Explore the Frontier of Global Academia

01.

arXiv (CS.AI) 2026-06-15 DOI: arXiv:2606.14117

A Two-Stage Statistical Framework for Evaluating Associative Interference in Large Language Models

Authors:

Achraf Cohen↗Andrew Kincaid↗

arXiv:2606.14117v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly evaluated for bias using adaptations of human psychological paradigms, yet methodological limitations-particularly the conflation of refusal behavior with task performance-have hindered clear interpretation. Here, we adapt the Implicit Association Test (IAT) to a controlled, forced-choice framework and introduce a two-stage modeling approach that separates response compliance from task-consistent classification. Across three contemporary LLMs (Claude Sonnet-4, Gemini 2.5 Pro, and GPT-5), we evaluate associative interference, defined as reduced task-consistency in incongruent relative to congruent conditions. While compliance with the structured response format was uniformly high, interference effects varied substantially across models and domains. Claude Sonnet-4 exhibited strong interference in the Gender–Career domain (DeltaP = 0.086, 95% CrI [0.026, 0.173]) and smaller but credible effects in Gender–Science. Gemini 2.5 Pro showed attenuated interference, and GPT-5 exhibited minimal or no detectable interference across domains. These findings demonstrate that IAT-style associative asymmetries are not a universal property of LLMs, but instead depend on model-specific characteristics. By isolating interference from compliance and modeling item-level variability, this study provides a principled framework for evaluating structured response patterns in LLMs. The results highlight the importance of model-specific assessment and suggest that associative interference can be substantially mitigated in modern systems.

Read & Discuss → View Source →

02.

arXiv (CS.LG) 2026-06-17 DOI: arXiv:2511.01352

MiniFool – Physics-Constraint-Aware Minimizer-Based Adversarial Attacks in Deep Neural Networks

Authors:

Lucie Flek↗Oliver Janik↗Philipp Alexander Jung↗Akbar Karimi↗Timo Saala↗Alexander Schmidt↗Matthias Schott↗Philipp Soldin↗Matthias Thiesmeyer↗Christopher Wiebusch↗Ulrich Willemsen↗

arXiv:2511.01352v2 Announce Type: replace Abstract: In this paper, we present a new algorithm, MiniFool, that implements physics-inspired adversarial attacks for testing neural network-based classification tasks in particle and astroparticle physics. While we initially developed the algorithm for the search for astrophysical tau neutrinos with the IceCube Neutrino Observatory, we apply it to further data from other science domains, thus demonstrating its general applicability. Here, we apply the algorithm to the well-known MNIST data set and furthermore, to Open Data data from the CMS experiment at the Large Hadron Collider. The algorithm is based on minimizing a cost function that combines a $\chi^2$ based test-statistic with the deviation from the desired target score. The test statistic quantifies the probability of the perturbations applied to the data based on the experimental uncertainties. For our studied use cases, we find that the likelihood of a flipped classification differs for both the initially correctly and incorrectly classified events. When testing changes of the classifications as a function of an attack parameter that scales the experimental uncertainties, the robustness of the network decision can be quantified. Furthermore, this allows testing the robustness of the classification of unlabeled experimental data.

Read & Discuss → View Source →

03.

arXiv (CS.LG) 2026-06-19 DOI: arXiv:2510.06048

BLISS: A Lightweight Bilevel Influence Scoring Method for Data Selection in Language Model Pretraining

Authors:

Jie Hao↗Rui Yu↗Wei Zhang↗Huixia Wang↗Jie Xu↗Mingrui Liu↗

arXiv:2510.06048v5 Announce Type: replace Abstract: Effective data selection is essential for pretraining large language models (LLMs), enhancing efficiency and improving generalization to downstream tasks. However, existing approaches often require leveraging external pretrained models, making it difficult to disentangle the effects of data selection from those of the external pretrained models. In addition, they often overlook the long-term impact of selected data if the model is trained to convergence, primarily due to the prohibitive cost of full-scale LLM pretraining. In this paper, we introduce BLISS (BileveL Influence Scoring method for data Selection): a lightweight data selection method that operates entirely from scratch, without relying on any external pretrained oracle models, while explicitly accounting for the long-term impact of selected data. BLISS leverages a small proxy model as a surrogate for the LLM and employs a score model to estimate the long-term influence of training samples if the proxy model is trained to convergence. We formulate data selection as a bilevel optimization problem, where the upper-level objective optimizes the score model to assign importance weights to training samples, ensuring that minimizing the lower-level objective (i.e., training the proxy model over the weighted training loss until convergence) leads to best validation performance. Once optimized, the trained score model predicts influence scores for the dataset, enabling efficient selection of high-quality samples for LLM pretraining. We validate BLISS by pretraining 410M/1B/2.8B Pythia and LLaMA-0.5B models on selected subsets of the C4 dataset. Notably, under the 1B model setting, BLISS achieves $1.7\times$ speedup in reaching the same performance as the state-of-the-art method, demonstrating superior performance across multiple downstream tasks.

Read & Discuss → View Source →

04.

arXiv (CS.CV) 2026-06-16 DOI: arXiv:2602.07343

Seeing Roads Through Words: A Language-Guided Framework for RGB-T Driving Scene Segmentation

Authors:

Ruturaj Reddy↗Hrishav Bakul Barua↗Junn Yong Loo↗Thanh Thi Nguyen↗Ganesh Krishnasamy↗

Robust semantic segmentation of road scenes under adverse illumination, lighting, and shadow conditions remain a core challenge for autonomous driving applications. RGB-Thermal fusion is a standard approach, yet existing methods apply static fusion strategies uniformly across all conditions, allowing modality-specific noise to propagate throughout the network. Hence, we propose CLARITY that dynamically adapts its fusion strategy to the detected scene condition. Guided by vision-language model (VLM) priors, the network learns to modulate each modality's contribution based on the illumination state while leveraging object embeddings for segmentation, rather than applying a fixed fusion policy. We further introduce two mechanisms - one which preserves valid dark-object semantics that prior noise-suppression methods incorrectly discard, and a hierarchical decoder that enforces structural consistency across scales to sharpen boundaries on thin objects. Experiments on the MFNet dataset demonstrate that CLARITY establishes a new state-of-the-art (SOTA), achieving 62.3% mIoU and 77.5% mAcc.

Read & Discuss → View Source →

05.

arXiv (CS.AI) 2026-06-16 DOI: arXiv:2606.15507

Frame-Conditioned Moral Computation in LLaMA 3.1-8B-Instruct: A Mechanistic Interpretability Audit of Ethical Reasoning

Authors:

Ali Dasdan↗Manan Shah↗W. Russell Neuman↗Chad Coleman↗Kund Meghani↗Safinah Ali↗

arXiv:2606.15507v1 Announce Type: new Abstract: Behavioral audits of Large Language Models on moral prompts measure what the model says, not the internal computation producing it. We use Transluce, an AI-driven mechanistic-interpretability platform, to examine LLaMA 3.1-8B-Instruct on 54 moral prompts in four batteries: 17 dilemmas, policy, and meta-ethical questions (B1); 6 role-playing scenarios (B3); and a controlled trolley contrast varying the switching mechanism with people fixed (B4, 15 prompts) or identity attributes with mechanism fixed (B5, 16 prompts). Two complementary metric families, five cluster-level metrics and a six-metric neuron-level panel, converge on a Situational Anchor Effect: domain-specific representations dominate the top of the activation list across every battery. The model's ethics-labeled capacity stays essentially constant; its salience (rank, priority, top-of-list presence) is highly sensitive to the interpretive frame the prompt selects. The B4-vs-B5 contrast confirms the model attends to whichever surface feature varies: aggregate ethics metrics are indistinguishable, but the dominant non-ethics distractor mirrors the design. A multi-temperature audit identifies a candidate ethics neuron (L16/N3837) stable across temperatures; a cross-model behavioral proxy on two frontier models yields preliminary evidence of divergence in self-reported moral focus, consistent with an Alignment Wrapper in which RLHF re-orders surface text without removing underlying domain-first frames. We unify these as Frame-Conditioned Moral Computation: the prompt's surface vocabulary selects a feature manifold, and the moral conclusion is downstream of that selection. Behavioral alignment must be supplemented by Mechanistic Alignment: a research program asking whether ethics-related features can be shown causally privileged under controlled frame variation, not merely loud in the explanation.

Read & Discuss → View Source →

06.

arXiv (CS.CV) 2026-06-15 DOI: arXiv:2606.14194

Hybrid Classical-Quantum (HCQ) Alzheimer's Classification via Supervised $\beta$-VAE and Quantum Kernels

Authors:

Tia Tiwari↗Vamshi Krishna Kancharla↗Neelam Sinha↗

This paper presents a two-stage Hybrid Classical-Quantum (HCQ) pipeline for binary Alzheimer's disease (AD) classification from 3D T1-weighted structural MRI volumes, where the classical and quantum components are designed to complement each other rather than operate independently. A supervised 3D $\beta$-variational autoencoder (VAE) is trained end-to-end under voxel-wise reconstruction, KL-divergence, and focal classification losses that compress each 3D MRI volume (resized from 152 x 184 x 152 to 96 x 96 x 96) into a 64-dimensional latent code. Partial Least Squares (PLS) regression selects the six components in the latent code that best separate Alzheimer's Disease (AD) from cognitively normal (CN) subjects and rescales them into rotation angles, which are encoded onto a six-qubit register using the ZZ quantum feature map to give us the respective quantum states. The input to a precomputed-kernel Support Vector Machine (SVM) is an N x N Gram matrix (N = 308), created by calculating the overlap between every pair of quantum states. The novelty of this work lies in the fact that the quantum kernel operates directly on disease-aware features that are learned end-to-end by a supervised autoencoder, rather than on pre-extracted inputs. On 308 ADNI-1 subjects, consisting of 137 AD and 171 CN subjects, the baseline achieved 67.2% accuracy and 0.759 AUC, while the stability-enhanced variant reached 72.1% accuracy and 0.799 AUC with cross-fold variance halved. 3D Grad-CAM further helped validate our model's focus on brain regions linked to Alzheimer's. The HCQ pipeline could serve as a general-purpose framework for diagnostic classification across biomedical imaging domains that present similar challenges for classical approaches.

Read & Discuss → View Source →

07.

arXiv (quant-ph) 2026-06-16 DOI: arXiv:2604.25953

Entanglement as a Witness of Quantum Coherence: A Bipartite Monty-Hall Protocol

Authors:

Jorge Meza-Dom\'inguez↗

arXiv:2604.25953v3 Announce Type: replace Abstract: We present a bipartite protocol inspired by the Monty Hall puzzle that operationally distinguishes quantum coherence from classical ignorance. A principal qutrit is entangled with an ancillary qutrit via a controlled unitary, preparing $|\Psi\rangle = \frac{1}{\sqrt{3}}(|A,0\rangle + |B,1\rangle + |C,2\rangle)$. A rank-1 projective discard then eliminates one basis state, leaving a coherent superposition of the two remaining states. Finally, the ancilla and qutrit are measured, yielding joint probabilities that encode the interplay between superposition and measurement back-action. We show that the conditional probability $P(B|anc=0)$ takes the value $1/4$ in both quantum mechanics and the classical ignorant-host model, making it unsuitable as a witness. The true quantum-classical separation emerges in conditional joint probabilities that correlate ancilla outcomes with specific discard operations. We define witnesses $\mathcal{W}_{i,j} = P(anc=i, qutrit=j \mid discard k)$ where $j$ differs from the ancilla-implied state. Quantum mechanics predicts $\mathcal{W} = 1/4$, while any classical epistemic model with perfect initial correlations yields $\mathcal{W} = 0$. We provide the explicit $9 \times 9$ unitary matrix, a complete analysis of all measurement outcomes, and a detailed proof of the violation. The witness is fully immune to white noise and robust against moderate dephasing. The protocol requires only a single pair of entangled qutrits and sequential measurements – no spatial separation, no multiple copies, and no complex sets of incompatible observables. This makes it suitable for advanced undergraduate laboratories and provides a pedagogically accessible test of the ontic-epistemic distinction in quantum foundations.

Read & Discuss → View Source →

08.

arXiv (CS.LG) 2026-06-12 DOI: arXiv:2605.11165

COSMOS: Model-Agnostic Personalized Federated Learning with Clustered Server Models and Pseudo-Label-Only Communication

Authors:

Ben Rachmut↗Luise Ge↗William Yeoh↗Ning Zhang↗Yevgeniy Vorobeychik↗

arXiv:2605.11165v2 Announce Type: replace Abstract: Federated learning (FL) in heterogeneous environments remains challenging because client models often differ in both architecture and data distribution. While recent approaches attempt to address this challenge through client clustering and knowledge distillation, simultaneously handling architectural and statistical heterogeneity remains difficult. We introduce COSMOS, a model-agnostic framework that enables server-side personalization using only pseudo-label communication. Clients train local models and predict on the public data; the server clusters clients by prediction similarity, trains a cluster-specific model for each group using its own compute, and distills the resulting models back to clients. We provide the first theoretical analysis showing that distillation from the learned cluster models can yield exponential personalization risk contraction, going beyond the convergence-to-stationarity guarantees typically provided in model-agnostic FL. Experiments across benchmarks demonstrate that COSMOS consistently outperforms all model-agnostic FL baselines while remaining competitive with state-of-the-art personalized FL methods. More broadly, our results highlight personalized server-side learning with pseudo-labels as a promising paradigm for scalable and model-agnostic federated learning in highly heterogeneous environments.

Read & Discuss → View Source →

09.

arXiv (CS.CV) 2026-06-17 DOI: arXiv:2502.00241

Mordal: Automated Pretrained Model Selection for Vision Language Models

Authors:

Shiqi He↗Insu Jang↗Mosharaf Chowdhury↗

Incorporating multiple modalities into large language models (LLMs) is a powerful way to enhance their understanding of non-textual data, enabling them to perform multimodal tasks. Vision language models (VLMs) form the fastest growing category of multimodal models because of their many practical use cases, including in healthcare, robotics, and accessibility. Unfortunately, even though different VLMs in the literature demonstrate impressive visual capabilities in different benchmarks, they are handcrafted by human experts; there is no automated framework to create task-specific multimodal models. We introduce Mordal, an automated multimodal model search framework that efficiently finds the best VLM for a user-defined task without manual intervention. Mordal achieves this both by reducing the number of candidates to consider during the search process and by minimizing the time required to evaluate each remaining candidate. Our evaluation shows that Mordal can find the best VLM for a given problem using $8.9\times$–$11.6\times$ lower GPU hours than grid search. We have also discovered that Mordal achieves about 69\% higher weighted Kendall's $\tau$ on average than the state-of-the-art model selection method across diverse tasks.

Read & Discuss → View Source →

10.

arXiv (CS.AI) 2026-06-17 DOI: arXiv:2605.08827

Mental Health AI Safety Claims Must Preserve Temporal Evidence

Authors:

Srimonti Dutta↗Ratna Kandala↗

arXiv:2605.08827v2 Announce Type: replace Abstract: The safety of mental health AI is often judged at the wrong temporal scale. Current evaluations typically score isolated responses, endpoint outcomes, or aggregate dialogue quality, while clinically consequential failures may arise from the order and accumulation of interactions themselves, including delayed escalation, repeated reinforcement, dependency formation, failed repair, and gradual deterioration across turns. This paper argues that this mismatch is not merely a limitation of evaluation coverage but a source of invalid safety conclusions. We introduce Temporal Safety Non-Identifiability, a formal account of why safety properties that depend on sequence, timing, accumulation, or recovery cannot be certified by protocols that discard those features. From this formalization, we develop SCOPE (Safety Claims Over Preserved Evidence) as a general principle for aligning safety claims with the evidence an evaluation actually retains, and instantiate it as SCOPE-MH, a mental-health instantiation of this reporting standard. We operationalize SCOPE-MH through a proof-of-concept on the AnnoMI dataset of expert-annotated motivational interviewing conversations, which reveals mechanisms of failure that per-turn behavior scoring does not represent. We propose SCOPE-MH as a diagnostic complement to existing evaluation infrastructure and argue that evaluation preserving temporal evidence is necessary, not optional, for safety-critical mental health AI deployment.

Read & Discuss → View Source →

11.

arXiv (CS.AI) 2026-06-24 DOI: arXiv:2606.23741

A Survey on Federated Causal Discovery and Inference

Authors:

Xianjie Guo↗Yuwei Wang↗Guodu Xiang↗Xiaoli Tang↗Kui Yu↗Han Yu↗Qiang Yang↗

arXiv:2606.23741v1 Announce Type: cross Abstract: Causal reasoning, which encompasses the discovery of causal structures and the inference of causal effects, is fundamental to data-driven decision making. In practice, data for reliable causal analysis are often distributed across institutions and cannot be centralized due to privacy regulations or communication constraints. Federated learning (FL) addresses this by enabling collaborative analysis without raw data sharing, giving rise to the rapidly growing field of federated causal discovery (FCD) and inference (FCI). However, the interdisciplinary nature of this field and the absence of a comprehensive survey present barriers to entry for researchers. This paper bridges that gap by providing a systematic review through multi-dimensional taxonomies. Grounded in the three core design decisions underlying any FCD solution, namely how structures are learned, how data are partitioned, and what structural knowledge each party obtains, we organize FCD along three axes: methodological paradigm, federation topology, and structural scope. We further examine key practical dimensions, including temporal dynamics, data heterogeneity, missing data, and non-identical variable sets. For FCI, we categorize methods by target estimand (average versus individualized/conditional treatment effects) and by estimation strategy, from classical weighting methods to modern deep generative architectures. Unlike prior works that treat FCD and FCI separately, we formalize their connection as complementary stages of a unified federated causal reasoning pipeline, where FCD supplies the structural knowledge required for valid effect estimation in FCI. Finally, we highlight their shared concerns regarding privacy, communication efficiency, theoretical guarantees, and application domains, and conclude by identifying open challenges for future research.

Read & Discuss → View Source →

12.

arXiv (CS.AI) 2026-06-12 DOI: arXiv:2601.06572

Hellinger Multimodal Variational Autoencoders

Authors:

Huyen Vo↗Isabel Valera↗

arXiv:2601.06572v4 Announce Type: replace-cross Abstract: Multimodal variational autoencoders (VAEs) are widely used for weakly supervised generative learning with multiple modalities. Predominant methods aggregate unimodal inference distributions using either a product of experts (PoE), a mixture of experts (MoE), or their combinations to approximate the joint posterior. In this work, we revisit multimodal inference through the lens of probabilistic opinion pooling, an optimization-based approach. We start from Hölder pooling with $\alpha=0.5$, which corresponds to the unique symmetric member of the $\alpha-divergence$ family, and derive a moment-matching approximation, termed Hellinger. We then leverage such an approximation to propose HELVAE, a multimodal VAE that avoids sub-sampling, yielding an efficient yet effective model that: (i) learns more expressive latent representations as additional modalities are observed; and (ii) empirically achieves better trade-offs between generative coherence and quality, outperforming state-of-the-art multimodal VAE models.

Read & Discuss → View Source →

13.

arXiv (CS.LG) 2026-06-17 DOI: arXiv:2602.11557

The Implicit Bias of Steepest Descent with Mini-batch Stochastic Gradient

Authors:

Jichu Li↗Xuan Tang↗Difan Zou↗

arXiv:2602.11557v2 Announce Type: replace Abstract: A variety of widely used optimization methods like SignSGD and Muon can be interpreted as instances of steepest descent under different norm-induced geometries. In this work, we study the implicit bias of mini-batch stochastic steepest descent in multi-class classification, characterizing how batch size, momentum, and variance reduction shape the limiting max-margin behavior and convergence rates under general entry-wise and Schatten-$p$ norms. We show that, without momentum, worst-case convergence and successful classification can only be guaranteed with full-batch gradient. In contrast, momentum enables small-batch convergence to an approximate max-margin solution through a batch-momentum trade-off, though it slows convergence. This approach provides fully explicit, dimension-free rates that improve upon prior results. Moreover, we prove that variance reduction can recover the exact full-batch implicit bias for any batch size, albeit at a slower convergence rate. Finally, we further investigate the batch-size-one steepest descent without momentum, and reveal its convergence to a fundamentally different bias via a concrete data example, which reveals a key limitation of purely stochastic updates. Overall, our unified analysis clarifies when stochastic optimization aligns with full-batch behavior, and paves the way for perform deeper explorations of the training behavior of stochastic gradient steepest descent algorithms.

Read & Discuss → View Source →

14.

arXiv (CS.LG) 2026-06-18 DOI: arXiv:2606.19302

Optimal scenario design for climate emulation

Authors:

Christopher B. Womack↗Shahine Bouabid↗Andrei Sokolov↗Popat Salunke↗Glenn Flierl↗Sebastian D. Eastham↗Noelle E. Selin↗

arXiv:2606.19302v1 Announce Type: cross Abstract: As deep learning for physical systems continues to grow in popularity, efforts to improve generalizability have primarily focused on designing architectures that embed physical constraints. However, for machine-learning surrogate climate models (emulators), we show that the low structural diversity in existing scenarios commonly used to generate training data places a ceiling on predictive skill. Here, we examine whether training datasets themselves can be optimized to improve generalization. We introduce a method to create datasets that produce emulators capable of generalizing to new, structurally different scenarios absent from the training data. We use a differentiable Simple Climate Model (SCM) to calculate the sensitivity of emulator loss to perturbations in the training data, iteratively updating the training data to maximize emulator skill. For an SCM, training on one scenario optimized in this fashion outperforms an emulator trained on six standard ScenarioMIP pathways. We achieve this higher predictive skill despite training on a smaller dataset, finding that our emulator successfully isolates distinct physical behaviors of different climate forcing agents (e.g., greenhouse gases vs. aerosols) without single-forcing runs. We then demonstrate that scenarios optimized using an SCM, when used to drive an intermediate-complexity climate model, produce a training dataset that yields a more skillful emulator than training on ScenarioMIP outputs. Our results suggest that, in the compute-constrained environment of running full-scale climate models, generating a small number of dynamically rich scenarios provides greater marginal value for emulation and characterizing system responses than expanding the suite of traditional emissions pathways.

Read & Discuss → View Source →

15.

arXiv (CS.AI) 2026-06-19 DOI: arXiv:2603.28387

The Scaffold Effect: How Prompt Framing Drives Apparent Multimodal Gains in Clinical VLM Evaluation

Authors:

Doan Nam Long Vu↗Simone Balloccu↗

arXiv:2603.28387v2 Announce Type: replace Abstract: Trustworthy clinical AI requires that performance gains reflect genuine evidence integration rather than surface-level artifacts. We evaluate 12 open-weight vision-language models (VLMs) on binary classification across two clinical neuroimaging cohorts, \textsc{FOR2107} (affective disorders) and \textsc{OASIS-3} (cognitive decline). Both datasets come with structural MRI data that carries no reliable individual-level diagnostic signal. Under these conditions, smaller VLMs exhibit gains of up to 58\% F1 upon introduction of neuroimaging context, with distilled models becoming competitive with counterparts an order of magnitude larger. A contrastive confidence analysis reveals that merely mentioning MRI availability in the task prompt accounts for 70-80\% of this shift, independent of whether imaging data is present, a domain-specific instance of modality collapse we term the scaffold effect. Expert evaluation reveals fabrication of neuroimaging-grounded justifications across all conditions, and preference alignment, while eliminating MRI-referencing behavior, collapses both conditions toward random baseline. Our findings demonstrate that surface evaluations are inadequate indicators of multimodal reasoning, with direct implications for the deployment of VLMs in clinical settings.

Read & Discuss → View Source →

16.

bioRxiv (Bioinfo) 2026-06-22 DOI: HASH:40b8af690f0e77920f94f604e1d9cf95

Dynamic balance of sparse flux vectors for efficient simulation of culture dynamics and metabolic network reduction

Authors:

Tapia García↗Torrealba↗Luna↗Pérez-Correa↗J. R↗Saa↗P. A↗

Dynamic Flux Balance Analysis (DFBA) enables simulation of microbial culture dynamics under changing environmental conditions, but remains computationally expensive for tasks such as parameter calibration and fermentation optimization when applied using genome-scale metabolic models (GEMs). To address this challenge, we introduce Dynamic Flux Vector Balancing (DFVB), a reformulation of DFBA that solves an equivalent problem using a pre-computed, sparse basis of flux solutions that reduces the dimensionality of the internal optimization problem without information loss. Notably, DFVB provides a compact, interpretable representation of flux states that can readily identify dynamically inactive pathways and enable simulation-based automatic metabolic network reduction. We showed that DFVB produces the same culture dynamics as DFBA across multiple model scales and conditions, and identifies inactive reactions more accurately than Flux Variability Analysis (FVA) when compared to transcriptomic data profiles. Furthermore, computational performance analyses demonstrated that integrating DFVB with solver warm-start strategies and model reduction enhances computational efficiency relative to DFBA, yielding up to 3-fold reductions in simulation time for large-scale metabolic models. Finally, kinetic parameter estimation of culture dynamics with DFVB in two fermentation scenarios using a large-scale yeast GEM reached equal or higher prediction fidelity and narrower confidence intervals than DFBA, indicating improved parameter identifiability and robustness. Together, these results position DFVB as a scalable, robust, and biologically coherent framework for dynamic metabolic modeling, easing the integration of GEMs for culture dynamics simulation.

Read & Discuss → View Source →

17.

arXiv (CS.LG) 2026-06-16 DOI: arXiv:2606.15123

Data-Centric Benchmarking of Exploit Generation in LLMs: Understanding the Impact of Fine-Tuning

Authors:

Yiwei Chen↗Lichi Li↗Kai Cheung↗Vinny Parla↗Ganesh Sundaram↗

arXiv:2606.15123v1 Announce Type: cross Abstract: We study the task of CVE-conditioned exploit generation, where a model drafts proof-of-concept (PoC) exploits given software vulnerability context. We adopt a data-centric approach, constructing a high-quality dataset via multi-stage preprocessing and introducing a scalable evaluation framework with LLM-as-judge and fine-grained rubrics. Under this unified setup, we benchmark 17 large language models across 8 evaluation criteria, providing systematic insights into their zero-shot capabilities. We further show that a compact 8B open-weight model, when fine-tuned on curated data, achieves over 42.5% improvement in exploit quality and rivals some proprietary models when combined with simple test-time rejection strategies. Our results highlight the importance of data quality, structured supervision, and evaluation design for reliable exploit generation, suggesting that these factors can be as critical as model scale in adapting LLMs to cybersecurity tasks.

Read & Discuss → View Source →

18.

arXiv (CS.AI) 2026-06-18 DOI: arXiv:2606.18521

Sparsity Curse: Understanding RLVR Model Parameter Space from Model Merging

Authors:

Chenrui Wu↗Zexi Li↗Jiajun Bu↗Jiangchuan Liu↗Haishuai Wang↗

arXiv:2606.18521v1 Announce Type: cross Abstract: Reinforcement Learning with Verifiable Reward (RLVR) has emerged as a powerful post-training paradigm that surpasses Supervised Fine-Tuning (SFT) in eliciting reasoning intelligence and resisting catastrophic forgetting. Recent studies further reveal that RLVR induces highly sparse and off-principal parameter updates compared to SFT. This naturally raises the question: does such sparsity make RLVR models more amenable to model merging? If so, model merging would offer a scalable, training-free path to aggregate diverse reasoning capabilities from independently trained RLVR models. Surprisingly, we find the opposite, uncovering a sparsity curse: the sparse RLVR updates are spread farther apart in parameter space, forming near-orthogonal shortcuts that make aggregation inherently fragile. This is likely rooted in the stochasticity of RL optimization and the diversity of emergent reasoning patterns. Unlike SFT models that converge to shared, flat basins and merge naturally, RLVR models suffer severe degradation under standard merging methods. Through systematic empirical analysis of the update geometry, we characterize the mechanisms behind this failure and propose Sensitivity-aware Resolving Merging (SAR-Merging), a merging recipe tailored for the unique structure of RLVR parameter spaces. SAR-Merging resolves conflicts in overlapping update regions via Fisher Information-based sensitivity arbitration, followed by magnitude-aware sparsification and rescaling to preserve fragile reasoning pathways. Experiments on mathematical and coding benchmarks demonstrate that SAR-Merging substantially outperforms existing merging methods on RLVR models, enabling both single-task enhancement and multi-capability fusion.

Read & Discuss → View Source →

19.

arXiv (CS.CL) 2026-06-16 DOI: arXiv:2505.14418

Hidden Ghost Hand: Unveiling Backdoor Vulnerabilities in MLLM-Powered Mobile GUI Agents

Authors:

Pengzhou Cheng↗Haowen Hu↗Zheng Wu↗Zongru Wu↗Tianjie Ju↗Zhuosheng Zhang↗Gongshen Liu↗

Graphical user interface (GUI) agents powered by multimodal large language models (MLLMs) have shown greater promise for human-interaction. However, due to the high fine-tuning cost, users often rely on open-source GUI agents or APIs offered by AI providers, which introduces a critical but underexplored supply chain threat: backdoor attacks. In this work, we first unveil that MLLM-powered GUI agents naturally expose multiple interaction-level triggers, such as historical steps, environment states, and task progress. Based on this observation, we introduce AgentGhost, an effective and stealthy framework for red-teaming backdoor attacks. Specifically, we first construct composite triggers by combining goal and interaction levels, allowing GUI agents to unintentionally activate backdoors while ensuring task utility. Then, we formulate backdoor injection as a Min-Max optimization problem that uses supervised contrastive learning to maximize the feature difference across sample classes at the representation space, improving flexibility of the backdoor. Meanwhile, it adopts supervised fine-tuning to minimize the discrepancy between backdoor and clean behavior generation, enhancing effectiveness and utility. Extensive evaluations of various agent models in two established mobile benchmarks show that AgentGhost is effective and generic, with attack accuracy that reaches 99.7\% on three attack objectives, and shows stealthiness with only 1\% utility degradation. Furthermore, we tailor a defense method against AgentGhost that reduces the attack accuracy to 22.1\%. Our code is available at \texttt{anonymous}.

Read & Discuss → View Source →

20.

arXiv (CS.AI) 2026-06-16 DOI: arXiv:2606.15485

The Perils of Agency: How Developers Perceive, Prioritize, and Address Risks in Agentic AI Products

Authors:

Hao-Ping Lee↗Jessica He↗David Piorkowski↗Thomas Serban von Davier↗Jodi Forlizzi↗Sauvik Das↗

arXiv:2606.15485v1 Announce Type: cross Abstract: Agentic AI systems act autonomously, use tools, adapt to context, and operate in complex real-world environments. However, these same characteristics can create or exacerbate product risks. We studied how industry developers (n=35) perceive, prioritize, and address the risks in their agentic AI products. We found that developers' perceptions of risk were closely tied to the qualities that made the product agentic, such as autonomy, tool use, and usage in a real-world context. Developers prioritized product and business risks before considering downstream societal risks like job displacement and end-user privacy. This prioritization also impacted developers' ability and motivation to mitigate agentic risks. Finally, developers lacked mature controls for containing agentic risks, often relying on constraining the same characteristics that make agents useful: e.g., autonomy and goal complexity. These findings reveal a capability vs. risk control tension in agentic AI development: developers need to address risks that emerge from agentic capabilities, yet they currently have limited support for doing so without constraining agentic functionality.

Read & Discuss → View Source →

21.

arXiv (CS.LG) 2026-06-16 DOI: arXiv:2508.11522

Finite-Width Neural Tangent Kernels from Feynman Diagrams

Authors:

Max Guillen↗Philipp Misof↗Jan E. Gerken↗

arXiv:2508.11522v4 Announce Type: replace Abstract: Neural tangent kernels (NTKs) are a powerful tool for analyzing deep, non-linear neural networks. In the infinite-width limit, NTKs can easily be computed for most common architectures, yielding full analytic control over the training dynamics. However, at infinite width, important properties of training such as NTK evolution or feature learning are absent. Nevertheless, finite width effects can be included by computing corrections to the Gaussian statistics at infinite width. We introduce Feynman diagrams for computing finite-width corrections to NTK statistics. These dramatically simplify the necessary algebraic manipulations and enable the computation of layer-wise recursion relations for arbitrary statistics involving preactivations, NTKs and certain higher-derivative tensors (dNTK and ddNTK) required to predict the training dynamics at leading order. We demonstrate the feasibility of our framework by extending stability results for deep networks from preactivations to NTKs and proving the absence of finite-width corrections for scale-invariant nonlinearities such as ReLU on the diagonal of the Gram matrix of the NTK. We numerically implement the complete set of equations necessary to compute the first-order corrections for arbitrary inputs and demonstrate that the results follow the statistics of sampled neural networks for widths $n\gtrsim 20$.

Read & Discuss → View Source →

22.

arXiv (CS.AI) 2026-06-12 DOI: arXiv:2606.12821

GeoNatureAgent Benchmark: Benchmarking LLM Agents for Environmental Geospatial Analysis Across Frontier and Open-Weight Foundation Models

Authors:

Gabriel Diaz-Ireland↗Diego Prieto-Herr\'aez↗Mario Garc\'ia Peces↗Javier Vel\'azquez↗Devika Jain↗

arXiv:2606.12821v1 Announce Type: new Abstract: Environmental scientists spend disproportionate effort on data wrangling rather than analysis, and AI agents that automate geospatial workflows remain unvalidated: no benchmark evaluates agents operating through structured tool calling against real APIs. We introduce the GeoNatureAgent Benchmark, the first benchmark for environmental analysis agents that operate via structured tool calls to a production-style geospatial API. It comprises 93 tasks across 18 categories, covering municipality analysis, multi-turn conversation, spatial reasoning, cross-indicator synthesis, error handling and recovery, ranking, comparison, multilingual understanding, habitat analysis, and task rejection. Tasks are evaluated against an open, self-hostable API serving three environmental indicators across Spain and Portugal via sixteen tools. We evaluate seven LLMs (Claude Sonnet 4, DeepSeek V3.2, GLM-5, Gemini 2.5 Pro, Qwen3-235B, GPT-OSS-120B, Llama 4 Scout) under three temperature-1.0 seeds, reporting capability and per-case cost as orthogonal axes. We find: (1) Claude Sonnet 4 leads at 60.8% +/- 0.8%, followed by DeepSeek V3.2 at 56.3% +/- 3.1%, with no other model above 51%; (2) the cost-accuracy Pareto frontier is occupied mostly by open-weight models, with DeepSeek V3.2 offering 93% of Claude's capability at 11x lower cost ($0.011/case); (3) comparison tasks remain universally unsolved (0% on close-value comparisons), exposing systematic reasoning limits; and (4) structured tool calling against a real API is more discriminative than general-purpose GIS benchmarks, with accuracies 25-35 points lower. We further show extensibility by integrating BigEarthNet V2 land cover for Portugal alongside Spanish CO2 and erosion indicators. The benchmark, harness, and self-hostable API are publicly available.

Read & Discuss → View Source →

23.

arXiv (CS.CL) 2026-06-11 DOI: arXiv:2509.23982

Toward Preference-aligned Large Language Models via Residual-based Model Steering

Authors:

Lucio La Cava↗Andrea Tagarelli↗

Preference alignment is a critical step in making Large Language Models (LLMs) useful and aligned with (human) preferences. Existing approaches such as Reinforcement Learning from Human Feedback or Direct Preference Optimization typically require curated data and expensive optimization over billions of parameters, and eventually lead to persistent task-specific models. In this work, we introduce Preference alignment of Large Language Models via Residual Steering (PaLRS), a training-free method that exploits preference signals encoded in the residual streams of LLMs. From as few as one hundred preference pairs, PaLRS extracts lightweight, plug-and-play steering vectors that can be applied at inference time to push models toward preferred behaviors. We evaluate PaLRS on various small-to-medium-scale open-source LLMs, showing that PaLRS-aligned models achieve consistent gains on mathematical reasoning and code generation benchmarks while preserving baseline general-purpose performance. Moreover, when compared to models aligned with DPO and SimPO, they perform better with great time-savings. Our findings highlight that PaLRS offers an effective, much more efficient and flexible alternative to standard preference optimization pipelines, offering a training-free, plug-and-play mechanism for alignment with minimal data.

Read & Discuss → View Source →

24.

medRxiv (Medicine) 2026-06-23 DOI: HASH:4f60e7f5d02b47d01e05d3e3179c503c

Unscreenable: The Burden, Structure, and Analytic Consequences of "Unable to Assess" Delirium Documentation in the Intensive Care Unit

Authors:

Gorenshtein↗Adiniaev↗Omar↗Barash↗Klang↗Daniel↗

Objective: To quantify the burden, structure, and downstream analytic consequences of "Unable to Assess" (UTA) delirium documentation in the intensive care unit (ICU). Design: Retrospective cross-sectional and repeated-measures study. Setting: A single US academic medical center (Medical Information Mart for Intensive Care IV [MIMIC-IV], 2008-2019). Patients: 72,944 adult ICU stays with at least 1 delirium screen. Interventions: None. Measurements and Main Results: Among 610,632 screens, 130,455 (21.4%; 95% CI, 21.0%-21.8%) were recorded as UTA, exceeding the 119,052 (19.5%) scored positive. The UTA fraction rose from 2.0% at a Richmond Agitation-Sedation Scale (RASS) score of 0 to 97.8% at RASS -4; 22.0% of UTA screens occurred in arousable patients, where UTA was associated with mechanical ventilation (odds ratio [OR], 3.43; 95% CI, 3.17-3.71) and non-English primary language (OR, 3.74; 95% CI, 3.43-4.08). Building the delirium label three ways from the same patients shifted prevalence modestly (32.1% to 30.8%) and prediction (area under the curve, 0.737 to 0.719) but most affected the delirium-mortality association: in a baseline-adjusted model the OR was 4.12 (95% CI, 3.88-4.36) under complete-case handling and fell to 2.16 (95% CI, 2.06-2.27) when UTA was recoded as negative. UTA was recoverable from the observed clinical state (area under the curve, 0.95). Conclusions: In this ICU cohort, Unable to Assess was the most common recorded delirium result other than Negative, exceeding positive screens; recoding it as negative roughly halved the apparent delirium-mortality association by relabeling deeply sedated, high-mortality patients. Delirium datasets should preserve and report UTA, whose concentration among arousable non-English-speaking patients is a measurable equity target.

Read & Discuss → View Source →

25.

arXiv (CS.CV) 2026-06-18 DOI: arXiv:2606.19249

Transformer Geometry Observatory TGO-I: Spectral Geometry Observatory

Authors:

Kaustubh Kapil↗Kishor P. Upla↗

Despite the widespread adoption of Vision Transformers (ViTs) and their success across numerous computer vision applications, the fundamental understanding of their dimensional and representational geometry remains relatively underexplored. To address this gap, we introduce Transformer Geometry Observatory (TGO), a systematic framework of experiments and analysis pipelines designed to investigate the representational geometry and dynamics of Vision Transformers. TGO-I, the first installment of the framework, focuses on the spectral geometry of ViT representations. Using a ViT-Small/16 model trained on ImageNet-100, we analyze Effective Rank, Stable Rank, Participation Ratio, Spectral Entropy, Spectral Flatness, Spectral Anisotropy, covariance structure, eigenspectra, and singular value spectra throughout training. Our results reveal a consistent increase in dimensional utilization, accompanied by decreasing anisotropy, increasing spectral entropy, increasing participation ratio, and progressively flatter eigenspectra. Contrary to the common intuition that training should concentrate information into a small number of dominant directions, we observe a progressive redistribution of variance across representational dimensions. This phenomenon is particularly pronounced in the final CLS token representation, which exhibits the highest effective dimensionality and lowest anisotropy within the network.

Read & Discuss → View Source →