This paper introduces Topo4Vec, a GeoAI framework using topological error simulation and spatial representation learning to automate geospatial vector data quality assessment, achieving peak accuracies of 0.99 for overlapping building footprints and 0.60 for street network errors.

We present a two-stage pipeline for player-centric ball action spotting in soccer videos. A Track-Aware Action Detector (TAAD) with temporal transformer produces per-player action logits, and a Denoising Sequence Transduction (DST) transformer converts game-state features and TAAD logits into structured events. Spatial-first attention ordering improves Macro-F1 by 1.87%. A weighted ensemble with agreement filtering raises Macro-F1 from 48.6 to 58.94 on the challenge.

This paper presents a post-hoc framework to detect characteristic patterns in images generated by image autoregressive models (IARs), enabling reliable tracing of generated images to their source model without modifying the generative process or outputs. It is applicable to already-published content without watermarks or models lacking watermark integration, aiding in misinformation prevention, fraud detection, and harmful content attribution. The method demonstrates effectiveness across various IARs and is accepted at ICLR 2026.

A novel method detects driving scenario complexity without any labels by using a Joint Embedding Predictive Architecture (JEPA) trained on structured agent state data from the nuPlan mini dataset. Temporal prediction error serves as a zero-shot complexity score, effectively distinguishing complex scenarios (e.g., unprotected turns, crosswalk interactions) from simple ones (e.g., lane-following). Downstream anomaly detection achieves Average Precision of 0.512, surpassing a 0.436 chance baseline.

A novel unsupervised framework using a memory-augmented autoencoder achieves state-of-the-art results on IMU-based human activity recognition, reaching 96.6% and 98.4% accuracy on DaLiAc and PAMAP2 datasets respectively, surpassing both supervised baselines and unsupervised approaches.

GeoISF is a novel large-scale LiDAR-to-image geo-localization pipeline that constructs an instance semantic forest using WordNet, enhancing temporal semantic representation and discriminative power. It bridges the modality gap between point clouds and satellite images, achieving 13.22 times better R@10 on KITTI than parallel methods.

This work evaluates the phonological perception of SLR models trained on American Sign Language (ASL) by probing phonological sensitivity using minimal pairs and evaluating representational alignment with human behavioral data. Results reveal emergent phonological sensitivity with architectural trade-offs: pose-based models are sensitive to handshape contrasts, while pixel-based models better capture location changes. Pose-based models learn latent representations that correlate with human perceptual similarity judgments (r~0.49).

While LLMs may agree with human annotators in coding text, reliability does not guarantee construct validity. This paper proposes 'grain calibration,' which decomposes a construct into clause-level components, tests each with extractive evidence, and combines results via an explicit rule. This reveals the process behind coding, shifting validation from output comparison to demonstrating that the model runs on the construct specified by its theory.

A new arXiv paper proposes SEAD, which uses entropy as a unified probe to address the issue of teacher supervision quality varying with student competence in on-policy distillation (OPD) at three scales: token partitioning, KL divergence annealing, and curriculum learning. It achieves a +4.8% average accuracy improvement on OLMo-3.

Study compares four methods for setting depth-wise alpha in sparse self-attention, finding that static per-layer stagger outperforms fixed and learned alpha, and all sparse variants extrapolate to 4x training length while dense baseline collapses.

This paper introduces turn-averaged sparse autoencoders (SAEs), which represent each human or assistant turn with a fixed number of features by reconstructing the average model activation across the turn. This addresses the issue of features scaling linearly with context length in standard SAEs. The turn-averaged features describe high-level characteristics more completely and simplify downstream uses like attribution graphs.

Researchers further pre-trained ModernBERT on all US court opinions, achieving significant improvements over vanilla ModernBERT on legal tasks. Despite ModernBERT's 500x larger pre-training data compared to original BERT, domain adaptation still yields gains. From-scratch pre-training underperforms continued pre-training. The model handles up to 8,192 tokens and is useful for legal text embeddings or reranking. All checkpoints are released. To appear at ICAIL 2026.

The study introduces a French OSCE dialogue dataset of 240 student-patient interactions, and builds a controllable LLM-based pipeline to generate synthetic OSCE dialogues. The pipeline uses retrieval-based grounding and a reflection loop for patient fidelity. A multi-level evaluation framework with LLM-as-a-Judge is proposed. Experiments show controllability modules improve patient fidelity and evaluation consistency. An interactive prototype with automatic feedback is implemented.

This study adopts a developmental perspective to examine false belief task (FBT) performance in Olmo2 and Pythia language models across training stages. Results show that above-chance FBT performance depends on model size and training volume, emerges late in pretraining, and is most improved by post-training interventions. However, FBT performance is fragile, and situation modeling accuracy generally precedes it. Larger models build partially coherent situation models but display surprising fragility.

This paper presents a method for automatically extracting lexical information from a machine-readable version of the Arabic-English Al-Mawrid dictionary. Using n-gram and keyword-in-context (KWIC) analysis to discover lexical patterns, and hand-crafted rule-based information extraction, the study achieved high precision across all information types, high recall for synonyms, but low recall for other types. The Al-Mawrid dictionary was found to contain substantial derivations, synonyms, domain labels, and hyponym/hypernym relations.

The classic paradigm of language identification in the limit models learning as a game between an adversary, who reveals strings from an unknown target language, and a learner tasked with identifying that language. The recently introduced framework of language generation in the limit shifted the objective to better reflect modern language modeling, requiring the learner to produce valid, unseen strings from the target language. Related work highlighted a fundamental tension: a broad coverage of the target often comes at the cost of validity. We introduce a new notion of precision and recast this problem as the classic recall-precision trade-off. We analyze generation in the limit under varying constraints on enumeration, novelty, and validity, aimed at reflecting settings closer to those encountered by large language models. A key contribution is our analysis of learners that are not eventually valid: we allow infinitely many mistakes, provided their frequency tends to zero so that precision remains one. We show that this relaxation can strictly increase recall when the adversary permanently withholds a large portion of the target language. We also study a continuous relaxation of the novelty constraint that requires only a fixed fraction of outputs to be novel. Taken together, our results move toward a more realistic model of language generation where occasional errors and repetitions are unavoidable, but their rates are controlled.

This research demonstrates how emotional valence influences the order-dependent structure of children's recognition memory. Correct recall of a sequence of emotionally-valenced toys depended not only on the valence of the toy itself but also on the valence of toys shown before and after it. While standard psychological models show low accuracy, a classical tensor network model incorporating valence achieves 77.98% accuracy, highlighting the value of quantum-inspired methods for modelling order-dependent phenomena like emotional memory. The study also introduces a novel protocol for exploring emotional temporal memory in children.

This paper proposes an end-to-end agentic pipeline combining deep time-series forecasting, variational anomaly detection, and LLM reasoning to generate prioritized, actionable maintenance recommendations for office building appliance-level energy monitoring. The system uses a hybrid SSA-LSTM forecasting model and per-appliance LSTM VAE with attention for anomaly detection, with a three-stage LangChain pipeline (Context, Diagnosis, Report agents) featuring dynamic retrieval. Evaluated on a 16-scenario benchmark, the best backend scores 90.4/100 and a local 7B model passes all scenarios.

A new paper applies singular learning theory to deep monomial networks, showing that critical points correspond to subnetworks, providing a mathematical explanation for neural networks' implicit bias towards simpler functions.

Proposes Counterfactual Residual Data Augmentation (CRDA) for tabular regression, leveraging residual invariance under feature perturbations to generate realistic training samples. Experiments show average MSE reduction of 22.9% for MLP and 6.4% for XGBoost, outperforming existing methods.

Classical universal approximation theorems ensure sigmoidal MLPs can represent any function, but do not guide weight initialization to encode data geometry. This paper proposes S-GAI, a spectral geometry-aware initialization framework for one-hidden-layer sigmoidal MLPs. It uses SVD to estimate class-wise spectral geometry from images: mean, principal directions, and spectral scales. An energy threshold selects retained directions, each represented by two sigmoid gates. These class-specific gates form a shared hidden layer initialized from the training set. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 show S-GAI yields more informative hidden states than Xavier initialization, with comparable final accuracy under full training. Freezing the hidden layer and training only the output layer still outperforms frozen random gates, confirming effective embedding of spectral geometry.

A novel online distributed sensing framework combining partial domain knowledge with deep neural networks for latent state estimation without requiring noise statistics. The proposed CA-NKCF outperforms traditional filters and model-free DNNs across linear, chaotic, and wireless environments, demonstrating robustness to model misspecification.

This position paper argues that reinforcement learning researchers often conflate two distinct uses of simulators: solving the simulator as an end in itself, and using it as a proxy for real-world deployment. Through experiments and examples, the authors show that failing to distinguish these settings can lead to misleading conclusions, and call for clearer empirical practices.

A preprint proves that in a mesh of sovereign agents without central coordination, the substrate for each agent must belong to the continuous-time liquid class to optimally process irregular, asynchronous observations. Two necessary conditions are identified: an adaptive timescale and sensitivity to elapsed inter-arrival times, which scale cannot compensate.

Existing image-generation benchmarks fail to assess the usability of scientific figures. SciDraw-Bench introduces 32 tasks across 8 figure types and 10 disciplines, with a four-dimensional evaluation protocol. Experiments show that a domain-specific system, SciDraw AI, outperforms general-purpose models across all dimensions, while text fidelity remains the hardest challenge.

This paper uses evolutionary game theory to model when a harm-minimizing AI agent can displace an approval-seeking (RLHF) agent in a competitive market, and whether that policy suffices to prevent community harm. It shows adoption is favored under certain prior distributions, a critical adoption level exists, and self-audit alone is insufficient without alignment of values and proper evaluation timeframe.

Critic-free RL with verifiable rewards (RLVR) like GRPO avoids training a value function but can be unstable when all rollouts in a group receive identical rewards. BV-Blend stabilizes advantage estimation by combining prompt-local statistics with historical moments from semantic clusters, using a confidence weight derived from a standard error of the mean proxy. Experiments show improved stability and performance on verifiable reasoning benchmarks.

COMPASS is the first unified multimodal framework that grounds composition-intent control in a single system, using a shared expert token τ_c for both perception and generation. It injects composition expertise into an MoE backbone, distills intent into τ_c, and reuses it as a conditioning signal for layout control. The companion Comp-11 dataset features an 11-class taxonomy and reasoning-augmented annotations. Experiments show significant improvements in composition understanding and generation consistency.

Researchers introduce ATHENA-R1, an AI agent trained via reinforcement learning to perform treatment reasoning across 212 biomedical tools. It outperforms GPT-5 on benchmarks and is preferred by experts and physicians.

VirtueMap is a framework that profiles Large Language Models using Aristotelian virtue ethics. It presents seven general ethical dilemmas, each with five possible responses, and asks humans or LLMs to rank them by virtue. Reference orderings are validated by over 100 respondents with at least 95% agreement. Applied to nine LLM families, it finds 90.3% mean rank consistency, with largest differences in Courage, Temperance, and Justice.