MetaWorld proposes a novel framework to scale multi-agent video world models from single-view videos, addressing data scarcity and world state alignment. It uses Monocular World-State Unrolling (MWSU) to decompose camera ego-motion and subject trajectory, a Subject-Aware World Generator for appearance-driven simulation, and World-State Alignment (WSA) via cross-attention to ensure cross-view consistency. Experiments show superior cross-view consistency and identity fidelity.

Plan2Map is a 208-case multimodal benchmark for reconstructing geospatial boundaries from UK planning records. The proposed GeoPlanAgent system achieves 0.736 mean IoU and 0.904 median IoU, substantially outperforming direct VLM baselines, with errors concentrated in localization and map registration.

A new study reveals that state-of-the-art vision-language models often produce consistent yet incorrect predictions across viewpoints in spatial reasoning tasks, indicating a reliance on prior knowledge rather than visual evidence. The authors introduce ViewDiag, a multi-view evaluation protocol, to diagnose this issue.

AVTrack is a human-centric audio-visual instance segmentation dataset designed for dynamic real-world scenarios, featuring camera motion, visual occlusions, and position changes. Evaluations show substantial performance degradation of existing methods, establishing it as a challenging benchmark for robust scene understanding.

This paper introduces COD10K-C, a robustness benchmark based on COD10K, featuring 8 corruption types and 5 severity levels, totaling 40 conditions and 81,040 evaluation pairs. Evaluations on SINet-v2, PFNet, ZoomNet, and a lightweight model RobustCODLite show significant performance drops under corruptions, with motion blur and Gaussian blur causing the largest declines. RobustCODLite retains 92.3% of its clean Dice score via corruption augmentation, frequency-prior branch, and uncertainty-consistency loss, outperforming other models. The benchmark and code will be released.

A study probing Qwen3-14B hidden states shows that linear probes achieving 100% accuracy in classifying reasoning types (deductive, inductive, abductive) actually detect task format confounds (source, option count, response length) rather than genuine reasoning modes. After deconfounding, accuracy drops to chance, and causal steering shows no functional link. The findings urge routine format deconfounding in mechanistic interpretability.

To address inefficiency caused by verbose textual reasoning at every decision step in LLM agents, researchers propose Adaptive Latent Agentic Reasoning (ALAR), a dual-mode framework using compact latent reasoning for routine turns and explicit chain-of-thought for harder decisions. Experiments show up to 43.6% token reduction in search and 84.6% in tool use while maintaining task accuracy.

Inspired by Hayek's market theory, researchers propose an agent economy where agents compete via auctions, exchange payments, and accumulate wealth. This decentralized system produces emergent multi-step reasoning strategies and outperforms monolithic baselines across five tasks, suggesting a new path to multi-agent intelligence.

A systematic audit of NL-to-FOL benchmarks FOLIO and MALLS reveals that approximately 39% and 36% of entries contain incorrect FOL formalizations. The authors release corrected annotations and propose an LLM-based framework that reduces human review effort by over 70%, achieving 90% dataset accuracy after inspecting fewer than 24% of instances.

Researchers introduce Padyam2Gadyam, a dataset for translating 13th-17th Century Telugu classical poetry into contemporary Telugu and English prose. It includes 600 poems with human-verified translations. Five LLMs were evaluated, showing significant room for improvement in both languages.

A study shows that in high-stakes scenarios like legal, medical, and financial advice, even a single conversation history can lead to differences in LLM outcomes. While LLMs struggle to infer user sociodemographics, conversation topics act as proxies and affect advice unpredictably.

This study investigates how lexical overlap influences representations in large language models, finding that lexical effects persist across all layers and are consistent across architectures and training regimes. A mid-depth region shows simultaneous degradation of both lexical and semantic signals, and these effects impact downstream tasks like summarization and model editing.

A study benchmarks 31 LLMs on environmental attitudes, finding many more progressive than average humans, but with context sensitivity and sycophancy issues.

IdiomX is a large-scale multilingual benchmark for idioms, with 190K+ examples across 12K+ idioms in English, Arabic and French. It defines four tasks: idiom detection, context-to-idiom retrieval, Arabic-to-English retrieval, and idiom interpretation. Experiments show contextual models improve detection, and hybrid retrieval boosts cross-lingual performance.

This paper introduces Geometry-Aware Tabular Diffusion (GATD), which augments tabular diffusion denoisers with pairwise angles and lengths computed from column value differences as inputs and auxiliary targets. The MLP instantiation achieves state-of-the-art benchmark performance while using 3.5x fewer parameters on average (up to 25x for classification tasks): on ten datasets, it wins 8/10 Shape, 7/10 Trend, and 9/10 downstream utility (F1/RMSE), reducing Shape and Trend error by 27% and 20%. Default loss weights transfer to GNN and Transformer denoisers, improving Shape on 27/30 and Trend on 25/30 architecture-dataset cells. A matched ablation shows supervision (not extra inputs or capacity) drives the gain. This shows explicit relational supervision is a portable inductive bias for tab…

Large Language Models (LLMs) are increasingly deployed as continuously evolving services, where frequent base-model updates may invalidate previously deployed task-specific Low-Rank Adaptation (LoRA) adapters. ReLoRA is a knowledge-reusing re-adaptation framework that efficiently restores service-ready LoRA adapters for evolving LLM services while preserving or improving task performance. It consists of adaptive LoRA initialization using Bayesian optimization and fine-tuning with scheduled regularization. Experiments show ReLoRA reduces time-to-readiness by up to 8.9× and improves accuracy by up to 4.6%.

Introducing StenCE, a pretraining framework that uses cross-modal contrastive learning to extract stenosis-related signals from ECGs by aligning them with angiogram features. It enables non-invasive early screening, achieving state-of-the-art performance in severe stenosis classification without relying on invasive angiography.

ESG and climate risk data remain fragmented across Scope 1-3 reporting environments, lacking provenance-aware auditability and drift detection. This paper proposes a deterministic framework integrating single source of truth orchestration, temporal anomaly detection, imbalance-aware ensemble learning, and explainability-oriented governance. A synthetic benchmark calibrated to GHG Protocol, PCAF, and ISSB is released. Evaluation against multiple baselines shows improved recall, F1, ROC AUC, calibration, and audit trace completeness.

The paper proposes TopoMamSurv, a novel framework for whole slide image survival analysis that addresses Transformer's computational bottleneck and Mamba's sensitivity to input order. Using topology-aware ordering, bidirectional Mamba module, and GCN integration, it achieves efficient long-range dependency modeling and spatial structure utilization. Validated on five TCGA datasets.

Research reveals that within-dataset class-split evaluation for anomaly detection can become ill-posed when the held-out anomaly class overlaps the normal mixture in representation space. Scores may collapse to chance or invert, and preferred direction depends on unknown anomaly class. A training-free diagnostic, neighborhood class leakage, is introduced and shown to predict instability across multiple datasets and latent spaces. The paper argues that class-split benchmarks should be treated as geometry-dependent stress tests rather than unconditional evidence of detection ability.

Researchers propose a lightweight CNN to defend EEG-based brain-computer interfaces against adversarial attacks, showing improved robustness over existing models.

The paper introduces a Spectral Alignment Decomposition explaining why the curvature exponent α varies across layer types (≈2 for convolutions, ≈1 for attention, <1 for MLP up-projections). It reduces this variation to geometric alignment between Kronecker factor eigenbases and gradient singular directions. A spectral transfer identity s=αγ is derived, predicting Hessian decay exponent s with ~2% median error across 93 layers without free parameters. An architecture-adaptive preconditioner T(σ;α) yields Spectral Newton, outperforming AdamW on vision benchmarks.

This paper introduces the Human-in-the-Loop Gated Bandit (HITL-GB) framework for dynamic pricing in short-term rental markets, where a contextual bandit generates price recommendations that a human agent can accept, modify, or reject. The authors show that historical pricing data collected under a deterministic policy is structurally equivalent to on-policy warm-up data, bypassing the cold-start phase. Validated on real-world STR data from 1,461 nights, the warm-up reduces effective cold-start from ~150 to ~30 episodes. The approach is domain-agnostic, applicable to clinical drug dosing, credit origination, content moderation, and radiology.

AI-Driven Research Systems (ADRS) couple LLMs with automated evaluation to discover algorithms, proofs, and designs. This paper introduces GAMBLe, a framework that decomposes ADRS behavior into four parameters (generator, assessor, discovery mechanism, budget) and the effective landscape. Experiments on 760+ runs reveal no total ordering of components; correct choices can improve performance by 13-67% and search efficiency by 6-39x.

This paper proposes a modular reference architecture for embedded agent systems, introducing a tiered design that decouples on-device agents from cloud-augmented agents, and integrates a cross-cutting governance layer, to address challenges in deploying LLM-based autonomous systems on resource-constrained microcontrollers.

Large Reasoning Models improve by generating intermediate reasoning traces, but this paper reveals that continuing to reason after reaching the correct answer can be harmful. Using a prefix-level trajectory evaluation, they find that stopping at the first correct prefix improves accuracy by up to 21%, while common early stopping methods fail to mitigate harmful overthinking, which is driven by logical drift and visual reinterpretation.

This paper explores three novel approaches to generate enemy morphologies for video games based on player collision information, addressing a gap in procedural content generation. All methods perform comparably or better than an evolutionary baseline adapted from robotics.

Traj-Evolve is a self-evolving multi-agent system for patient trajectory modeling from longitudinal EHRs. It uses an Experience Pool (ExPool) for non-parametric memory and multi-agent reinforcement learning (MARL) for parametric optimization. On a lung cancer prediction task, it outperforms 9 baselines, with ExPool improving specificity and MARL improving sensitivity.

ChatHealthAI is a multimodal reasoning framework that aligns structured EHR representations from a pretrained EHR foundation model with the semantic space of a frozen LLM via a task-aware resampler. It integrates longitudinal patient data with refined clinical event descriptions to enable interpretable natural-language reasoning while maintaining predictive performance. Evaluated on three tasks from the EHRSHOT benchmark, ChatHealthAI improves reasoning quality and interpretability without sacrificing accuracy.

A new study compares encoder-only Transformer and LSTM for upstream streamflow inference in ungauged basins using NOAA National Water Model simulations. LSTM outperformed Transformer overall, and incorporating downstream information boosted median NNSE by over 60%. The findings highlight the importance of architectural inductive bias.