arXiv Research Digest

Recent studies have made notable progress in video representation learning by transferring image-pretrained models to video tasks . Fine-tuning heavy modules may compromise inter-video semantic separability . While reducing the tunable parameters hinders their intra-video temporal consistency, which is required for stable representations of the same object within a video .

Vision-language models are trained on massive paired text-image corpora but lack embodied experience . We report two experiments comparing descriptions generated by 18 VLMs to those of over 2000 human observers across 15 high-level scene understanding tasks .

OVI-MAP decouples instance reconstruction from semantic inference . System operates in real time and outperforms state-of-the-art open-vocabulary mapping baselines .

Clinicians often need to retrieve patient-specific information from electronic health records . We present a locally deployable Clinical Contextual Question Answering framework that answers clinical questions directly from EHRs without external data transfer .

Extended-thinking models expose a second text-generation channel alongside the user-visible answer . In 55.4% of these cases the model's thinking tokens contain hint-related keywords that the visible answer omits entirely .

Reconstructing 3D geometry and appearance from a sparse set of fixed cameras is a foundational task with broad applications . We show that this bound can be broken by exploiting opportunistic object motion . We address these by formulating a joint pose and shape optimization using 2D Gaussian splatting .

Large language model (LLM)-based coding agents achieve impressive results on controlled benchmarks yet routinely produce pull requests that real maintainers reject . Learning to Commit is a framework that closes this gap through Online Repository Memory . Given a repository with a strict chronological split, the agent performs contrastive reflection on earlier commits .

Weight tying, i.e. sharing parameters between input and output embedding matrices, is common practice in language model design . Yet its impact on the learned embedding space remains poorly understood . In this paper, we show tied embedding matrix aligns more closely with output matrices than with input embeddings of comparable untied models . This unembedding bias arises because output gradients dominate early in training .

High-fidelity 3D reconstruction of vehicle exteriors improves buyer confidence in online automotive marketplaces . Generating these models in cluttered dealership drive-throughs presents severe technical challenges . We propose an end-to-end pipeline utilizing a two-pillar camera rig .

This paper investigates relationship between utterance sentiment and language choice in English-Tamil code-switched text . We apply a fine-tuned XLM-RoBERTa model for token-level language identification on 35,650 romanized YouTube comments from the DravidianCodeMix dataset .

MA-Bench is a benchmark comprising 1,000 videos and a three-tier evaluation architecture that progressively examines micro-action perception, relational comprehension, and interpretive reasoning . MA-bench contains 12,000 structured question-answer pairs, enabling systematic assessment of both recognition accuracy and action interpretation . Results of 23 representative MLLMs reveal that there are significant challenges in capturing motion granularity .

Reconstructing accurate 3D models of large-scale real-world scenes from unstructured, in-the-wild imagery remains a challenge in computer vision . We obtain reference models from dense, geospatially accurate pseudo-synthetic renderings derived from Google Earth Studio .

The field of co-speech gesture generation has seen significant advances but producing holistic, semantically grounded gestures remains a challenge . Existing approaches rely on external semantic retrieval methods, which limit their generalisation capability due to dependency on predefined linguistic rules . We introduce a Contrastive Flow Matching-based model that uses mismatched audio-text conditions as negatives, training the velocity field to follow the correct motion trajectory while repelling semantically incongruent trajectories .

AutoWeather4D is a feed-forward 3D-aware weather editing framework designed to explicitly decouple geometry and illumination . The Geometry Pass leverages explicit structural foundations to enable surface-anchored physical interactions . The Light Pass analytically resolves light transport, accumulating the contributions of local illuminants into the global illumination to enable dynamic 3D local relighting .

Existing datasets do not capture when adverse events (AEs) are officially recognized by regulatory authorities . This study addresses this gap by developing a time-indexed reference dataset for the European Union (EU) It incorporates the timing of AE inclusion in product labels along with regulatory metadata .

Named entity recognition (NER) enables the automatic extraction of medical concepts, but benchmarks for Portuguese remain scarce . We aimed to evaluate BERT-based models and large language models for clinical NER in Portuguese . The mmBERT-base model achieved the best performance (micro F1 = 0.76), outperforming all other models .

Computed tomography (CT) provides rich 3D anatomical details but is often constrained by high radiation exposure, substantial costs, and limited availability . While standard chest X-rays are cost-effective and widely accessible, they only provide 2D projections with limited pathological information . We propose AXON, a multi-stage diffusion-based framework that reconstructs high-fidelity 3D CT volumes .

Quantum language models have shown competitive performance on sequential tasks . We introduce the first mechanistic interpretability study of quantum language models . We find single-qubit models are exactly classically simulable and converge to the same geometric strategy as matched classical baselines .

Large vision-language models tend to hallucinate when visual inputs are corrupted at test time . We propose ClipTTT, a method to adapt LVLMs under degraded conditions on the fly with a single test sample . We leverage image-text alignment strength of a pre-trained CLIP model as a stable guidance signal .

High-quality 4D reconstruction enables photorealistic and immersive rendering of the dynamic real world . Unlike static scenes that can be fully captured with a single camera, dynamic scenes typically require dense arrays of tens or hundreds of synchronized cameras .

ClimateCheck 2026 is the second iteration of a shared task addressing this challenge . The competition ran from January to February 2026 on the CodaBench platform .

Cervical dystonia (CD) is the most common form of dystia . Current assessment relies on subjective clinical rating scales, such as the Toronto Western Spasmodic Torticollis Rating Scale . We developed an assessment tool that combines a pretrained head-pose estimation algorithm for rotational symptoms with a deep learning model trained exclusively on 16,000 synthetic avatar images .

Incivility in the German Bundestag signals political polarization and institutional conflict . An insult towards individuals is the most frequent cause of CtO . Male members and those belonging to opposition parties receive more calls to order than female and coalition-party counterparts .

Existing approaches favor node mapping that maximizes $F_1$ score over node relations and attributes, regardless whether the similarity is intentional or accidental . We propose a node-matching algorithm that allows comparison of multiple Uniform Meaning Representations (UMR) of one sentence .

Reentrancy remains one of the most critical classes of vulnerabilities in smart contracts . Traditional tools and ML models achieve up to 0.87 F1, while the best LLMs reach 0.96 in a zero-shot setting . Most tools fail on multiple scenarios, the top performer achieving an F1 of 0.76 .

This paper presents an autoethnographic analysis of cryptoasset auditing challenges . It builds on top of prior research on a comprehensive framework addressing existence, ownership, valuation, and internal control verification .

Null-Space Compression (NSC) Merging is a label-free, output-agnostic method that sets merge weights from adapter geometry . NSC achieves state-of-the-art performance across twenty heterogeneous vision tasks with balanced gains where prior methods overfit subsets of tasks .

As multimodal models like CLIP become integral to downstream systems, the need to remove sensitive information is critical . We introduce SALMUBench (Sensitive Association-Level Multimodal Unlearning), a benchmark built upon a synthetic dataset of 60K persona-attribute associations .

We propose a semi-structured discrete-time multi-state model to analyse mortgage delinquency transitions . This model combines an easy-to-understand structured additive predictor with a flexible neural network component . We demonstrate the method using the Freddie Mac Single-Family Loan-Level Dataset, employing an out-of-time test design .

Merging multiple Low-Rank Adaptation (LoRA) modules is promising for constructing general-purpose systems, but challenging because LoRA update directions span different subspaces and contribute unevenly . We propose TARA-Merging (Task-Rank Anisotropy Alignment), which aligns merging weights using a preference-weighted cross-entropy pseudo-loss . This ensures broad subspace coverage and mitigates anisotry via direction-wise reweighting

Transfer-based anti-money laundering (AML) systems monitor token flows through transaction-graph abstractions . This assumption, the bedrock of industrial forensics, fundamentally collapses in composable smart-contract ecosystems . We formalize two structural mechanisms that undermine completeness of transfer-layer attribution .

Optimal dispatch of energy storage systems in distribution networks involves jointly improving operating economy and voltage security under time-varying conditions . We develop a topology-aware Reinforcement Learning architecture based on Twin Delayed Deep Deterministic Policy Gradient (TD3)

Contrastive learning produces coherent semantic feature embeddings by encouraging positive samples to cluster closely while separating negative samples . We extend conformal prediction to this setting by introducing minimum-volume covering sets equipped with learnable generalized multi-norm constraints . Negative exclusion is maximized through learned set geometry optimized on a held-out training split .

Hypertrophic cardiomyopathy (HCM) requires accurate risk stratification to inform decisions regarding ICD therapy and follow-up management . This study develops a robust, explainable machine learning (ML) risk score leveraging routinely collected echocardiographic, clinical, and medication data, typically contained within Electronic Health Records (EHRs)

Generalized Hopfield models with higher-order or exponential interaction terms are known to have substantially larger storage capacities than the classical quadratic model . On the other hand, associative memories for sparse patterns, such as the Willshaw and Amari models, already outperform the classical Hopfield model in the sparse regime . In this paper we combine these two mechanisms .

In extreme-sparsity regimes, foreground signals are overwhelmingly dominated by background observations . Background-driven gradients destabilize the feature backbone during sequential domain shifts . This exposes a structural limitation of continual learning approaches relying solely on output-level distillation . We propose a dual-stage invariant continual learning framework via joint distillation, enforcing structural and semantic consistency on both backbone representations and detection predictions .

Geometric Evolution Graph Convolutional Network (GEGCN) is a novel framework that enhances graph representation learning by modeling geometric evolution on graphs . GEGCN employs a Long Short-Term Memory to model the structural sequence generated by discrete Ricci flow .

We propose Process-Aware Policy Optimization (PAPO) to address two limitations of existing reward designs . Outcome reward models evaluate only final-answer correctness, treating all correct responses identically regardless of reasoning quality . Process reward models offer richer supervision but directly using PRM scores causes reward hacking, where models exploit verbosity to inflate scores while accuracy collapses .

CR-Eyes is a computationally rational model that simulates visual sampling and gameplay behavior in Atari games . It is a step toward scalable, theory-grounded user models that support design and evaluation of interactive systems .

Can an expensive AI model effectively direct a cheap one to solve software engineering tasks? We study this question by introducing ManagerWorker, a two-agent pipeline where an expensive "manager" model analyzes issues, dispatches exploration tasks, and reviews implementations . Our findings reveal both the promise and the limits of multi-agent direction .

Large language models' performance on spatial reasoning benchmarks reflects structured internal spatial representations or reliance on linguistic heuristics . We examine internal representations using linear probing, sparse autoencoder based feature analysis, and causal interventions . We find that task relevant spatial information is encoded in intermediate layers and can causally influence behavior .

GUIDE (GUI Unbiasing via Instructional-Video Driven Expertise) is a plug-and-play framework that resolves GUI agent domain bias by autonomously acquiring domain-specific expertise from web tutorial videos through a retrieval-augmented automated annotation pipeline .

While Late Interaction models exhibit strong retrieval performance, many of their underlying dynamics remain understudied, potentially hiding performance bottlenecks . In this work, we focus on two topics: a length bias that arises when using multi-vector scoring and the similarity distribution beyond the best scores pooled by the MaxSim operator .

Autoregressive vision-language models have long dominated multimodal understanding, reasoning, and graphical user interface (GUI) grounding . We evaluate whether discrete DVLMs can serve as a viable alternative to AR models for GUI grounding . The authors propose a hybrid masking schedule that combines linear and deterministic masking .

Cardiac MRI late gadolinium enhancement (LGE) enables non-invasive identification of left atrial (LA) scar, whose spatial distribution is strongly associated with atrial fibrillation (AF) severity and recurrence . However, automatic LA scar segmentation remains challenging due to low contrast, annotation variability, and lack of anatomical constraints, often leading to non-reliable predictions .

The gaze-and-pinch framework offers a high-fidelity interaction modality for spatial computing in virtual reality . It remains vulnerable to coordination errors--timing misalignments between gaze fixation and pinch gestures . We investigate two heuristics--STICKY selection (temporal buffer) and MAGNETIC selection (spatial field)

Large Language Models (LLMs) have shown impressive capabilities across software engineering tasks, including question answering . We introduce StackRepoQA, the first multi-project, repository-level question answering dataset constructed from 1,318 real developer questions and accepted answers across 134 open-source Java projects .

Buffer zones are essential in production systems to decouple sequential processes . Automating these zones requires solving the Buffer Storage, Retrieval, and Reshuffling Problem (BSRRP)

A new form of human trafficking has emerged across Chinese borders, where individuals are lured to Southeast Asia with fraudulent job offers and then coerced into operating online scams . Despite its massive economic and human toll, this scam-driven trafficking remains underexplored in academic research .

CADSmith is a multi-agent pipeline that generates CadQuery code from natural language . It then undergoes an iterative refinement process through two nested correction loops . The outer loop combines exact measurements from the OpenCASCADE kernel with holistic visual assessment from an independent vision-language model .

AIRA$_2$ addresses three structural performance bottlenecks in AI research agents: synchronous single-GPU execution constrains sample throughput . A Hidden Consistent Evaluation protocol delivers a reliable evaluation signal . ReAct agents that dynamically scope their actions and debug interactively .

Auditory Attention Decoding (AAD) aims to identify the attended speech stream in a multiple speaker scenario from neural recordings . Entrainment-based AAD approaches assume access to clean speech sources and electroencephalography (EEG) signals . In this study, we propose CA-TCN, a Causal-Anticausal Temporal Temporal Convolutional Network that directly classifies the attended speaker .

Multimodal Large Language Models (MLLMs) achieve strong multimodal reasoning performance . As outputs grow longer, models drift away from image evidence and fall back on textual priors, resulting in ungrounded reasoning and hallucinations . We propose Visual Re-Examination (VRE), a self-evolving training framework that enables MLLMs to autonomously perform visual introspection during reasoning without additional visual inputs . VRE promotes iterative self-improvement by leveraging the

Legal reasoning requires understanding of the context in which rules operate . CALRK-Bench provides a new stress test for evaluating context-aware legal reasoning . The benchmark is based on the legal system in Korean .

Supervised fine-tuning on visual instruction data often improves perceptual capabilities in vision-language models . We propose Input-Adaptive Depth Aggregation (IADA) to make cross-depth retrieval input-adaptive, modality-aware, and efficiently parameterized through a low-rank bottleneck .

Most existing approaches to AI in pharmacy collapse three epistemologically distinct operations into a single technical layer: document preservation, semantic interpretation, and contextual presentation . This conflation is a root cause of recurring fragilities including loss of provenance, interpretive opacity, alert fatigue, and erosion of accountability . This paper proposes the PATOS--Lector--PRISMA infrastructure as a normative information architecture for responsible pharmaceutical knowledge management .

Findsylls unifies classical syllable detectors and end-to-end syllabifiers under a common interface for syllable segmentation, embedding extraction, and multi-granular evaluation . The toolkit implements and standardizes widely used methods (e.g., Sylber, VG-HuBERT) and allows their components to be recombined .

GeoGuide is a novel framework that leverages pretrained 3D models to integrate geometry-semantic consistency for open-vocabulary 3D segmentation . Extensive experiments on ScanNet v2, Matterport3D, nuScenes demonstrate the superior performance of GeoGuide .

Language endangerment poses a major challenge to linguistic diversity worldwide . Automatic speech recognition (ASR) has shown increasing potential to assist in the transcription of endangered language data . This study focuses on Ikema, a severely endangered Ryukyuan language spoken in Okinawa, Japan .

In a previous work, the Sequence Encoder for online dynamical system identification (Elaarabi et al., 2025a) and its combination with PINN (PINN-SE) were introduced and tested on both synthetic and real data case scenarios . The results show that combining multiple encoders with the previously proposed method is feasible . Training the model on synthetic data generated based on experimental data can help the model to generalize well for real experimental data .

Large Language Model (LLM) technology suggests a meta-semantic picture of how words and complex expressions come to have the meaning that they do . It has previously been argued that LLMs adopt a form of holism about meaning . Recent work in mechanistic interpretability presents a challenge to these arguments .

Recent research often relies on prompt engineering with large language models to simulate novice student behaviour, but it is difficult to keep the AI-simulated student at a stable novice knowledge level . Many LLMs are trained to be broadly capable, so even when prompted to "act like a novice," the LLMs can still produce expert-level explanations . We propose a knowledge-level simulation approach based on machine unlearning .

For undirected weighted graphs, we design a randomized algorithm that achieves a $(k+1)$-approximation, for any \emph{real} number $k \ge 1$ The round complexity of algorithm is $O/O, where $n$ denotes the number of nodes and $D$ is the unweighted diameter of the graph . This result yields a smooth trade-off between approximation and round complexity .

Federated Learning (FL) is a paradigm for training machine learning (ML) models in collaborative settings while preserving participants' privacy by keeping raw data local . A key requirement for the use of FL in production is reliability .

FluxEDA introduces a managed gateway-based execution interface with structured request and response handling . It also maintains persistent backend instances . These features allow upper-layer agents and programmable clients to interact with heterogeneous EDA tools .

Large language models have shown promise in generating RTL code from natural-language descriptions . We propose IncreRTL, a LLM-driven framework for incremental RTL generation under requirement evolution . By constructing requirement-code traceability links to locate and regenerate affected code segments .

Existing multi-agent frameworks allow each agent to simultaneously plan, execute, and evaluate its own actions -- a structural deficiency we term the "Logic Monopoly" Empirical evidence quantifies the resulting "Reliability Gap": 84.30% average attack success rates across ten deployment scenarios . The remedy is not better alignment of individual models but a social contract for agents .

We present the first polylogarithmic-round algorithm for sampling a random spanning tree in the (Broadcast) Congested Clique model . This is an exponential improvement over the previous best algorithm of Pemmaraju, Roy, and Sobel (PODC 2025)

Neuromorphic VLSI systems take inspiration from biology to enable efficient emulation of large-scale spiking neural networks . To establish large neuromorphic systems, a sophisticated routing infrastructure is needed .

Practical deployments must contend with unpredictable network latencies when information dispersal is integrated into consensus protocols, a prominent and latency-sensitive use case . Rafture is the first solution to incorporate post-dissemination pruning, allowing systems to adapt storage costs after dissemination completes .

Interleaved ADCs are critical for applications requiring multi-gigasample per second (GS/s) rates . Performance is often limited by offset, gain, and timing skew mismatches across sub-ADCs . We propose exact but compact expressions that describe the impact of each of those non-idealities on the output spectrum .

Storage scalability is paramount in the era of big data blockchain . A storage-scalable blockchain can effectively scale out state storage to an arbitrary number of nodes . Supermassive Blockchain achieves better storage scalability compared to prior approaches .

This paper maps the conceptual lineage from Gray's requestor/server model to modern blockchain architectures . We examine consensus mechanisms, cryptographic foundations, rollup-based Layer-2 protocols, and cross-chain interoperability through this historical lens . We outline future directions toward Web4: an intelligent, decentralized internet integrating blockchain, artificial intelligence, and the Internet of Things .

The TRAP protocol solves rational agreement by combining accountable consensus with a one-shot BFTCR finalization phase . SNARE (Scalable Nash Agreement via Reward and Exclusion) is the adaptation of TRAP to $n=5f{+}1$ and prove $ε$-$(k,t)$-robustness .

Parallel applications with irregular and time-varying workloads often suffer from load imbalance . Dynamic load balancing techniques address this challenge by redistributing work during execution .

Fully Homomorphic Encryption (FHE) is rapidly emerging as a promising foundation for privacy-preserving cloud services . FHE incurs much higher computational overhead, making it more susceptible to transient hardware faults . Data corruptions are likely to remain silent because FHE service has no access to the underlying plaintext .

KV-cache reuse offers a promising solution by storing previously computed KV states for shared input prefixes . Yet, the effectiveness of cache reuse is limited by low cache hit rates due to naive eviction policies and high CPU-GPU data transfer overhead . We propose PCR, a system designed to maximize KV cache reuse efficiency through intelligent prefetching and pipelined data movement .

Simulation remains a cornerstone of computer architecture research . Full end-to-end application execution is prohibitively time-consuming . The industry-standard solution, SimPoint, mitigates this cost by selecting a small number of representative code regions .

Computer system simulation studies rely on executing a limited number of short application regions . To preserve representativeness, existing methods employ either random sampling or phase-based characterization to identify representative regions . We show that the ranked set sampling (RSS) technique - well established in the statistical literature - maps naturally to architectural simulation .

Most of our interactions with computers are interactive and often urgent . Applications from simulations to data analysis and machine learning require more parallel computational capability and more interactivity . This chapter overviews the progress made so far along with some vectors of what the path forward will bring .

Cycle-accurate simulators are widely used to study systolic accelerators, yet their accuracy and usability are limited by weak validation against real hardware and poor integration with modern ML compilation stacks . This paper presents SCALE-Sim TPU, a validated and extended version of SCale-Sim v3 for TPU-style accelerators .

In the 1980s, high-performance computing (HPC) became another tool for research in the open (non-defense) science and engineering research communities . HPC came with a high price tag; the first Cray-2 machines, released in 1985, cost between $12 million and $17 million . In the 1990s, with demand for HPC increasing due to vast datasets, more complex modeling, and growing computational needs of scientific applications, researchers began experimenting with building H

ExaCB is a framework for continuous benchmarking developed in the context of the JUPITER exascale system . Integrating benchmarking into CI workflows enables reproducible evaluation, early detection of regressions and continuous validation .

Quantum Error Correction codes use multiple physical qubits to form a logical qubit to achieve a lower logical error rate . The main challenge for QEC is to achieve error correction with high accuracy within the tight $1μs$ decoding time budget imposed by superconducting qubits .

Convolutional neural networks (CNNs) require a large number of multiply-accumulate operations . We formalise such sequences of steps and apply our formalism to a state of the art decomposition of convolutions . A Python-based simulator allows to analyse in-depth computed strategies .

Accurate and reliable safety metrics are paramount for functional safety verification of ASICs in automotive systems . Traditional FMEDA metrics depend on precision of failure mode distribution (FMD) and diagnostic coverage (DC) estimations . This reliance can often lead to significant, unquantified uncertainties and a dependency on expert judgment, compromising the quality of the safety analysis .

As AI agents transition from human-supervised copilots to autonomous platform infrastructure, the ability to analyze their reasoning behavior across populations of investigations becomes a pressing infrastructure requirement . This paper introduces the Agent Execution Record (AER), a structured reasoning provenance primitive .

DataFlex is a unified data-centric dynamic training framework built upon LLaMA-Factory . It provides extensible trainer abstractions and modular components, enabling a drop-in replacement for standard LLM training . It unifies key model-dependent operations such as embedding extraction, inference, and gradient computation .

Graph Neural Networks (GNNs) face two challenges when scaled to deep architectures: oversmoothing, where node representations converge to indistinguishable vectors, and oversquashing, where information from distant nodes fails to propagate through bottlenecks . We prove that exact optimization for either problem is NP-hard .

PruneFuse is a novel strategy that leverages pruned networks for data selection and later fuses them with the original network to optimize training . It uses structured pruning to create a smaller pruned network that is well-suited for the data selection task . This small network is then trained and selects the most informative samples from the dataset .

Graph Neural Networks (GNNs) have achieved remarkable performance on tasks involving relational data . However, small perturbations to the graph structure can significantly alter GNN output . We propose PEANUT, a simple, gradient-free, restricted black-box attack that injects virtual nodes to capitalize on this vulnerability .

DPD-Cancer is a deep learning method based on a Graph Attention Transformer (GAT) framework . It is designed for small molecule anti-cancer activity classification and the quantitative prediction of cell-line specific responses .

Short-term (0-24 hours) precipitation forecasting is highly valuable to socioeconomic activities and public safety . The highly complex evolution patterns of precipitation events, the extreme imbalance between precipitation and non-precipitation samples, and the inability of existing models to efficiently and effectively utilize large volumes of multi-source atmospheric observation data hinder improvements in precipitation forecasting accuracy .

Large Language Models (LLMs) have advanced Graph Neural Networks (GNNs) by enriching node representations with semantic features . We propose a robustness assessment framework that systematically evaluates LLM-enhanced GNNs under poisoning attacks . The framework enables comprehensive evaluation across multiple dimensions .

Test-time adaptation aims to mitigate performance degradation under distribution shifts by updating model parameters during inference . AcTTA reformulates conventional activation functions into parameterized forms that shift their response threshold and modulate gradient sensitivity . This functional reparameterization enables continuous adjustment of activation behavior without modifying network weights or requiring source data .