MoMo is a two-stage imitation-learning framework that enables robots to adjust their motion mode (steady, dynamic, intermediate) during manipulation tasks. Using a spatiotemporal action tokenizer and a behavior-cloning transformer, the method takes task and a continuous motion-mode condition as inputs. Experiments on six real-robot tasks demonstrate compositional generalization to unseen task-mode combinations, showing that motion mode can be reused across tasks.

This paper introduces MMOCIC, a multi-objective framework that blends coevolved behaviors with compliance behaviors to balance team progress and norm adherence. By decoupling learning from compliance, it achieves high performance with fewer collisions. Demonstrated on a swimmer rescue mission with up to 8 robots in hardware and 12 in simulation.

This paper proposes a systems framework defining trustworthy embodied intelligence as sustained safe success—the ability to reliably perform tasks under variability while keeping risk acceptable. Four interdependent layers (model, system, evidence, deployment) support this objective. A non-normative hierarchy of trustworthiness levels is introduced to grade claims for bounded deployment, comparison, research, and standardization.

BG-REAL is a benchmark focusing on background manipulation detection and localization, built from Open Images V7 with 7,000 samples covering six edit families and matched authentic controls. Evaluation reveals that re-encoding artifacts are a shared shortcut risk across baselines, with false positive rates ranging from 0.57 to 1.00.

This paper introduces Lag-Aware Cross-Hand Alignment (LACA), a lightweight module that explicitly estimates directional temporal-offset distributions between hand-specific feature streams. LACA improves dual-hand action segmentation performance on HA-ViD and ATTACH datasets with minimal parameter increase. A future-free variant LACA-C enables real-time perception.

This study presents federated physics-informed neural networks (PINNs) that combine federated learning with linear elasticity equations to model brain tumor biomechanics while preserving patient privacy. Simulations across three clinical sites sharing only model weights achieved 91.4% accuracy, outperforming the pooled baseline of 90.0%, with significant improvements in pituitary adenoma accuracy.

Video shadow removal in the wild remains challenging due to complex illumination, diverse shadow appearances, and limited training data. WildShadowRemover adapts a pretrained video diffusion model via LoRA fine-tuning, introduces detail injection and frequency-decomposed modulation modules, and uses monocular depth priors. It outperforms existing methods in shadow removal quality and temporal consistency.

Rad-JEPA 3D is a self-supervised pretraining framework that learns volumetric CT representations by predicting latent features from a masked view. Its core hybrid H-Mamba encoder combines Mamba state-space and grouped-query attention, and Hidden States Orthogonal Regularization (HSOR) improves representation quality. Pretrained on ~120,000 CT scans, it achieves SOTA with only 4.0B parameters on closed-ended VQA and spatial reasoning benchmarks.

This study presents a hybrid computer vision method combining Mask R-CNN and a color-based algorithm to quantify exposed skin in images for dermal exposure assessment. Using 170 indoor-painting images, the method achieved approximately 80% agreement with human estimates, offering a scalable approach for semi-quantitative exposure analysis. Future work includes body-part recognition, PPE detection, and video-based analysis.

DVPSFormer is a unified online architecture for efficient 4D scene understanding, featuring explicit scene discretization (ESD) and a discrete-to-continuous (D2C) depth head for single-pass metric depth decoding, along with an online majority voting (OMV) mechanism for temporal consistency in instance tracking. It achieves state-of-the-art results on Cityscapes-DVPS and SemKITTI-DVPS benchmarks, providing a streamlined solution for online robotic perception.

TraceCLIP is a training-free framework that recovers latent patch-level semantic evidence by isolating patch-specific terms from CLS attention output in CLIP, and converts these into a semantic-geodesic topology gate to calibrate patch affinity for dense reconstruction. It achieves 1.3 to 4.5 mIoU gains on eight zero-shot semantic segmentation benchmarks without extra training or external models.

This paper explores deep neural networks for estimating BMI, weight, and height from a single daily-life image. The authors introduce a new dataset of 6,105 diverse in-the-wild images with ground truth labels, and evaluate multi-modal inputs including RGB, depth maps, pose-affinity maps, and edge maps. Experimental results show that full-body images outperform half-body or face-only images.

This paper proposes Knowledge-guided Disentanglement with Atomic Actions (KDA), which uses LLMs to decompose action labels into atomic actions, and employs knowledge injection and disentanglement modules to enhance video features for fine-grained action recognition, achieving state-of-the-art on multi-label benchmarks.

AgentGUI is a user-friendly, locally hosted GUI for observing and steering AI agents across multiple concurrent, long-running sessions. It features rich trajectory visualizations, effective manual and automated steering, and integration with open-source and frontier agent frameworks. A user study shows a 38% faster identification of key elements from agent traces (p=0.023), and its automated drift prevention improves task completion rates of small local agents by up to 34 percentage points.

This paper investigates prompt scope and demonstration selection as experimental variables in local LLM machine translation. Comparing three local instruction-tuned LLMs against dedicated MT baselines, the study finds that dedicated systems remain strongest, especially for Germanic languages. Few-shot prompting helps some models but hurts others, and family-scope prompting exposes failures in smaller models. The results call for evaluating LLM translation beyond language pairs and metrics, considering prompt scope, retrieval strategy, and multi-target compliance.

This paper presents Robostreet Flow, a freight architecture that jointly optimizes the vehicle, convoy formation, and operating model to reduce cost per ton-mile by 56%. The platform uses a battery-electric 6x4 tractor with drag coefficient 0.35, carbon-composite monocoque reducing weight by 50%, and a four-truck convoy where three followers operate at SAE Level 4 autonomy. Close following at 8m gap reduces follower drag by 42-48%, and fleet-average energy consumption is 1.27 kWh/mi, 20.5% lower than an isolated vehicle.

Researchers have built a corpus of over 60 million diarized lines of speech from religious radio broadcasts, recorded from 785 webstreams over one month in July 2025, enabling analysis of social and political issues in religious media.

A new method called V-Steer uses inference-time editing of cached value vectors to enforce instruction hierarchies in LLMs, improving primary constraint accuracy from under 18% to 92% on role conflict benchmarks without training.

This study addresses the challenge of distinguishing AI-generated poetry from human-written poetry by proposing a zero-shot detection pipeline that extracts key attributes to improve classification accuracy and reduce training needs.

This paper evaluates content moderation configurations using end-to-end metrics Usefulness and Harmful Exposure, comparing Input-only, Response-only, and Input+response hard blocking on product and public datasets. Response-only achieves highest filter-only Usefulness, while Input+response yields lower Harmful Exposure. Replacing Response-only blocking with Response+rewrite recovers most blocked traffic without increasing harmful exposure. Probe routing substantially reduces latency compared to LLM routing. Rewrites balance filter passage and usefulness by generalizing triggering language, but may omit safety-supporting info in sensitive domains.

DuplexGen is a framework that calibrates LLM predictions with a small set of human preference annotations to generate scenario-adaptive turn-taking dialogues. In six cooperative and competitive tasks, it aligns much closer with human preferences than uncalibrated prompting or generic human-human training data, showing that human calibration is key.

With the growing volume of scientific submissions, large language models (LLMs) are being used to assist peer review. Existing automated novelty assessment often compares claimed contributions against prior literature, but human reviewers frequently challenge novelty claims because the methodological evidence in the paper does not adequately support them. This paper introduces intra-paper claim verification, a framework that uses an LLM to extract novelty claims, retrieve relevant methodological evidence, and assess whether methods substantiate the stated contributions. Evaluation criteria are derived from 182 ICLR 2025 human reviews. Experiments show significant alignment between framework assessments and human reviewer concerns, especially for novelty-related issues.

This paper presents a two-part contribution for large-scale chatbot validation: a methodology for creating high-fidelity synthetic customer agents (SCAs) as digital twins, and an SCA-based validation framework combining automated LLM-as-a-Judge evaluation, human expert testing, and adversarial probing. The approach was validated at a leading UK bank, providing a scalable pathway toward regulatory compliance.

This paper studies a stochastic multi-objective bandit problem where an agent selects a slate of $k$ arms each round under semi-bandit feedback. The goal is to maintain a small set of actions that jointly approximate the Pareto frontier, formalized via dominated hypervolume. They define $\alpha$-approximate hypervolume regret and propose THV-UCB, an optimistic greedy algorithm, achieving gap-free and gap-dependent regret bounds.

Reinforcement learning with verifiable rewards (RLVR) faces a rollout generation bottleneck. SARA introduces a sequential adaptive allocation method that uses Bayesian posteriors and a two-threshold SPRT-style rule to identify effective or saturated prompt groups early, reallocating saved budget to new prompts. Experiments show SARA matches baseline performance while reducing rollouts by 22%-67%.

This paper introduces ULoRA, a unified family of LoRA initializations that connects existing methods as endpoints of a continuum. Tuned ULoRA matches or exceeds full fine-tuning on GLUE tasks, and the automatic variant ULoRA-Auto selects per-layer exponents without search, approaching the upper bound.

This paper proposes Weak-to-Strong On-Policy Distillation (W2S-OPD), a framework that improves a strong student model by distilling from multiple weak models. It constructs a proxy teacher in logit space using contrast pairs, outperforming standard OPD on math and code benchmarks.

Input-dependent controller coefficients are often treated as evidence of dynamic inference or computational savings. This paper introduces Frozen-Controller Auditing (FCA) to decouple these concepts, showing that dynamic parameterization alone does not imply dynamic inference or efficiency. Experiments reveal that static profiles retain over 98% performance and dynamic models can be slower.

This research investigates the transferability of supervised fine-tuning (SFT) lessons across alignment training, model organisms, and toy models. Through three experiments, the authors demonstrate that lessons in behavior generalization, capability preservation, and robustness can transfer across domains, suggesting researchers should borrow techniques from other areas.

Researchers propose modifications to encoder-decoder transformers for automated molecular structure elucidation from IR spectroscopy without requiring pre-determined chemical formulas. Using a Mixture-of-Experts (MoE) decoder with non-additive aggregation (linear-order statistics and Choquet integral) and an auxiliary contrastive alignment loss, they improve Top-K prediction accuracy by over 10 percentage points over baseline IR-only models. Analysis confirms IR spectra encode most relevant chemical information, broadening AI's utility in analytical chemistry.