This paper introduces HALO, a visuomotor policy with attention-based memory retrieval for long-horizon robot control, addressing spurious correlations and error accumulation in imitation learning.

The paper extends Parametric Control Barrier Function (Parametric-CBF) by embedding causality inference to explicitly reason over inter-vehicle influence, enabling an adaptive safety-critical controller that avoids overly conservative behavior and improves task efficiency in multi-vehicle interactions.

This paper proposes RGB, an RL-guided whole-body MPPI framework that uses a pretrained RL policy as a sampling prior and MPPI for online correction, achieving robust and precise humanoid control without retraining. Simulations on a Unitree G1 humanoid demonstrate stable 280Hz control and improved precision over pure RL.

AeroCast is a probabilistic trajectory prediction framework combining Transformer encoder with Mixture Density Network to predict Gaussian mixture distributions over future 3D displacements. It reduces error by 50% on a quadrotor corpus and runs at 0.1ms per sample.

A novel dataset and framework, SurveilNav, enables robots to collaborate with multi-view surveillance systems for object goal navigation. By integrating active camera scheduling, joint 2D/3D mapping, VLM-based value estimation, and collaborative target verification, it overcomes the limitations of single-robot perception and fixed-camera blind spots. Experiments on HM3D demonstrate state-of-the-art performance in exploration efficiency and navigation success rate.

Proposes ADM-Fusion, an end-to-end deep learning multi-sensor fusion method using an adaptive sensor mixture-of-experts framework with content-aware routing to dynamically weigh sensor inputs. It features separate translation and rotation branches coupled via cross-task attention. Trained on CARLA-LOC simulated dataset and fine-tuned on KITTI real-world data, it demonstrates robust performance under sensor degradation while matching state-of-the-art methods.

This paper introduces a novel invariant Kalman filtering approach for extended pose estimation in multi-IMU articulated rigid-body systems. By proposing a relative L-extended pose Lie group representation and incorporating joint kinematic constraints as noise-free pseudo-measurements within an iterated IEKF, the method achieves faster convergence and over 50% reduction in RMSE compared to existing filters on both a UR5e robot and a human leg.

A new method called Latent Sequence Optimization (LSO) enables precise physics-based motion tracking by optimizing over sequences of latents in Behavioral Foundation Models, validated on a real humanoid robot.

Most vision-language-action (VLA) models are reactive, predicting the next action from the current observation alone, which limits generalization under distribution shift. This paper proposes Reflective VLA, which conditions decisions on a context of observation-action-consequence triplets, exposing deployment-specific action-effect mappings. On LIBERO-Plus and LIBERO-Plus-Hard, it improves success rates by 5.4 and 4.2 percentage points, with ablations confirming action consequences as the key.

This paper proposes a novel neural network quantization method that learns quantization-aware linear paths to find midpoints in low-loss subspaces, achieving performance comparable to quantization-aware training without using the straight-through estimator or explicit discretization during training.

This study evaluates Multimodal Large Language Models (MLLMs) on assistive AI tasks including currency recognition, scene text QA, and multilingual reading. The authors built NetraLink, a system using a head-mounted GoPro to collect real-world egocentric data, and created a benchmark. Findings reveal strengths and limitations of current MLLMs for vision-language assistive technologies.

Visual storytelling requires image sequences aligned with narrative prompts and consistent characters. Existing training-free methods rely on structured prompts that repeat full descriptions each time, deviating from natural storytelling. FreeStory introduces entity-grounded feature reuse to maintain character consistency under free-form prompts, and presents FreeStoryBench, achieving state-of-the-art performance among training-free methods.

Wan-Streamer is a native-streaming, end-to-end interactive foundation model designed for real-time, low-latency, full-duplex audio-visual interaction. It seamlessly models language, audio, and video within a single Transformer using block-causal attention for incremental streaming, without external modules. It achieves ~200ms model-side and ~550ms total interaction latency, enabling sub-second duplex communication.

Chorus II introduces a cross-request sparsity reuse framework that reuses sparse attention masks from similar historical requests to avoid online mask prediction, with optional feature reuse and guidance enhancement, achieving a 2.16× speedup while preserving generation quality.

Yuvion VL is a family of multimodal large language models purpose-built for content and AI safety, treating safety as an inherently adversarial and multimodal problem. It features an automated data pipeline with adversarial-aware synthesis and multi-stage quality control, a three-stage training pipeline including continued pretraining for cross-modal alignment, instruction post-training, and reasoning post-training, plus a novel Confuse-then-Contrast Fine-Tuning framework. The YVRE benchmark set evaluates safety, adversarial robustness, and real-world capabilities. Yuvion VL-32B achieves industry-leading safety performance, surpassing open-source and closed-source models while maintaining general capabilities.

This paper proposes a Noise-Aware Boundary-Enhanced Generative Learning (NBGL) framework for ultrasound speckle reduction. It integrates a speckle reduction branch and a boundary enhancement branch, with a noise-aware interaction weight generation module that uses 3D Laplacian filtering and median absolute deviation estimation to adapt to varying noise levels. Evaluated on 141 3D transvaginal ultrasound volumes, NBGL outperforms state-of-the-art methods across six noise levels.

This study developed a codebook for self-stigma across cognitive, affective, and behavioral domains and analyzed Reddit posts from people who use drugs. Results show that self-stigma is prevalent, and behavioral indicators often precede core ones, challenging progressive models.

Large language models have transformed code generation, raising concerns about authorship, assessment integrity, and software trust. SemEval-2026 Task 13 Subtask A defines detection as binary classification over code snippets, with emphasis on out-of-distribution generalization across unseen programming languages and domains. The authors propose SALSA (Single-pass Autoregressive LLM Structured Classification), which maps each class to a dedicated output token and trains the model to emit a single-token label. By combining balanced sampling, parameter-efficient fine-tuning, and conservative training, the system achieves OOD F1=0.789 on the official leaderboard, significantly outperforming the CodeBERT baseline (F1=0.305).

The rapid growth of scientific submissions has pushed traditional peer review toward its scalability limits, motivating the exploration of large language models (LLMs) as intelligent automated evaluation assistants. Although recent studies show that LLMs can generate fluent critiques and approximate reviewer scores, their reliability, robustness, and security as decision-support systems remain insufficiently understood. This survey offers a systems-level analysis of LLM-based scientific peer review, focusing on two core evaluative functions: critique generation and score prediction. We present a structured taxonomy of modeling approaches (including prompt-based, supervised, retrieval-augmented, and alignment-optimized approaches), and synthesize empirical findings across existing benchmar…

A new study introduces a dataset of 32,534 double-marked real student responses to GCSE mock exams, covering five subjects and handwritten work. Top LLMs agree with examiners more closely than examiners agree with each other, handling subjective and handwriting tasks effectively, with little dependence on model size.

Dustin is a sparse verification framework for long-context speculative decoding that combines draft model lookahead signals with target model historical attention to identify critical tokens, achieving 27.85x self-attention speedup and 9.17x end-to-end decoding speedup at 32k sequence length on Qwen2.5-72B with negligible accuracy loss.

A new arXiv paper investigates the geometric relationship between detection and control directions in language models. While models can perfectly detect hallucination (AUC=1.0), the direction for detection and the direction for causing refusal have a cosine of only 0.12, indicating that detection does not imply controllability. This gap generalizes across models and sizes, originates in pretraining, and a 15-degree rotation can partially bridge it.

A new framework using error-aware TF-IDF retrieval to correct ASR errors, achieving significant improvements in WER on Persian FLEURS.

AgentOdyssey is a novel evaluation framework that procedurally generates open-ended text games to test agents' ability to learn continuously during deployment. It challenges the traditional ML assumption of no learning at test time, interleaving learning and inference throughout. The framework measures world knowledge acquisition, episodic memory, exploration, action diversity, and model cost. Experiments show even the strongest agents fall far below human performance, with short-term memory emerging as a key beneficial mechanism.

A new study shows that a small group of Wikipedia editors can significantly influence how large language models discuss animal welfare with just 125 edits. Using gradient-based attribution methods, the research traced the impact of these edits, finding that animal welfare-related Wikipedia content dominates model responses to relevant queries.

Researchers propose G-SPIN, a structured ASR correction framework that combines phonetic graph modeling with contextual language understanding. It uses a graph neural network to generate acoustically plausible candidate sets, a masked language model for scoring, and an instruction-tuned large language model for final re-ranking, enabling lightweight, modular inference-time correction.

MacroLens is a new multi-task benchmark covering 4,416 U.S. small- and micro-cap equities over 2021-2026, integrating prices, accounting data, macroeconomic series, SEC filings, and news. It addresses four key assumption violations in financial time-series evaluation, includes seven tasks and 1,130 macroeconomic events, and evaluates 19 methods with a five-step feature-context ablation. The benchmark is publicly available on Hugging Face.

A study finds that holographic memory models fail at zero-shot compositional queries in knowledge graphs due to capacity and interference effects, not the binding algebra.

A new paper proposes a Supervised Reinforcement Learning (SRL) framework for coordinating distributed energy resources (DERs). The approach pre-trains policies on demonstration data, then fine-tunes with offline and online RL, outperforming benchmarks even with low-quality data.

Learned world models are useful only over horizons on which their rollout error remains controlled. This paper studies trust-horizon certification for latent world models with known group symmetries. Using split-conformal calibration, the authors show that exact equivariance transports the calibrated trust-horizon curve over the group orbit, making rollout errors and trust horizons orbit-constant. Experiments on 2D and 3D tasks demonstrate that equivariant models achieve safe and non-vacuous orbit-valid certificates from a single calibration sector, while non-equivariant baselines incur additional costs. The certificate is a conservative distributional audit, not a global reachability guarantee.