This paper introduces PIMbot, a framework that manipulates outcomes in multi-robot RL via two complementary levers: incentive manipulation of the reward channel and policy manipulation of an agent's own actions. An adaptive multi-objective controller balances these levers online. Experiments in Gazebo simulation and on NVIDIA Jetson Orin Nano embedded device demonstrate effectiveness, positioning PIMbot as a stress-test tool for vulnerabilities in multi-robot cooperation.

Researchers propose a method to certify reachable Cartesian steps under joint limits, achieving zero violations and 100% goal reaching in adversarial scenarios.

Robots learning reward functions from demonstrations often suffer from underspecified features due to imperfect demonstrations. This paper proposes a framework that detects underspecified features by analyzing variation across demonstrations (low variation indicates well-specified, high variation indicates underspecified). The robot then explains its uncertainty in natural language and requests targeted corrective demonstrations. Evaluations in simulation and with a real Franka robot show that explanation-guided queries significantly improve reward recovery over random querying and passive data collection.

Agentic-VLA introduces an agentic training framework that enables Vision-Language-Action (VLA) models to adapt efficiently online via three key innovations: Adaptive Reward Synthesis, Language-Guided Exploration, and Experience Memory. Evaluated on LIBERO benchmark, it achieves +12.3% on long-horizon tasks, +28.5% in 1-shot learning, and cross-task transfer from 0% to 31.2%. It also shows 2.4x faster convergence and retains advantages on the dual-arm RoboTwin 2.0 benchmark.

Event cameras are well suited for visual odometry under high-speed motion and challenging lighting conditions due to their low latency, high temporal resolution, and high dynamic range. Deep Event Visual Odometry (DEVO) demonstrated that monocular event-only odometry can achieve strong performance by combining sparse patch tracking, learned patch selection, recurrent correspondence refinement, and differentiable bundle adjustment. In this project, we extend DEVO with a sparse point-cloud export pipeline. Rather than modifying the core odometry formulation, our approach exposes the internal 3D structure already estimated by DEVO and converts it into an explicit point-cloud representation for visualization and further processing. In addition, we implement a practical workflow for data expor…

A systematic review of teleoperated endovascular intervention robots evaluated technical feasibility, communication infrastructure, and clinical outcomes. Sixteen studies out of 2501 were included. Teleoperated catheters and guidewires can navigate distances up to 7000 km with latency of 30-163 ms. Small-scale human trials show 100% procedural success, but most evidence is from animal or phantom models. The technology may reduce radiation exposure, expand patient access, and optimize resource allocation. Multi-center trials and studies in low/middle-income countries are needed.

Researchers develop a millimeter-wave radar system to measure body shape through clothing, preserving privacy and enabling fast, contactless assessments for clinical risk stratification.

A novel approach using motion as the central modality for video representation, training a masked autoencoder on point-tracks in a self-supervised manner. The resulting TIME embedding, trained solely on synthetic motion data, achieves performance on par with state-of-the-art models using up to 4 orders of magnitude less data.

CoMoGen is a controllable video generation framework that generates realistic interactive dynamics from a single binary mask sequence conditioned on an input image. It introduces a lightweight MaskAdapter to encode mask sequences into latent residual signals injected into a Multi Modal Diffusion Transformer (MMDiT) via a cosine-weighted schedule. By identifying 'Motion Layers' in the attention space of MMDiT and fine-tuning only those layers with Low-Rank Adaptation (LoRA), CoMoGen reduces computational cost without architectural changes. Experiments show state-of-the-art performance in motion fidelity and perceptual realism.

Video recordings of child-caregiver interactions allow study of attentional dynamics, but manual annotation is time-consuming. GBAT is a deep-learning toolkit that automates synchronization, gaze annotation, and pose/hand action categorization, enhancing efficiency for large-scale developmental research.

This paper presents a lightweight modification to the DETR-based fusion transformer baseline for the MaCVi 2026 Vision-to-Chart data association challenge. A dedicated MLP (QueryMLP) is trained to explicitly predict the buoy's waterline contact point in the image from chart measurements and IMU orientation data. The predicted pixel coordinates are appended to the baseline decoder query vector, providing a direct spatial prior per buoy and reducing the geometric reasoning burden on the transformer decoder. The approach achieves an Overall score of 0.7386, F1=0.8055, and mIoU=0.6718 on the held-out test set, placing second among all submissions.

VideoOdyssey is a benchmark for ultra-long-context and omni-modal video understanding, featuring videos averaging 109 minutes across 11 domains and 54 subcategories. It measures cognitive load via continuous certificate length and offers five granular levels. Evaluations show current MLLMs struggle with continuous reasoning, fine-grained perception, and non-verbal omni-modal understanding.

This study challenges the assumption that high benchmark scores reflect true visual understanding in vision-language models (VLMs). By removing a large fraction of image tokens with minimal performance drop, the authors reveal a mismatch between accuracy and visual grounding. Through multi-level analyses including global degradation, localized occlusion, question reformulation, answer-space expansion, decision-level analysis, and layer-wise vision-token geometry, they find that models are less sensitive to fine-grained visual evidence than expected, and that visual tokens become more similar in deeper layers. The results indicate that current benchmarks are insufficient for evaluating fine-grained visual grounding in VLMs.

GEM-4D is a geometry-grounded video world model that improves robot manipulation by injecting dense 4D correspondence supervision distilled from a pretrained geometry foundation model. It jointly captures appearance and geometric structure without additional inference cost. An inverse dynamics module converts consistent video rollouts into executable robot trajectories. GEM-4D achieves state-of-the-art performance on video prediction and geometric consistency, boosting real-world manipulation success from 61% to 81%.

A study finds that large language models (LLMs) exhibit persistent asymmetries when answering questions about religious conversion. Models tend to support joining Catholicism, Bahá'í, and Sikhism while subtly discouraging leaving these faiths, and show the opposite for atheists, agnostics, and Jehovah's Witnesses. The study tested 20 models across 182 religious pairings, with reproducible results.

This study evaluates seven LLMs (including Gemini, Claude, and GPT families) on inferring individual domain knowledge from long-term Slack logs. Using 27,188 messages from 43 users, zero-shot estimates were compared with self-reported skill ratings from 27 participants. Gemini 2.5 Flash achieved the lowest error (MAE 21.13%), while GPT models showed larger discrepancies. Accuracy depends weakly on message volume, highlighting limits and the need for privacy-aware deployments and richer knowledge representations.

Large Language Models (LLMs) are optimized to produce distributionally plausible continuations rather than to explicitly verify whether generated propositions are entailed by source documents. This inductive bias enables generalization, but it does not encode whether responses are grounded with respect to a reference. Existing hallucination detection approaches improve factuality through retrieval augmentation, self-consistency, or claim verification, but generally do not learn directly over alignment topology. To leverage alignment topology as an inductive bias, researchers construct aligned bipartite graphs between reference information and LLM outputs and train a graph neural network (GNN) to model alignment structure using message passing. The method achieves state-of-the-art results…

To improve reasoning in diffusion language models (DLMs), researchers propose LIFT, a fine-tuning algorithm that adapts to token learnability across diffusion steps, outperforming baselines on six benchmarks with up to 3x relative gains on AIME'24 and AIME'25.

This study introduces a red-teaming framework to measure the Overton Window (range of political opinions) of open-source LLMs on controversial topics, and quantifies how simple natural-language jailbreaks expand that range. Key findings include a systematic left-leaning bias, inverse relationship between model size and expressivity, significant regional differences, and varying jailbreak potency across model families. The framework aids auditing political steerability and designing countermeasures.

This paper proposes a knowledge-aware Text-to-SQL framework that constructs a task-specific knowledge base including schema semantics, abbreviations, business logic, and query patterns, and injects them into both training and inference. Experiments on seven benchmarks demonstrate substantial performance improvements for both open-source and closed-source large language models, especially in low-resource domain-specific settings.

This survey systematically catalogs publicly available text and speech resources for Hausa (80-100 million speakers) and Fongbe (2 million speakers in Benin). Findings show Hausa has broader text resource diversity across news, encyclopedic, and educational domains, while Fongbe has seen recent academic speech data collection initiatives. Both languages are represented in Masakhane benchmarks for NER and POS tagging. The paper provides task-specific recommendations and identifies priority gaps including domain-diverse Fongbe text and dedicated Hausa speech corpora.

Tensor Cache is a two-level cache pairing sliding-window softmax attention (L1) with a fixed-size outer-product fast-weight memory (L2) fed by KV pairs evicted from the window. It enables long-context access beyond the window with minimal memory overhead, improving the memory-quality frontier.

Existing neural solvers for multi-objective combinatorial optimization problems (MOCOPs) suffer from limited weight-conditioned context modeling and inefficient training due to random sampling in preference optimization. WeCon introduces Gated Residual Fusion (GRF) in the encoder and Residual Fusion (RF) in the decoder to enhance weight-instance interaction, along with Efficient Preference Optimization (EPO) for higher-quality training pairs. Experiments show WeCon achieves comparable HyperVolume (HV) to state-of-the-art POCCO-W while reducing inference time by 40%.

Research shows chain-of-thought reasoning is not always beneficial; early entropy dynamics can indicate when reasoning helps. The authors propose EDRM, an adaptive routing framework that uses entropy trajectories, achieving 41-55% token reduction across 15 benchmarks while improving accuracy.

MedExpMem is a novel experience memory framework that enables medical vision-language models to accumulate differential diagnosis expertise from diagnostic failures, achieving up to 7% accuracy improvement across 11 radiology subspecialties.

This paper introduces ManiF-SMC, a method for approximate machine unlearning that pushes erased samples away from their learned manifold centroids towards semantic neighbors in retained data, operating purely in representation space. A self-mode-connectivity module adaptively generates margins for triplet loss, achieving state-of-the-art unlearning effectiveness.

Research reveals that small language models rely on a positional shortcut—copying the last number before the answer delimiter—rather than logical reasoning when performing arithmetic with chain-of-thought (CoT) prompting. This copying accounts for 89-92% of model accuracy and overrides genuine computation. The findings challenge the faithfulness of CoT-based oversight.

FuRA is a novel full-rank parameter-efficient fine-tuning method that preserves pretrained robust features via spectral preconditioning, outperforming full fine-tuning and LoRA on LLM and VLM fine-tuning; its 4-bit quantized variant QFuRA also surpasses QLoRA.

FusionSense is a fusion-aware intelligent sensing framework for energy-constrained autonomous edge systems. It uses a three-step training procedure to create lightweight near-sensor classifiers that jointly reduce compute and communication while scaling linearly with sensor count. On a SynDrone dual-modality setup, it achieves up to 33x lower energy at 1% FoI prevalence, 92.3% reduction in quality loss at 30% data reduction, and 1.5x higher energy savings than prior baselines.

A new method uses geometric features from layer-wise MLP updates to train a sparse linear probe for better uncertainty quantification in language models, outperforming standard MSP by up to 21 AURC points on selective abstention.