This paper proposes a self-supervised representation learning framework for contact detection in legged robots using only joint encoders, eliminating the need for force sensors. It outperforms supervised and baseline methods and provides public code.

FlowPRO proposes a reward-free offline reinforced fine-tuning framework for flow-matching Vision-Language-Action (VLA) models. Its core algorithm RPRO combines a contrastive optimizer with an explicit proximal regularizer to eliminate reward hacking. Using a teleoperated intervention-and-rollback paradigm to collect paired trajectories, combined with smooth interpolation and batch mixing, it achieves dense per-state supervision. On four long-horizon bimanual tasks, FlowPRO achieves the highest success rate, outperforming four baselines.

This paper proposes a novel method for learning from demonstrations (LfD) on Riemannian manifolds using neural ordinary differential equations (ODEs). While traditional LfD operates in Euclidean spaces, robot states like orientation naturally evolve on curved spaces. The method efficiently estimates geodesics via neural ODEs, enabling natural motion generation between arbitrary points on the manifold, and decodes the geodesics back to task space for robot deployment. Simulation experiments validate the framework's effectiveness.

MoDex is a diffusion-based policy that enables a dexterous hand to sequentially grasp multiple objects without releasing those already held. By conditioning on opposition space and point cloud, it uses only a subset of finger degrees of freedom per grasp. Two-stage training (imitation learning + RL fine-tuning) improves success in simulation and real world.

VASO is a framework that uses formal verification to guide the self-evolution of LLM-generated robot skill contracts. On Clearpath Jackal and PX4 quadcopter tasks, it achieves 97.2% formal-specification compliance with fewer than 100 optimization samples, outperforming execution-feedback, prompt-optimization, and fine-tuning baselines. It is the first framework to close the loop between formal verification and self-evolving skills for physical AI agents.

This paper proposes an efficient method for computing distances between points and curves on Lie groups, using G-polynomial curves to reduce the problem to polynomial root finding. It significantly cuts computation time while maintaining accuracy, with practical formulas for SE(3) and experimental validation on a robotic manipulator. The code is publicly available.

Researchers propose a novel 4-segment, 8-joint quaternion-joint cable-driven redundant manipulator configuration that achieves a broader workspace at lower hardware cost. Residual reinforcement learning outperforms the state-of-the-art FABRIK algorithm by three orders of magnitude in positional and orientational accuracy, with a simpler control implementation. This work provides new tools for designing such manipulators and control systems.

OLIVE is a parameter-efficient online adaptation framework for exoskeletons that uses low-rank residual decomposition to personalize control during deployment. It relies solely on on-body sensor feedback (EMG, IMU, vibration) and a dynamic rank scheduler to adapt to terrain complexity. Experiments show improvements in gait smoothness, effort reduction, and motion stability.

This study develops deep learning models for automated staging of age-related macular degeneration (AMD) using OCT/OCTA data. Among 271 participants, three models were tested: biomarker-based, 2D en face projections, and 3D volumes. All models showed strong performance, with the biomarker-based model achieving the best overall results (QWK=0.85) and particular strength in early AMD detection.

A deep learning method restores capillary anatomy from a single OCTA volume, significantly improving image quality and addressing 3D vascular architecture for the first time.

Biomazon is a 20 m multimodal benchmark dataset covering the Amazon Basin that pairs GEDI RH and AGBD targets with multi-sensor predictors for joint prediction of the full GEDI RH profile and aboveground biomass density. It provides standardized spatial splits and evaluation protocols, along with a baseline framework and comprehensive ablation studies on model scale, modality contributions, and auxiliary embeddings. Biomazon aims to advance structurally consistent RH-profile prediction and structure-biomass modeling in tropical forests.

This paper presents RePHO, a physics-guided reconstruction framework that recovers physically plausible human-object interactions from monocular videos. It starts with a kinematic estimate and refines it via reinforcement learning in a physics simulator, using an adaptive sampling strategy to handle noisy estimates. Results show clear improvements on two benchmarks.

LightVesselNet is an efficient neural network with only 75K parameters designed for retinal vessel segmentation in resource-constrained settings. It uses a compact encoder-decoder with channel and spatial attention, multi-scale feature aggregation at the bottleneck, subpixel upsampling, and edge residual connections. Experiments on five public datasets (DRIVE, STARE, CHASEDB1, FIVES, HRF) show competitive sensitivity (0.8096–0.8640) and Dice scores (0.7686–0.8649) while being more efficient than state-of-the-art models. Cross-dataset evaluation confirms generalization. It is a strong candidate for low-resource clinical deployment and mobile screening.

This paper presents TopoPult-SSL, a two-stage framework for cross-device meibomian gland segmentation. Stage 1 adapts without target gland masks, using eyelid outlines and clinical metadata as weak priors; Stage 2, when target masks are available, distills complementary teachers into a compact student via supervised self-distillation. On MGD-1k to CAMG benchmark, the distilled model achieves Dice 0.716, surpassing UA-MT and ensemble teacher with a single pass. The gland-mask-free variant reaches Precision 0.694, significantly outperforming SAM/MedSAM.

Researchers propose a framework for cross-model safety steering that transfers a safety direction from a source LLM to a target image/video generator via a lightweight alignment, without requiring unsafe data on the target side. The approach achieves comparable safety improvements to native directions while maintaining generation quality.

This paper introduces the personal camera roll visual question answering task, constructs the camroll dataset with 50 users, 31,476 images, and 2,500 QA pairs, and designs camroll-agent, a conversational AI agent with hierarchical memory and efficient tools. Experiments show it outperforms baselines, highlighting the need for specialized approaches to personalized visual memory.

A new method called NIV automatically converts static fonts into variable fonts by predicting per-point displacements using neural networks, enabling continuous variation along design axes like weight and width. The model generalizes to unseen styles and complex glyphs, including CJK characters, and outputs standard variable font files.

Researchers introduce VideoKR, the first large-scale training corpus specifically designed to strengthen knowledge- and reasoning-intensive video understanding. It comprises 315K video reasoning examples over 145K newly collected, CC-licensed, expert-domain videos. They develop a human-in-the-loop, skill-oriented example generation pipeline and curate VideoKR-Eval, a new expert-annotated benchmark. Experiments show that models post-trained on VideoKR under a standard SFT→GRPO pipeline outperform prior approaches on knowledge-intensive video reasoning while remaining competitive on general video reasoning.

LANTERN is a lightweight memory layer that proactively archives conversation turns and restores details after compaction via hybrid retrieval, requiring zero LLM calls and <25ms latency per turn. It recovers 78.3% of lost facts, outperforming MemGPT, and improves accuracy of production LLMs by 8.4 percentage points on average.

This paper proposes a novel Multi-Granularity Reasoning Network (MGRN) for Natural Language Inference (NLI). It explicitly leverages hierarchical semantic features to mimic the human cognitive process from lexical matching to logical reasoning, capturing complex semantic relationships. Experiments show MGRN consistently outperforms strong baselines.

This paper proposes a framework for sentence-level interpretability of rubric-based scoring, combining Shapley-value attributions with LLM-generated rationales. Tested on the CLASS Feedback quality dimension using the NCTE corpus, fine-tuned PLMs outperform LLMs in accuracy but show label compression. SHAP provides more faithful and transferable explanations than LLM rationales.

Existing multimodal safety benchmarks focus solely on visual inputs and cannot assess Omni Large Language Models (LLMs) that process vision, audio, and text. We introduce MCBench, a benchmark with 1196 scenarios spanning four safety categories that require integrating multiple modalities for accurate safety assessment. Each unsafe scenario is paired with a minimally different safe counterpart to assess model sensitivity. Our evaluations of state-of-the-art models reveal significant challenges. Omni LLMs struggle with subtle or non-physical risks but perform better when salient visual or acoustic cues are present. Analysis of reasoning traces shows that, although models can extract modality-specific information, they often fail to integrate these cues effectively for safety judgments. Our…

A systematic study of parameter-efficient fine-tuning using Low-Rank Adaptation (LoRA) applied to Qwen2.5-3B for building a domain-specific conversational assistant in telecommunications customer support. The research introduces a combinatorial synthetic data generation approach and evaluates 16 LoRA configurations, revealing a divergence between quantitative validation loss and qualitative human-aligned rankings, and provides an energy-performance trade-off analysis.

This paper studies generic triple-latent sequence models that maintain a running token state and compressed pair-memory pathway to capture higher-order token interactions without benchmark-specific parsing. The triple-latent family improves a small Transformer baseline on byte-level WikiText-2 and on a tokenizer-based MiniMind language-model benchmark, while a recall-focused gated key-value retrieval extension improves associative recall but remains seed-sensitive and much slower in the current reference implementation.

This paper proposes a Variance-Aware Reward Framework using Group Relative Policy Optimization (GRPO) for post-training LLMs on heart-focused medical question answering. The method replaces weighted binary criterion aggregation and single Likert scoring with continuous analytical reward functions, providing richer optimization signals. On the heart subset of HealthBench, the best variant improves accuracy from 0.362 to 0.502 and F1 from 0.532 to 0.668 over the Qwen3-14B base model, remaining competitive with GPT-OSS-120B.

This research proposes a hybrid pre-training objective combining JEPA latent-space prediction loss with standard masked language modeling (MLM) to improve language representations. Experiments show that the hybrid encoder produces more uniform embeddings and better semantic-lexical balance, but similar downstream accuracy to pure MLM.

Researchers propose a bilayer SIR/SIRS framework to model cross-contamination between AI models and data corpora, finding synthetic text detection and herd immunity as key intervention strategies.

DiffSlack introduces a differentiable projection layer that uses learnable slack variables to enforce nonlinear inequality constraints in neural networks, achieving improved performance in vehicle path planning with collision avoidance and curvature limits, validated in simulations and real-world tests.

Researchers propose a differentiable framework to automatically search for optimal token reduction operators in multimodal foundation models, achieving competitive accuracy-efficiency trade-offs even under aggressive visual token reduction.

Recent analyses of multi-pathway Deep Linear Networks use Gradient Flow to predict a 'winner-takes-all' specialization in which path symmetry breaks and each feature concentrates in a single pathway. This work shows that discrete Gradient Descent (GD) with a large step size tells a different story. We prove that single-path solutions are sharp minima, whereas distributing signals across pathways reduces sharpness by a factor that decreases with both the number of pathways and depth. Consequently, while early training reproduces the depth-driven symmetry breaking predicted by GF, oscillations at the Edge of Stability subsequently override this tendency and drive the network into a re-balancing phase, where signals redistribute across pathways. Together, these results clarify how depth shap…