Using Google AI credits, the author parsed about a million Wiktionary entries to create an etymology database, enabling search by affixes, cognates, derivatives, meaning, language of origin, and part of speech. The project is a work in progress with unverified data.

This study uses a two-stage large language model pipeline to automatically detect internal inconsistencies in discharge summaries from electronic health records. Applied to 3,000 MIMIC-IV-Note discharge summaries, the pipeline surfaced 3,460 candidate inconsistencies affecting 69.7% of admissions. Expert review revealed failure modes in temporal reasoning, evolving-diagnosis context, and outpatient prescribing knowledge. A graded ontology is proposed to categorize inconsistencies, providing a foundation for large-scale EHR inconsistency analysis.

A new study demonstrates that frontier language models can perform 'invisible reasoning' using semantically irrelevant filler tokens, improving accuracy on synthetic reasoning tasks by up to 13 percentage points. This computation leaves no interpretable trace in output tokens, posing challenges for AI safety.

This paper proposes LEACL, a framework integrating LLMs and automatic curriculum learning to tackle long-horizon manipulation tasks. LLMs decompose tasks into subtasks and generate specifications, enabling learning with sparse rewards only. On the LIBERO benchmark, LEACL outperforms human-designed dense rewards in success rates.

PRISM is a novel policy representation that explicitly models polynomial interactions among physical variables, improving robotics motor control performance and enabling sensorless compliance without force or tactile sensors.

This paper presents a generalized data-association-free SLAM framework that jointly estimates data associations, robot poses, landmark positions, and landmark semantics using deep learning semantic information (e.g., class labels and feature vectors). It adopts a semi-incremental estimation scheme for better accuracy and efficiency, and provides principled guidance for landmark-number estimation. Evaluations on synthetic and real datasets show superior performance over strong baselines with two types of semantic information.

A new framework that identifies elastic object parameters from a single non-destructive interaction, enabling policy learning in simulation and zero-shot transfer to real robots.

Researchers propose a two-stage framework for LLM-based dialogue robots that generates context-aware prefaces before speech onset, reducing response latency. Field experiments show timing trade-offs between filler types.

A new study reveals a super-exponential scaling law between required demonstrations and target precision in robotic manipulation. The precision limit is not a fixed constant but an emergent property of the robot system, including its sensors and expert policy. Improving system components can lower the precision limit, expanding achievable precision. The work provides a theoretical framework and practical metric for high-precision robotic systems.

Researchers propose the World-Cognition Model (WCM), a human-centered embodied agent built on the SLAK architecture (Sensing, Logic, Action, Knowledge) and an asynchronous runtime. It separates perception, reasoning, control, and memory, and introduces a human-in-the-loop teaching mode that enables users to interactively teach robots difficult or long-horizon tasks. WCM achieves a 73.8% average success rate across nine real-world human-robot interaction tasks, including held-out tasks and a long-horizon task learned through teaching.

This paper presents an end-to-end, fully AMD-accelerated technology stack for embodied manipulation, spanning data-center training silicon, Radeon PRO simulation/rendering GPUs, and Ryzen AI edge compute, unified by the open ROCm software stack. Four progressive demos demonstrate that VLA training and deployment do not require a CUDA-locked ecosystem.

A new algorithm uses hand pose data to reduce force sensing errors in flexible tactile gloves, improving minimum detectable force by up to 18.3% across multiple designs and users.

Researchers present a replay-constrained simulation framework using MuJoCo to personalize powered knee-ankle prosthesis controllers. By replaying recorded hip kinematics and ground reaction forces, the framework avoids modeling complex human neuromuscular control. Deep reinforcement learning simultaneously personalizes stiffness, damping, and equilibrium angle for both joints, maximizing a biomimicry reward from onboard sensors. Experiments with three transfemoral amputees showed strong predictive validity (Pearson r=0.96-0.997) and improved biomimicry rewards by 42-59% over unpersonalized baselines.

A new model, GeoMVC, uses geometric interaction and multi-view consensus to detect misogynistic memes, achieving top rankings in Malayalam and Chinese at the CC-MMD challenge.

StepX-Edge is a 0.9B-parameter on-device UI vision-language model that balances accuracy and efficiency through co-design of architecture, training, and deployment. Innovations include ULVE visual encoding, PDP connector, five-stage curriculum training, and a two-stage quantization scheme. It surpasses 2B-level models on benchmarks like ScreenQA and Chinese OCRBench, and runs efficiently on Snapdragon 8 Gen5 devices. Training data and code will be open-sourced.

The paper proposes LowAux-RDNet, a method that introduces a training-only low-pass reflection auxiliary objective (LowAux) on top of RDNet. This low-frequency constraint, combined with scene-balanced real pairs from RRW, improves generalization across diverse reflection distributions. A unified benchmark on five datasets yields macro-average PSNR of 27.546 dB, with SSIM, NCC, and LMSE all reaching state-of-the-art.

EditCLEVR is a new benchmark for evaluating compositional faithfulness in object-centric learning. It uses paired-scene interventions—exact attribute edits or no-edit rerenders—to test whether models correctly alter object representations under semantic edits. The benchmark introduces Scene-Graph Intervention Accuracy (SGIA) and Delta-SGIA metrics, requiring full after-scene correctness and single-change detection, respectively. Baseline evaluations reveal that compositional out-of-distribution degradation persists even with ground-truth masks, and locality or stability alone can overstate semantic faithfulness.

This paper proposes Delta-InvFormer, a differential Transformer-based neural network, to predict 2D light intensity distributions from visible-light camera video frames of tokamak plasma, significantly accelerating traditional methods while maintaining competitive accuracy.

This study proposes a Language-guided Segmentation-assisted Diagnostic Transformer (LSDT) for four-class prostate cancer classification on mpMRI, constructing a Prostate Cancer Histopathology Spectrum Dataset (PCa-HSD) and leveraging zero-shot segmentation for anatomical priors. In five-fold cross-validation on 344 patients, it achieves 0.633 average accuracy and 0.768 JointRecall, significantly improving fine-grained classification.

This paper introduces the Multimodal Interactive Motion Encoder (MIME), the first dedicated multimodal encoder for two-person interactive motion. MIME uses stream-based co-attention with explicit interaction features and curriculum-based contrastive training to capture individual and shared structures. On Inter-X text-motion retrieval, MIME achieves a 12.8% relative improvement in text-to-motion R@1 over baselines. As a frozen auxiliary prior, MIME improves semantic alignment in TIMotion and InterMask while maintaining FID, demonstrating that interaction-aware multimodal encoding enhances multi-person motion retrieval and transfers to downstream motion generation.

FogDrive is a multi-modal autonomous driving dataset built with the CARLA simulator, featuring 660 scenes (~133k frames) across four synchronized cameras, LiDAR, and radar. It models physically consistent fog at three visibility levels (160m, 100m, 50m) using Koschmieder's model for cameras and Beer-Lambert law for LiDAR, providing matched clean and foggy variants for each scene. A quality audit over 8k frames confirms 95.1% annotation precision and over 99% recall for vehicles within 40m. Baseline evaluations show that mixing multi-density fog during training improves 3D detection, while 2D image quality metrics are poor predictors of downstream performance.

MegaSlide-DiT proposes a system-level approach to adapt a 105B-parameter DiT on a single H200 GPU by keeping model state in host memory and streaming transient shards, combined with 3D deformable slide attention that reduces memory and compute complexity to linear.

Foundation vision models trained on natural images can be applied to medical tasks without domain-specific pre-training. This study compares three aggregators for renal tumour and cyst detection using DINOv3 features on KiTS23, finding that gated attention multiple instance learning (MIL) achieves the best performance, with AUROC of 0.74 for tumour and 0.80 for cyst. The prototype head fails on cyst detection, highlighting a trade-off between interpretability and performance.

Cross-lingual mismatch remains a key source of overall degradation in modern speaker verification. The TidyVoice2026 Challenge targets this setting with text-independent verification, comprising 3,666 training and 808 development speakers in 40 languages and 2,200 evaluation speakers in 38 unseen languages, without language labels at test time. Starting from the official SimAM-ResNet34 baseline pretrained on VoxBlink2 and VoxCeleb2 and fine-tuned on TidyVoice, we revisit Nuisance Attribute Projection (NAP) as a simple language-normalization step in the embedding space. We estimate a compact language subspace from cross-language same-speaker differences and project embeddings onto its orthogonal complement before cosine scoring with Adaptive Symmetric score normalization. This reduces deve…

Researchers propose CHiPS, a lightweight character-level authorship attribution method for Romanian texts. It combines a character histogram classifier (CH-SVM) and a positional signal classifier (FFT12-LR), requiring no tokenization, syntactic analysis, pretrained language models, or transformer fine-tuning. In closed-set experiments, CHiPS-F achieves 0.9310 accuracy on a 400-file dataset, but the authors emphasize the contribution is not about achieving the best classification performance, but rather exploring how far restricted, transparent character evidence can go under strict leakage control.

PatiGonit22K is an expanded Bengali mathematical word problem dataset containing 22,441 problems, including simple and multi-operation equations, culturally adapted and verified to support NLP research in low-resource languages.

This paper introduces LENS, a protocol to evaluate whether machine unlearning can suppress disinformation-aligned narrative frames in LLMs. Experiments on four multilingual models show effective suppression but also reveal entity recovery side effects.

Proposes AutoThinkSQL framework that integrates an auto-thinking mechanism into SFT and DPO for Text-to-SQL, dynamically bypassing reasoning for simple queries and invoking deep CoT for complex ones, achieving gains on Spider and BIRD benchmarks while reducing output tokens and latency.

This study examines how different reviewer guidelines—official conference guidelines vs. reviewer-imitating guidelines generated by LLMs from high-quality human reviews—affect automated peer review. Results show official guidelines align best with human judgments, while strict rubric scoring degrades performance.

MioFFAn is an open-source, customizable annotation framework that addresses the scarcity of high-quality datasets for automatically translating mathematical expressions into executable code (Formula Formalization). Based on the MioGatto architecture, it features equation selection, aided symbolic code specification, custom taxonomies, and partial automation via LLMs, enabling human-in-the-loop iterative refinement.