AI can generate code quickly but lacks architectural insight. This article argues that as implementation becomes cheap, architecture becomes the scarce resource that turns code abundance into coherent systems. Good architecture is more critical than ever in the AI era.

Geomagnetic sensing provides an infrastructure-free, absolute orientation reference robust to GNSS denial and visual degradation. However, no large-scale outdoor robotics dataset supports its systematic study in SLAM. Mag4D-SLAM fills this gap with 14 sequences totaling over 18 km of synchronized LiDAR, camera, IMU, tri-axis magnetometer, and GNSS measurements with SE(3) ground-truth poses. Collected along structured campus trajectories under paired day/night conditions in both forward and reverse directions, it enables analysis of magnetic field repeatability, drift-free global heading estimation, and location-discriminative magnetic signatures. The dataset supports research on yaw drift mitigation, magnetic loop closure, and long-term localization.

ACME is a large-scale, multi-modal dataset for social navigation, collected across 8 sites in 5 countries using 7 robot embodiments. It provides 29.35 hours of onboard data and 43.5 hours of overhead pedestrian tracking, focusing on goal-driven navigation with robot speech interaction in diverse cultural settings.

This research demonstrates how deep reinforcement learning enables snake-like robots to adaptively move in dynamically changing viscous fluids, overcoming limitations of classical control methods. By using an asymmetric actor-critic framework with knowledge distillation from privileged simulator information, the robot learns non-sinusoidal gaits that improve speed and efficiency across a wide viscosity range, outperforming traditional sinusoidal and kinematic controllers.

A two-stage model-based learning pipeline for robotic ultrasound probe guidance: a latent conditional diffusion world model predicts future ultrasound observations, and a goal-conditioned temporal transformer predicts probe motions, achieving 70% carotid and 65% thyroid guidance success rates in real-world experiments.

This paper reformulates socially appropriate robot action generation as a preference modeling problem inspired by recommender systems. The proposed StARS framework integrates collaborative filtering with learnable scene representations to produce user-specific appropriateness scores. Evaluated on MannersDB+ and SocNav1, StARS consistently improves performance and annotator agreement, enabling personalized action selection.

Researchers introduce Pigey, a physical agent orchestrator that decomposes robotic capabilities into a high-level manager and low-level policy, achieving over 4x improvement on LIBERO-PRO and near-100% success on real-world reasoning tasks.

This paper introduces Ordered Action Tokenization (OAT), a learned action tokenizer that maps continuous robot action chunks to an ordered sequence of discrete tokens, achieving high compression, total decodability, and ordered token space. It demonstrates strong performance across various tasks and backbones.

GRACE is a novel method that guides pretrained diffusion policies with Model Predictive Path Integral (MPPI) control using only forward cost evaluations, enabling the handling of nondifferentiable constraints such as binary collision checks and joint limits. It achieves higher success rates in simulation and successfully avoids obstacles on a real 7-DoF manipulator.

This survey reviews progress reward modeling for robotic learning, addressing the insufficiency of terminal success signals. It proposes a unified framework covering the interface of progress models, signal construction methods, and supporting data/benchmarks, while summarizing limitations and future directions.

A new framework leverages generative AI to convert text prompts into realistic and diverse human motion sequences, enabling humanoid robots to learn multiple task execution styles without real-world demonstrations. Simulated experiments show successful task completion and strong adaptability to complex motion variations.

This paper presents ISPCloak, an optimization-free adversarial attack framework that leverages hardware-intrinsic statistical signatures from image signal processing (ISP) pipelines to cloak AI-generated images as authentic photographs, effectively bypassing current deepfake detectors. By projecting images into the RAW domain, injecting realistic sensor noise, and reconstructing the ISP pipeline, the method generates imperceptible adversarial examples that universally deceive detection mechanisms.

The paper introduces the Semantic Gaussian Allocation Transformer (SAGFormer), which explicitly allocates a fixed number of Gaussians to address the allocation bottleneck in 3D semantic occupancy prediction, improving prediction accuracy and semantic consistency.

A reproducible workflow maps farmland extent and visible boundaries from 1 m NAIP RGB imagery using a residual U-Net (ResUNet) with Dice-dominant loss (accuracy 0.8808, IoU 0.8605, Dice 0.9234) and a text-prompted SAM 3 branch. Fusion via logical OR improves Dice on difficult patches (orchard rows: 0.858 to 0.955; fragmented parcels: 0.804 to 0.903). The output is a semantic farmland-extent layer for agricultural monitoring where field maps are unavailable.

SCALE is a self-supervised framework that generates violation-prone layout variants to train a DRC-VLM, improving DRV fix rates by 12-25% on sub-2nm nodes.

Recent conditional video generation models can convert 3D engine renderings like depth maps into photorealistic videos, but suffer from appearance inconsistencies when the camera revisits a location after context eviction in long-horizon autoregressive generation. This paper proposes a training-free method that leverages correspondences from the 3D engine: temporal correspondence retrieves pose-matched historical latent chunks into the KV cache as loop-closure memory, while spatial correspondence biases token-level attention toward geometrically corresponding regions. Evaluations on loop-closure trajectories from TartanAir and TartanGround show improved revisit consistency without sacrificing video quality, outperforming existing training-free baselines.

This paper proposes Trans-Unet, a framework that transforms 3D point-cloud data into a 2D grid domain and integrates CNNs with self-attention mechanisms to accurately predict brain folding patterns with high fidelity and accuracy.

A new study examines how the size of a gallery database affects the accuracy of photo lineups in facial identification processes. Comparing galleries of 500, 5,000, and 24,000 images, researchers found that larger galleries increase both the likelihood of witness misidentification and witness confidence in that incorrect identification. This process has been linked to at least nine wrongful arrests, raising questions about the use of such images in lineups and whether lineup results alone should constitute probable cause for arrest.

A compact segmentation framework called Risk-routed Implicit Boundary Refinement (RIBR) is proposed, using implicit neural representation as a risk-routed residual correction rather than a full-mask predictor. It combines boundary-refinement implicit residuals, risk-routed residual control, and geometry- and speckle-aware boundary regularization. Evaluated on nine ultrasound datasets, RIBR achieves the best overall macro-average and consistently reduces boundary error under a compact parameter budget.

Large Vision-Language Models (LVLMs) struggle with robust visual reasoning. Existing benchmarks focus on symbolic math or simple vision tasks, lacking assessment of complex reasoning and logical consistency. This paper introduces ConVBench, a benchmark with logically equivalent question pairs across six categories, and two metrics: logical consistency and robust accuracy. The authors also propose ConVLM, using GRPO-based reinforcement learning with a novel consistency reward to improve LVLM reasoning.

Oxygen-TryOn is a unified foundation model for any-item virtual try-on, built specifically for the task rather than adapted from general-purpose image editors. It leverages a dedicated data engine and try-on-specific training to generate photorealistic images of a subject wearing reference items across diverse fashion categories. Unlike prior systems limited to single garments or garment-centric approaches, Oxygen-TryOn supports full/half-body views, variable references, and free multi-item composition while preserving identity and appearance. It employs a three-stage training recipe (CPT, SFT, RL) with a hybrid reward model. On public benchmarks and its internal bench, it achieves state-of-the-art single-item try-on and leads on multi-item, matching or surpassing leading proprietary and…

This paper presents the first systematic study of faithfulness in document-grounded podcast generation using LLMs, introducing a turn-level evaluation framework and the catch-n-repair method to improve faithfulness.

Historical documents are often illegible due to damage. Existing restoration methods using masked language modeling fail with named entities needing external knowledge. A new framework ARI combines LLMs with retrieval-augmented generation (RAG) to infer proper nouns. Experiments on Korean historical texts show significant improvements, validated by expert assessments.

New research shows that for closed-book QA using LoRA adapters, data quality matters more than model capacity. A single curation pass boosted accuracy from 57.7% to 85.7% on a 15-document corpus, outperforming even retrieval-augmented baselines.

Copyright-Bench benchmark evaluates LLM agents' compliance with copyright law in commercial tasks. Agents often choose copyrighted works despite public-domain alternatives, and open-weights models show increased violation rates under certain user preferences and time pressure.

Khondo is the first benchmark for splitting document packets of Bangladeshi government forms, featuring bilingual (Bangla-English) vision-native data with five concatenation schemes across 14 domains. Zero-shot MLLM evaluation shows good page clustering but poor page order reconstruction, revealing page order as the primary challenge.

This pilot study uses Natural Language Autoencoders (NLA) to probe whether Qwen2.5-7B-Instruct internally represents Colombian identity, socioeconomic status, or stereotype-related information when processing Colombian-Spanish and English prompts. Using 30 prompts arranged as 15 matched Spanish-English pairs spanning explicit Colombian cues, implicit Colombian cues, and neutral controls, it reports descriptive rates and qualitative evidence rather than statistically powered effects.

Teachers, conference chairs, and public readers judge writing from limited evidence, seeing only a finished document. Humanly is a writing platform that records the entire writing process and generates a sealed writing certificate with configurable anomaly review, supporting scenarios like assignments, peer review, and certification. User studies show its utility across roles, and a red-teaming study confirms its typing detector distinguishes human from automated typing.

A new study systematically compares expert-assigned versus automatic MeSH terms as classifier features for systematic review screening, finding that the apparent gap strongly depends on evaluation design, with implications for benchmark interpretation.

This paper introduces a consensus-based evaluation framework that measures relative preference among LLM-generated responses by having a panel of diverse LLMs vote on anonymized outputs. The Relative Intelligence Index (RII) aggregates inter-model agreement as a proxy for response quality. The study reveals consistent preference patterns across domains, but emphasizes that these reflect inter-model alignment, not human judgment. This scalable method offers an alternative for comparing LLMs when multiple valid answers exist.