EvoPlan is a neuro-symbolic framework combining LLM fluency and classical PDDL planner guarantees. It has three parts: offline mining of global STL constraints from demonstrations, an evolutionary PDDL planner, and a constrained execution loop. All LLM calls use local open-weight models, enabling on-robot deployment without cloud dependency. Tested on Bench2Drive, HA-VLN-CE, and ALFWorld benchmarks.

This review synthesizes 183 contributions from 2017-2026 covering VLA architectures, training recipes, action representations, bimanual coordination (2022-2026), UAV navigation and control (2017-2026), language grounding, and cross-cutting concerns. It shows that strategies from bimanual VLAs transfer to aerial systems and identifies fourteen research directions.

This paper proposes CILC, a system that uses Secure Multi-Party Computation (SMPC) to detect loop closure candidates in multi-agent SLAM without exchanging global descriptors in the clear, protecting against compromised agents. Experiments validate real-time performance on visual and LiDAR descriptors with reduced information leakage.

RoboSnap is a real-to-sim framework that turns a single RGB image into a simulation-ready scene using a layered design: collision-aware foreground assets for stable robot interaction and 3D Gaussian splatting for faithful background appearance. Experiments on DROID scenes and real-robot tasks show reliable trajectory replay, task-specific synthetic data generation, and meaningful sim-real correlation. The work also introduces DROID-Sim, a companion dataset of 564 real-world scenes.

NativeMEM is a novel VLA model that compresses historical frames into single tokens using native memory compression, eliminating external memory modules. It boosts success rates from 32.4% to 84.0% in simulation and up to 98.7% on real robots, using only 20% training data.

Pelican-VLA 0.5 is a unified Vision-Language-Action (VLA) model that integrates vision-language understanding, future-frame generation, and action prediction within a single architecture. It achieves attention-level generalization through learnable Reasoning Slots inserted between perception and action, enabling the action pathway to focus on instruction-relevant objects and contact regions without explicit supervision. This behavior persists across unseen scenes and robot embodiments, outperforming other open-source VLA baselines.

This paper proposes ProMoE-FL, a prototype-conditioned mixture-of-experts framework for robust missing-modality feature synthesis in multimodal federated learning. It builds a global client-aware prototype bank capturing clinically meaningful modality priors across institutions, and uses direction-aware expert routing to dynamically synthesize missing features. Experiments on four chest X-ray datasets show consistent outperformance over state-of-the-art methods in both homogeneous and heterogeneous settings.

Video Diffusion Models (VDMs) achieve high quality but are computationally expensive. Recent few-step distillation accelerates inference but ignores varying computational demands across noise levels. This paper proposes a post-training framework that integrates dynamic structural sparsification into distillation, jointly optimizing denoising steps and model sparsity to create a step-specific Mixture-of-Models (MoM). A Progressive Training Strategy with Output Rollout Mechanism ensures stability, and a specialized inference engine enables efficient deployment. On Wan-14B, it removes 24% per-step FLOPs on top of 4-step distillation, achieving 1.2x wall-clock gain and 30x speedup over the 50-step teacher with competitive quality.

SpaR3D-MoE is an end-to-end framework enabling adaptive 3D spatial reasoning from sparse RGB inputs via adaptive spatiotemporal manifold sampling and a geometry-inductive mixture-of-experts. It achieves SOTA on VSI-Bench (63.5 average, +7.8 over baseline), with 35.4% and 51.4% improvements on Route Plan and Relative Direction tasks respectively.

This paper introduces the Difficulty-Aware Medical Instructional Video Question Answering (DA-MIVQA) shared task for NLPCC 2026. It extends previous benchmarks by explicitly distinguishing questions based on the type and complexity of evidence required. Simple questions can be answered from subtitle text, while complex questions require visual grounding, procedural understanding, and cross-modal integration. Three tracks are included: DA-TAGSV, DA-VCR, and DA-TAGVC. The dataset is collected from public medical instructional channels, covering first aid, emergency response, rehabilitation, nursing, and general medical education, with manual difficulty annotations.

This study investigates the relationship between counterfactual fairness (CF) and group fairness (GF) in image classification. By constructing new datasets with high-quality image editing, it finds that CF does not imply GF in images, contrary to tabular data results. The discrepancy is attributed to a latent attribute correlated with the sensitive attribute. The proposed Counterfactual Knowledge Distillation (CKD) method reduces reliance on this attribute, allowing CF-achieving models to also satisfy GF.

Line segment detection is key in visual SLAM, 3D reconstruction, and industrial inspection. Deep learning models require several megabytes of memory, exceeding MCU capacity. This work proposes MiLSD, a detector for sub-megabyte budgets. It compares three output representations in a compact backbone, finding F-Clip (center with length and angle) most effective. 8-bit quantization preserves full precision; 4-bit degrades angle regression. With 1 MB activation budget, MiLSD improves sAP10 from 10.6 to 24.1 on ShanghaiTech Wireframe. The paper maps accuracy-memory trade-offs for embedded vision.

This paper introduces LipSSD, a Lipschitz-constrained Single Shot MultiBox Detector (SSD) that improves adversarial robustness by design. On Pascal VOC, adversarially trained LipSSD improves mAP@50 on unseen attacks by up to 15 points over standard adversarially trained SSD, and shows robustness gains on safety-critical datasets LARD and KITTI while largely preserving clean performance.

This study fine-tunes two pretrained latent diffusion models, Protogen v3.4 and Stable Diffusion v1.4, on a curated dataset of high-resolution Ulos motifs to generate culturally consistent yet novel designs. Protogen v3.4 significantly outperforms Stable Diffusion v1.4 in terms of FID and IS, highlighting a fidelity-diversity tradeoff. A guidance scale of 5–9 is recommended for optimal balance.

CoFINN is a physics-informed deep learning framework for predicting compressible flow fields, embedding finite-volume conservation physics directly into the training process. It outperforms traditional data-driven CNNs and classical physics-informed methods, reducing drag prediction error by up to 34% on transonic airfoil flows and 15% on average, especially beneficial in limited-data regimes.

The paper proposes a unified deep learning pipeline integrating semantic segmentation, regression-based severity estimation, and disease classification for plant disease severity quantification. On the Apple Tree Leaf Disease Segmentation dataset, U-Net with MobileNetV2 achieves 98.20% pixel accuracy, 0.70 mIoU, and 99.41% detection accuracy at 14.7 ms per image, suitable for real-time use. The computed severity index strongly correlates with expert annotations (r=0.968), demonstrating reliability for automated crop monitoring.

A new evaluation framework predicts item psychometric parameters from text embeddings, revealing that difficulty is highly predictable while discrimination and pseudo-guessing are limited by reliability ceilings. The study highlights the need for repeated cross-validation and scale-free metrics in benchmark construction.

This paper analyzes over-calling in multi-teacher on-policy distillation for tool-using language models and proposes Soft Clamp, a per-token divergence calibration method that reduces over-calling from 13.7% to 9.0% while matching decision accuracy.

This work proposes Riemannian Mean Pooling (RMP), which extracts per-token pullback metrics from a learned encoder's analytical Jacobian and aggregates them via Fréchet mean on the SPD manifold. RMP outperforms Euclidean mean pooling on CoLA, CREAK, and RTE, while staying at chance on FEVER-Symmetric. Ablations show that even a randomly initialized encoder with Fréchet aggregation beats Euclidean pooling on most datasets, locating the gain in geometric aggregation.

Large language models increasingly improve reasoning via test-time computation, but most methods treat problems in isolation. MILES introduces modular memory units with learnable selection heads to accumulate reusable experience across sequential problems, achieving superior accuracy-efficiency tradeoffs.

A new multi-factor scoring framework integrates five dimensions to evaluate LLM response quality, revealing strengths in reasoning but significant weaknesses in factual consistency and ambiguity handling.

Large language models (LLMs) systematically prefer Standard American English (SAE) continuations even when the preceding context is in African American English (AAE), effectively rewriting AAE into SAE. The authors propose an end-to-end framework to audit and mitigate this bias, including conditional Dialect Group Invariance (cDGI) and activation steering. They also release REAL-AAE, the largest real-AAE parallel corpus to date.

A gradient-based speech-to-text alignment method applicable to any differentiable ASR model, including CTC, transducer, attention-based encoder-decoders, and speech LLMs. No training or model modification required. It aligns on the input grid for finer precision. Evaluated on 16 models, it provides usable alignment for all, outperforming native alignment on streaming models, at the cost of one backward pass per token.

This paper proposes a programmatic solution to generate advertising headlines for e-commerce using a self-critical masked language model optimized with reinforcement learning policy gradients. The method conditions on multiple products jointly and outperforms existing Transformer and LSTM+RL methods in overlap metrics and quality audits, even surpassing human-written headlines in grammar and creativity.

Large language models (LLMs) achieve promising results on medical question answering benchmarks, yet their use in public health is constrained by hallucinations and the rapid evolution of official guidance. Retrieval-Augmented Generation (RAG) mitigates these risks by grounding responses in an explicitly maintained corpus, but end-to-end performance depends critically on retrieval configuration and on evaluation beyond multiple-choice formats. We extend PubHealthBench into a retrieval-augmented setting and systematically evaluate retrieval and generation choices. Hybrid retrieval consistently improves recall and ranking quality. Providing retrieved context substantially increases multiple-choice accuracy across a diverse set of LLMs, enabling smaller open-weight models to match or outperf…

This paper proposes a multimodal solution that integrates audio and automatically generated multilingual text transcripts via cross-modal transformers, achieving significant performance gains in sentiment polarity classification. Knowledge distillation further enhances an audio-only model without extra inference cost.

TSF uses LLMs offline to build task-semantic fields, enhancing online forecasting without LLM overhead. Achieves 6.4% average MAE reduction with only 1.8-3.0k additional parameters.

This paper investigates how the spectral properties of the attention score matrix are influenced by positional encoding. Analyzing seven pretrained models, they find that previous-token heads under RoPE exhibit rotational spectra, while those under learned-absolute and ALiBi do not. Dynamic analysis shows spectral signatures emerge after behavior, and causal experiments demonstrate that no spectral channel is necessary but bans delay formation.

Inertia-1 is a fully open exploration of wearable motion foundation models using over 18.2 million hours of accelerometer data. It systematically studies data, model, and training choices, evaluating on 15 datasets for tasks like human activity recognition, freezing-of-gait detection, and disease prediction. The work provides state-of-the-art recipes and an open cookbook for wearable motion representation learning.

A new method, FedEAS, addresses label skew in federated learning by adaptively allocating per-class generation budgets to clients based on their local label distributions. It recovers most accuracy gains of full class balancing while reducing the generation budget by 94.1%.