11 frontier models, our method consistently outperforms existing baselines and surpasses the strongest individual model while incurring less than half its cost.

---

### Optimizing Resource-Constrained Non-Pharmaceutical Interventions for Multi-Cluster Outbreak Control Using Hierarchical Reinforcement Learning

Authors: Xueqiao Peng, Andrew Perrault
Published: 2026-03-19
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.19397v2

Non-pharmaceutical interventions (NPIs), such as diagnostic testing and quarantine, are crucial for controlling infectious disease outbreaks but are often constrained by limited resources, particularly in early outbreak stages. In real-world public health settings, resources must be allocated across multiple outbreak clusters that emerge asynchronously, vary in size and risk, and compete for a shared resource budget. Here, a cluster corresponds to a group of close contacts generated by a single infected index case. Thus, decisions must be made under uncertainty and heterogeneous demands, while respecting operational constraints. We formulate this problem as a constrained restless multi-armed bandit and propose a hierarchical reinforcement learning framework. A global controller learns a continuous action cost multiplier that adjusts global resource demand, while a generalized local policy estimates the marginal value of allocating resources to individuals within each cluster. We evaluate the proposed framework in a realistic agent-based simulator of SARS-CoV-2 with dynamically arriving clusters. Across a wide range of system scales and testing budgets, our method consistently outperforms RMAB-inspired and heuristic baselines, improving outbreak control effectiveness by 20%-30%. Experiments on up to 40 concurrently active clusters further demonstrate that the hierarchical framework is highly scalable and enables faster decision-making than the RMAB-inspired method.

---

### Bridging Conformal Prediction and Scenario Optimization: Discarded Constraints and Modular Risk Allocation

Authors: Giuseppe C. Calafiore
Published: 2026-03-19
Categories: eess.SY, cs.LG
arXiv: https://arxiv.org/abs/2603.19396v1

Scenario optimization and conformal prediction share a common goal, that is, turning finite samples into safety margins. Yet, different terminology often obscures the connection between their respective guarantees. This paper revisits that connection directly from a systems-and-control viewpoint. Building on the recent conformal/scenario bridge of \citet{OSullivanRomaoMargellos2026}, we extend the forward direction to feasible sample-and-discard scenario algorithms. Specifically, if the final decision is determined by a stable subset of the retained sampled constraints, the classical mean violation law admits a direct exchangeability-based derivation. In this view, discarded samples naturally appear as admissible exceptions. We also introduce a simple modular composition rule that combines several blockwise calibration certificates into a single joint guarantee. This rule proves particularly useful in multi-output prediction and finite-horizon control, where engineers must distribute risk across coordinates, constraints, or prediction steps. Finally, we provide numerical illustrations using a calibrated multi-step tube around an identified predictor. These examples compare alternative stage-wise risk allocations and highlight the resulting performance and safety trade-offs in a standard constraint-tightening problem.

---

### TuLaBM: Tumor-Biased Latent Bridge Matching for Contrast-Enhanced MRI Synthesis

Authors: Atharva Rege, Adinath Madhavrao Dukre, Numan Balci, Dwarikanath Mahapatra, Imran Razzak
Published: 2026-03-19
Categories: eess.IV, cs.LG
arXiv: https://arxiv.org/abs/2603.19386v1

Contrast-enhanced magnetic resonance imaging (CE-MRI) plays a crucial role in brain tumor assessment; however, its acquisition requires gadolinium-based contrast agents (GBCAs), which increase costs and raise safety concerns. Consequently, synthesizing CE-MRI from non-contrast MRI (NC-MRI) has emerged as a promising alternative. Early Generative Adversarial Network (GAN)-based approaches suffered from instability and mode collapse, while diffusion models, despite impressive synthesis quality, remain computationally expensive and often fail to faithfully reproduce critical tumor contrast patterns. To address these limitations, we propose Tumor-Biased Latent Bridge Matching (TuLaBM), which formulates NC-to-CE MRI translation as Brownian bridge transport between source and target distributions in a learned latent space, enabling efficient training and inference. To enhance tumor-region fidelity, we introduce a Tumor-Biased Attention Mechanism (TuBAM) that amplifies tumor-relevant latent features during bridge evolution, along with a boundary-aware loss that constrains tumor interfaces to improve margin sharpness. While bridge matching has been explored for medical image translation in pixel space, our latent formulation substantially reduces computational cost and inference time. Experiments on BraTS2023-GLI (BraSyn) and Cleveland Clinic (in-house) liver MRI dataset show that TuLaBM consistently outperforms state-of-the-art baselines on both whole-image and tumor-region metrics, generalizes effectively to unseen liver MRI data in zero-shot and fine-tuned settings, and achieves inference times under 0.097 seconds per image.

---

### Automated Membership Inference Attacks (AutoMIA): Discovering MIA Signal Computations using LLM Agents

Authors: Toan Tran, Olivera Kotevska, Li Xiong
Published: 2026-03-19
Categories: cs.CR, cs.LG
arXiv: https://arxiv.org/abs/2603.19375v2

Membership inference attacks (MIAs), which enable adversaries to determine whether specific data points were part of a model's training dataset, have emerged as an important framework to understand, assess, and quantify the potential information leakage associated with machine learning systems. Designing effective MIAs is a challenging task that usually requires extensive manual exploration of model behaviors to identify potential vulnerabilities. In this paper, we introduce AutoMIA -- a novel framework that leverages large language model (LLM) agents to automate the design and implementation of new MIA signal computations. By utilizing LLM agents, we can systematically explore a vast space of potential attack strategies, enabling the discovery of novel strategies. Our experiments demonstrate AutoMIA can successfully discover new MIAs that are specifically tailored to user-configured target model and dataset, resulting in improvements of up to 0.18 in absolute AUC over existing MIAs. This work provides the first demonstration that LLM agents can serve as an effective and scalable paradigm for designing and implementing MIAs with SOTA performance, opening up new avenues for future exploration.

---

### Automated Membership Inference Attacks: Discovering MIA Signal Computations using LLM Agents

Authors: Toan Tran, Olivera Kotevska, Li Xiong
Published: 2026-03-19
Categories: cs.CR, cs.LG
arXiv: https://arxiv.org/abs/2603.19375v1

Membership inference attacks (MIAs), which enable adversaries to determine whether specific data points were part of a model's training dataset, have emerged as an important framework to understand, assess, and quantify the potential information leakage associated with machine learning systems. Designing effective MIAs is a challenging task that usually requires extensive manual exploration of model behaviors to identify potential vulnerabilities. In this paper, we introduce AutoMIA -- a novel framework that leverages large language model (LLM) agents to automate the design and implementation of new MIA signal computations. By utilizing LLM agents, we can systematically explore a vast space of potential attack strategies, enabling the discovery of novel strategies. Our experiments demonstrate AutoMIA can successfully discover new MIAs that are specifically tailored to user-configured target model and dataset, resulting in improvements of up to 0.18 in absolute AUC over existing MIAs. This work provides the first demonstration that LLM agents can serve as an effective and scalable paradigm for designing and implementing MIAs with SOTA performance, opening up new avenues for future exploration.

---

### Factored Levenberg-Marquardt for Diffeomorphic Image Registration: An efficient optimizer for FireANTs

Authors: Rohit Jena, Pratik Chaudhari, James C. Gee
Published: 2026-03-19
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.19371v1

FireANTs introduced a novel Eulerian descent method for plug-and-play behavior with arbitrary optimizers adapted for diffeomorphic image registration as a test-time optimization problem, with a GPU-accelerated implementation. FireANTs uses Adam as its default optimizer for fast and more robust optimization. However, Adam requires storing state variables (i.e. momentum and squared-momentum estimates), each of which can consume significant memory, prohibiting its use for significantly large images. In this work, we propose a modified Levenberg-Marquardt (LM) optimizer that requires only a single scalar damping parameter as optimizer state, that is adaptively tuned using a trust region approach. The resulting optimizer reduces memory by up to 24.6% for large volumes, and retaining performance across all four datasets. A single hyperparameter configuration tuned on brain MRI transfers without modification to lung CT and cross-modal abdominal registration, matching or outperforming Adam on three of four benchmarks. We also perform ablations on the effectiveness of using Metropolis-Hastings style rejection step to prevent updates that worsen the loss function.

---

### AURORA: Adaptive Unified Representation for Robust Ultrasound Analysis

Authors: Ufaq Khan, L. D. M. S. Sai Teja, Ayuba Shakiru, Mai A. Shaaban, Yutong Xie, Muhammad Bilal, Muhammad Haris Khan
Published: 2026-03-19
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.19364v1

Ultrasound images vary widely across scanners, operators, and anatomical targets, which often causes models trained in one setting to generalize poorly to new hospitals and clinical conditions. The Foundation Model Challenge for Ultrasound Image Analysis (FMC-UIA) reflects this difficulty by requiring a single model to handle multiple tasks, including segmentation, detection, classification, and landmark regression across diverse organs and datasets. We propose a unified multi-task framework based on a transformer visual encoder from the Qwen3-VL family. Intermediate token features are projected into spatial feature maps and fused using a lightweight multi-scale feature pyramid, enabling both pixel-level predictions and global reasoning within a shared representation. Each task is handled by a small task-specific prediction head, while training uses task-aware sampling and selective loss balancing to manage heterogeneous supervision and reduce task imbalance. Our method is designed to be simple to optimize and adaptable across a wide range of ultrasound analysis tasks. The performance improved from 67% to 85% on the validation set and achieved an average score of 81.84% on the official test set across all tasks. The code is publicly available at: https://github.com/saitejalekkala33/FMCUIA-ISBI.git

---

### Warm-Start Flow Matching for Guaranteed Fast Text/Image Generation

Authors: Minyoung Kim
Published: 2026-03-19
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.19360v1

Current auto-regressive (AR) LLMs, diffusion-based text/image generative models, and recent flow matching (FM) algorithms are capable of generating premium quality text/image samples. However, the inference or sample generation in these models is often very time-consuming and computationally demanding, mainly due to large numbers of function evaluations corresponding to the lengths of tokens or the numbers of diffusion steps. This also necessitates heavy GPU resources, time, and electricity. In this work we propose a novel solution to reduce the sample generation time of flow matching algorithms by a guaranteed speed-up factor, without sacrificing the quality of the generated samples. Our key idea is to utilize computationally lightweight generative models whose generation time is negligible compared to that of the target AR/FM models. The draft samples from a lightweight model, whose quality is not satisfactory but fast to generate, are regarded as an initial distribution for a FM algorithm. Unlike conventional usage of FM that takes a pure noise (e.g., Gaussian or uniform) initial distribution, the draft samples are already of decent quality, so we can set the starting time to be closer to the end time rather than 0 in the pure noise FM case. This will significantly reduce the number of time steps to reach the target data distribution, and the speed-up factor is guaranteed. Our idea, dubbed {\em Warm-Start FM} or WS-FM, can essentially be seen as a {\em learning-to-refine} generative model from low-quality draft samples to high-quality samples. As a proof of concept, we demonstrate the idea on some synthetic toy data as well as real-world text and image generation tasks, illustrating that our idea offers guaranteed speed-up in sample generation without sacrificing the quality of the generated samples.

---

### A Novel Solution for Zero-Day Attack Detection in IDS using Self-Attention and Jensen-Shannon Divergence in WGAN-GP

Authors: Ziyu Mu, Xiyu Shi, Safak Dogan
Published: 2026-03-19
Categories: cs.CR, cs.AI, cs.LG
arXiv: https://arxiv.org/abs/2603.19350v1

The increasing sophistication of cyber threats, especially zero-day attacks, poses a significant challenge to cybersecurity. Zero-day attacks exploit unknown vulnerabilities, making them difficult to detect and defend against. Existing approaches patch flaws and deploy an Intrusion Detection System (IDS). Using advanced Wasserstein GANs with Gradient Penalty (WGAN-GP), this paper makes a novel proposition to synthesize network traffic that mimics zero-day patterns, enriching data diversity and improving IDS generalization. SA-WGAN-GP is first introduced, which adds a Self-Attention (SA) mechanism to capture long-range cross-feature dependencies by reshaping the feature vector into tokens after dense projections. A JS-WGAN-GP is then proposed, which adds a Jensen-Shannon (JS) divergence-based auxiliary discriminator that is trained with Binary Cross-Entropy (BCE), frozen during updates, and used to regularize the generator for smoother gradients and higher sample quality. Third, SA-JS-WGAN-GP is created by combining the SA mechanism with JS divergence, thereby enhancing the data generation ability of WGAN-GP. As data augmentation does not equate with true zero-day attack discovery, we emulate zero-day attacks via the leave-one-attack-type-out method on the NSL-KDD dataset for training all GANs and IDS models in the assessment of the effectiveness of the proposed solution. The evaluation results show that integrating SA and JS divergence into WGAN-GP yields superior IDS performance and more effective zero-day risk detection.

---

### A Mathematical Theory of Understanding

Authors: Bahar Taşkesen
Published: 2026-03-19
Categories: cs.LG, cs.IT, econ.TH
arXiv: https://arxiv.org/abs/2603.19349v1

Generative AI has transformed the economics of information production, making explanations, proofs, examples, and analyses available at very low cost. Yet the value of information still depends on whether downstream users can absorb and act on it. A signal conveys meaning only to a learner with the structural capacity to decode it: an explanation that clarifies a concept for one user may be indistinguishable from noise to another who lacks the relevant prerequisites. This paper develops a mathematical model of that learner-side bottleneck. We model the learner as a mind, an abstract learning system characterized by a prerequisite structure over concepts. A mind may represent a human learner, an artificial learner such as a neural network, or any agent whose ability to interpret signals depends on previously acquired concepts. Teaching is modeled as sequential communication with a latent target. Because instructional signals are usable only when the learner has acquired the prerequisites needed to parse them, the effective communication channel depends on the learner's current state of knowledge and becomes more informative as learning progresses. The model yields two limits on the speed of learning and adoption: a structural limit determined by prerequisite reachability and an epistemic limit determined by uncertainty about the target. The framework implies threshold effects in training and capability acquisition. When the teaching horizon lies below the prerequisite depth of the target, additional instruction cannot produce successful completion of teaching; once that depth is reached, completion becomes feasible. Across heterogeneous learners, a common broadcast curriculum can be slower than personalized instruction by a factor linear in the number of learner types.

---

### Anatomical Heterogeneity in Transformer Language Models

Authors: Tomasz Wietrzykowski
Published: 2026-03-19
Categories: cs.LG, cs.CL
arXiv: https://arxiv.org/abs/2603.19348v1

Current transformer language models are trained with uniform computational budgets across all layers, implicitly assuming layer homogeneity. We challenge this assumption through empirical analysis of SmolLM2-135M, a 30-layer, 135M-parameter causal language model, using five diagnostic metrics: weight predictability (R2), ablation degradation, recovery speed, weight manipulation robustness, and structural analysis. We find profound anatomical heterogeneity: (1) Layer weights follow strong mathematical regularity (R2 = 0.91) with a universal oscillatory delta pattern (correlation ~= -0.50), yet predicted weights cause catastrophic failure due to nonlinear error accumulation. (2) Layer importance spans a 10^7 range, from a critical core (L8-11, up to +63,419% PPL degradation) to anti-layers (L14, L17) whose removal improves performance. (3) Recovery speed correlates with layer importance, indicating differential training requirements. (4) Only weight scaling (alpha = 0.9) preserves model quality among five tested manipulation strategies. (5) Growth Transformer Training, allocating budget by layer importance, achieves ~54% cost reduction. A proof-of-concept experiment confirms this: 4.7x lower validation loss than uniform training at identical parameter count, while being 13% faster.

---

### Exploring the Agentic Frontier of Verilog Code Generation

Authors: Patrick Yubeaton, Siddharth Garg, Chinmay Hegde
Published: 2026-03-19
Categories: cs.AR, cs.LG
arXiv: https://arxiv.org/abs/2603.19347v3

Large language models (LLMs) have made rapid advancements in code generation for popular languages such as Python and C++. Many of these recent gains can be attributed to the use of ``agents'' that wrap domain-relevant tools alongside LLMs. Hardware design languages such as Verilog have also seen improved code generation in recent years, but the impact of agentic frameworks on Verilog code generation tasks remains unclear. In this work, we present the first systematic evaluation of agentic LLMs for Verilog generation, using the recently introduced CVDP benchmark. We also introduce several open-source hardware design agent harnesses, providing a model-agnostic baseline for future work. Through controlled experiments across frontier models, we study how structured prompting and tool design affect performance, analyze agent failure modes and tool usage patterns, compare open-source and closed-source models, and provide qualitative examples of successful and failed agent runs. Our results show that naive agentic wrapping around frontier models can degrade performance (relative to standard forward passes with optimized prompts), but that structured harnesses meaningfully match and in some cases exceed non-agentic baselines. We find that the performance gap between open and closed source models is driven by both higher crash rates and weaker tool output interpretation. Our exploration illuminates the path towards designing special-purpose agents for verilog generation in the future.

---

### Beyond Weighted Summation: Learnable Nonlinear Aggregation Functions for Robust Artificial Neurons

Authors: Berke Deniz Bozyigit
Published: 2026-03-19
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2603.19344v1

Weighted summation has remained the default input aggregation mechanism in artificial neurons since the earliest neural network models. While computationally efficient, this design implicitly behaves like a mean-based estimator and is therefore sensitive to noisy or extreme inputs. This paper investigates whether replacing fixed linear aggregation with learnable nonlinear alternatives can improve neural network robustness without sacrificing trainability. Two differentiable aggregation mechanisms are introduced: an F-Mean neuron based on a learnable power-weighted aggregation rule, and a Gaussian Support neuron based on distance-aware affinity weighting. To preserve the optimisation stability of standard neurons, hybrid neurons are proposed that interpolate between linear and nonlinear aggregation through a learnable blending parameter. Evaluated in multilayer perceptrons and convolutional neural networks on CIFAR-10 and a noisy CIFAR-10 variant with additive Gaussian corruption, hybrid neurons consistently improve robustness under noise while F-Mean hybrids also yield modest gains on clean data. The three-way hybrid achieves robustness scores of up to 0.991 compared to 0.890 for the standard baseline, and learned parameters converge consistently to sub-linear aggregation (p $\approx$ 0.43--0.50) and high novelty utilisation ($α$ $\approx$ 0.69--0.79). These findings suggest that neuron-level aggregation is a meaningful and underexplored design dimension for building more noise-tolerant neural networks.

---

### Spectral Tempering for Embedding Compression in Dense Passage Retrieval

Authors: Yongkang Li, Panagiotis Eustratiadis, Evangelos Kanoulas
Published: 2026-03-19
Categories: cs.IR, cs.AI, cs.CL
arXiv: https://arxiv.org/abs/2603.19339v2

Dimensionality reduction is critical for deploying dense retrieval systems at scale, yet mainstream post-hoc methods face a fundamental trade-off: principal component analysis (PCA) preserves dominant variance but underutilizes representational capacity, while whitening enforces isotropy at the cost of amplifying noise in the heavy-tailed eigenspectrum of retrieval embeddings. Intermediate spectral scaling methods unify these extremes by reweighting dimensions with a power coefficient $γ$, but treat $γ$ as a fixed hyperparameter that requires task-specific tuning. We show that the optimal scaling strength $γ$ is not a global constant: it varies systematically with target dimensionality $k$ and is governed by the signal-to-noise ratio (SNR) of the retained subspace. Based on this insight, we propose Spectral Tempering (\textbf{SpecTemp}), a learning-free method that derives an adaptive $γ(k)$ directly from the corpus eigenspectrum using local SNR analysis and knee-point normalization, requiring no labeled data or validation-based search. Extensive experiments demonstrate that Spectral Tempering consistently achieves near-oracle performance relative to grid-searched $γ^*(k)$ while remaining fully learning-free and model-agnostic. Our code is publicly available at https://github.com/liyongkang123/SpecTemp.

---

### DAPA: Distribution Aware Piecewise Activation Functions for On-Device Transformer Inference and Training

Authors: Maoyang Xiang, Bo Wang
Published: 2026-03-19
Categories: cs.LG
arXiv: https://arxiv.org/abs/2603.19338v1

Non-linear activation functions play a pivotal role in on-device inference and training, as they not only consume substantial hardware resources but also impose a significant impact on system performance and energy efficiency. In this work, we propose Distribution-Aware Piecewise Activation (DAPA), a differentiable and hardware-friendly activation function for Transformer architectures by exploiting the distribution of pre-activation data. DAPA employs a non-uniform piecewise approximation that allocates finer segments to high-probability regions of the distribution, improving generalizability over prior piecewise linear methods. The resulting approximation is further quantized using Distribution-Weighted Mean Square Error to reduce latency and resource utilization for hardware deployment. Our HLS implementation demonstrates that DAPA speeds up GELU computation by 16$\times$ and decreases DSP utilization by 16$\times$ while maintaining comparable or better performance across vision Transformers and GPT-2 models.

---

### Diffusion-Guided Semantic Consistency for Multimodal Heterogeneity

Authors: Jing Liu, Zhengliang Guo, Yan Wang, Xiaoguang Zhu, Yao Du, Zehua Wang, Victor C. M. Leung
Published: 2026-03-19
Categories: cs.CV, cs.AI
arXiv: https://arxiv.org/abs/2603.19337v1

Federated learning (FL) is severely challenged by non-independent and identically distributed (non-IID) client data, a problem that degrades global model performance, especially in multimodal perception settings. Conventional methods often fail to address the underlying semantic discrepancies between clients, leading to suboptimal performance for multimedia systems requiring robust perception. To overcome this, we introduce SemanticFL, a novel framework that leverages the rich semantic representations of pre-trained diffusion models to provide privacy-preserving guidance for local training. Our approach leverages multi-layer semantic representations from a pre-trained Stable Diffusion model (including VAE-encoded latents and U-Net hierarchical features) to create a shared latent space that aligns heterogeneous clients, facilitated by an efficient client-server architecture that offloads heavy computation to the server. A unified consistency mechanism, employing cross-modal contrastive learning, further stabilizes convergence. We conduct extensive experiments on benchmarks including CIFAR-10, CIFAR-100, and TinyImageNet under diverse heterogeneity scenarios. Our results demonstrate that SemanticFL surpasses existing federated learning approaches, achieving accuracy gains of up to 5.49% over FedAvg, validating its effectiveness in learning robust representations for heterogeneous and multimodal data for perception tasks.

---

### Do Post-Training Algorithms Actually Differ? A Controlled Study Across Model Scales Uncovers Scale-Dependent Ranking Inversions

Authors: Xiaoyi Li
Published: 2026-03-19
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2603.19335v1

Post-training alignment has produced dozens of competing algorithms -- DPO, SimPO, KTO, GRPO, and others -- yet practitioners lack controlled comparisons to guide algorithm selection. We present OXRL, a unified framework implementing 51 post-training algorithms with identical infrastructure, enabling the first large-scale apples-to-apples evaluation. Our study spans 8 algorithms across 4 model scales (0.5B--7B), 3 evaluation domains, and a 20-variant DPO taxonomy (100 runs at 1.5B, 5 seeds each), totaling $\sim$240 training runs on H100 GPUs. Three headline findings emerge. (1)~Algorithm rankings are unstable across scale: at 1.5B, online RL (SGRPO) tops all methods at 58.0\%~$\pm$0.57 on GSM8K; by 7B, the worst small-scale method (SimPO) becomes the best (85.8\%), a complete ranking inversion driven by model scale rather than LoRA regularization (confirmed via 2$\times$2 factorial). (2)~Loss function modifications yield negligible gains: none of 20 DPO variants significantly outperform vanilla DPO after Bonferroni correction; the sole significant outlier, SimPO, is worse ($-$11.5~pp, $p < 10^{-4}$). (3)~Algorithm leverage is task-specific: the 19.3~pp GSM8K spread collapses to 0.54~pp on MATH ($36\times$) and 0.47~pp on general-domain benchmarks ($41\times$), confirming that algorithm choice matters primarily within the training distribution. These findings yield a hierarchy of leverage for practitioners: model scale (${\sim}$50~pp) $\gg$ training paradigm (${\sim}$10~pp) $\gg$ online vs.\ offline (${\sim}$9~pp) $\gg$ loss function (${\sim}$1~pp). We release all code, configs, and evaluation data as a living community benchmark.

---

### POET: Power-Oriented Evolutionary Tuning for LLM-Based RTL PPA Optimization

Authors: Heng Ping, Peiyu Zhang, Zhenkun Wang, Shixuan Li, Anzhe Cheng, Wei Yang, Paul Bogdan, Shahin Nazarian
Published: 2026-03-19
Categories: cs.AR, cs.AI
arXiv: https://arxiv.org/abs/2603.19333v1

Applying large language models (LLMs) to RTL code optimization for improved power, performance, and area (PPA) faces two key challenges: ensuring functional correctness of optimized designs despite LLM hallucination, and systematically prioritizing power reduction within the multi-objective PPA trade-off space. We propose POET (Power-Oriented Evolutionary Tuning), a framework that addresses both challenges. For functional correctness, POET introduces a differential-testing-based testbench generation pipeline that treats the original design as a functional oracle, using deterministic simulation to produce golden references and eliminating LLM hallucination from the verification process. For PPA optimization, POET employs an LLM-driven evolutionary mechanism with non-dominated sorting, power-first intra-level ranking, and proportional survivor selection to steer the search toward the low-power region of the Pareto front without manual weight tuning. Evaluated on the RTL-OPT benchmark across 40 diverse RTL designs, POET achieves 100% functional correctness, the best power on all 40 designs, and competitive area and delay improvements.

---

### Generation Models Know Space: Unleashing Implicit 3D Priors for Scene Understanding

Authors: Xianjin Wu, Dingkang Liang, Tianrui Feng, Kui Xia, Yumeng Zhang, Xiaofan Li, Xiao Tan, Xiang Bai
Published: 2026-03-19
Categories: cs.CV, cs.RO
arXiv: https://arxiv.org/abs/2603.19235v3

While Multimodal Large Language Models demonstrate impressive semantic capabilities, they often suffer from spatial blindness, struggling with fine-grained geometric reasoning and physical dynamics. Existing solutions typically rely on explicit 3D modalities or complex geometric scaffolding, which are limited by data scarcity and generalization challenges. In this work, we propose a paradigm shift by leveraging the implicit spatial prior within large-scale video generation models. We posit that to synthesize temporally coherent videos, these models inherently learn robust 3D structural priors and physical laws. We introduce VEGA-3D (Video Extracted Generative Awareness), a plug-and-play framework that repurposes a pre-trained video diffusion model as a Latent World Simulator. By extracting spatiotemporal features from intermediate noise levels and integrating them with semantic representations via a token-level adaptive gated fusion mechanism, we enrich MLLMs with dense geometric cues without explicit 3D supervision. Extensive experiments across 3D scene understanding, spatial reasoning, and embodied manipulation benchmarks demonstrate that our method outperforms state-of-the-art baselines, validating that generative priors provide a scalable foundation for physical-world understanding. Code is publicly available at https://github.com/H-EmbodVis/VEGA-3D.

---

### Matryoshka Gaussian Splatting

Authors: Zhilin Guo, Boqiao Zhang, Hakan Aktas, Kyle Fogarty, Jeffrey Hu, Nursena Koprucu Aslan, Wenzhao Li, Canberk Baykal, Albert Miao, Josef Bengtson, Chenliang Zhou, Weihao Xia et al.
Published: 2026-03-19
Categories: cs.CV, cs.GR
arXiv: https://arxiv.org/abs/2603.19234v2

The ability to render scenes at adjustable fidelity from a single model, known as level of detail (LoD), is crucial for practical deployment of 3D Gaussian Splatting (3DGS). Existing discrete LoD methods expose only a limited set of operating points, while concurrent continuous LoD approaches enable smoother scaling but often suffer noticeable quality degradation at full capacity, making LoD a costly design decision. We introduce Matryoshka Gaussian Splatting (MGS), a training framework that enables continuous LoD for standard 3DGS pipelines without sacrificing full-capacity rendering quality. MGS learns a single ordered set of Gaussians such that rendering any prefix, the first k splats, produces a coherent reconstruction whose fidelity improves smoothly with increasing budget. Our key idea is stochastic budget training: each iteration samples a random splat budget and optimises both the corresponding prefix and the full set. This strategy requires only two forward passes and introduces no architectural modifications. Experiments across four benchmarks and six baselines show that MGS matches the full-capacity performance of its backbone while enabling a continuous speed-quality trade-off from a single model. Extensive ablations on ordering strategies, training objectives, and model capacity further validate the designs.

---

### Cubic Discrete Diffusion: Discrete Visual Generation on High-Dimensional Representation Tokens

Authors: Yuqing Wang, Chuofan Ma, Zhijie Lin, Yao Teng, Lijun Yu, Shuai Wang, Jiaming Han, Jiashi Feng, Yi Jiang, Xihui Liu
Published: 2026-03-19
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.19232v1

Visual generation with discrete tokens has gained significant attention as it enables a unified token prediction paradigm shared with language models, promising seamless multimodal architectures. However, current discrete generation methods remain limited to low-dimensional latent tokens (typically 8-32 dims), sacrificing the semantic richness essential for understanding. While high-dimensional pretrained representations (768-1024 dims) could bridge this gap, their discrete generation poses fundamental challenges. In this paper, we present Cubic Discrete Diffusion (CubiD), the first discrete generation model for high-dimensional representations. CubiD performs fine-grained masking throughout the high-dimensional discrete representation -- any dimension at any position can be masked and predicted from partial observations. This enables the model to learn rich correlations both within and across spatial positions, with the number of generation steps fixed at $T$ regardless of feature dimensionality, where $T \ll hwd$. On ImageNet-256, CubiD achieves state-of-the-art discrete generation with strong scaling behavior from 900M to 3.7B parameters. Crucially, we validate that these discretized tokens preserve original representation capabilities, demonstrating that the same discrete tokens can effectively serve both understanding and generation tasks. We hope this work will inspire future research toward unified multimodal architectures. Code is available at: https://github.com/YuqingWang1029/CubiD.

---

### MonoArt: Progressive Structural Reasoning for Monocular Articulated 3D Reconstruction

Authors: Haitian Li, Haozhe Xie, Junxiang Xu, Beichen Wen, Fangzhou Hong, Ziwei Liu
Published: 2026-03-19
Categories: cs.CV
arXiv: https://arxiv.org/abs/2603.19231v1

Reconstructing articulated 3D objects from a single image requires jointly inferring object geometry, part structure, and motion parameters from limited visual evidence. A key difficulty lies in the entanglement between motion cues and object structure, which makes direct articulation regression unstable. Existing methods address this challenge through multi-view supervision, retrieval-based assembly, or auxiliary video generation, often sacrificing scalability or efficiency. We present MonoArt, a unified framework grounded in progressive structural reasoning. Rather than predicting articulation directly from image features, MonoArt progressively transforms visual observations into canonical geometry, structured part representations, and motion-aware embeddings within a single architecture. This structured reasoning process enables stable and interpretable articulation inference without external motion templates or multi-stage pipelines. Extensive experiments on PartNet-Mobility demonstrate that OM achieves state-of-the-art performance in both reconstruction accuracy and inference speed. The framework further generalizes to robotic manipulation and articulated scene reconstruction.

---