A passive distributed paradigm using deep sub-wavelength acoustic waveguides decouples performance from structural flexibility, achieving high-resolution tactile sensing with sparse microphone arrays.

A novel occupancy-grounded pipeline for hierarchical 3D scene graphs anchors room nodes to tracked free-space regions from occupancy decomposition, providing explicit polygonal footprints. Evaluated on 12 Matterport3D scenes, it recovers more room instances than the state-of-the-art Hydra method but at lower precision, and wall-accurate boundaries remain an open challenge.

Avatar V is a production-scale framework for generating avatar videos that are behaviorally recognizable. It conditions on full token sequences of reference videos, using Sparse Reference Attention, motion representation, and identity-aware super-resolution. Trained on over 100M clips with a five-stage pipeline, it achieves state-of-the-art results on cross-scene benchmarks, outperforming existing systems.

Vision-language models (VLMs) confidently answer image-based questions even without images, inflating benchmark scores without reflecting visual grounding. This study classifies mirages into two regimes: textual biases (relying on language priors) and spurious images (constructing false visual content in latent space). Using a contrastive probing framework and a Prior Harnessing Index (PHI), the authors show these regimes are distinct and require different mitigation strategies—text cleaning addresses only the first, while faithful grounding needs representational interventions.

Researchers propose Temporal Backtracking Search (TBS) to overcome the single-shot bottleneck in generative video reasoning. TBS iteratively generates, verifies, and restarts along the temporal axis, significantly outperforming Best-of-N sampling across algorithmic, navigation, and robotics domains. In out-of-distribution settings, TBS achieves 22.7% success versus 0.7% for BoN.

This paper addresses the lack of explainability in rPPG Transformers by introducing a quantitative evaluation framework. The authors adapt four attribution methods to RhythmFormer's bi-level routing attention, propose a skin coverage metric, and extend the SaCo faithfulness coefficient to rPPG regression. They reveal a multi-hop leakage effect under sparse top-k routing and show that Beyond Intuition mitigates it, achieving the best skin coverage (0.83) and faithfulness (F=0.92) on UBFC-rPPG. The work advances rPPG XAI toward auditable numerical evidence.

Researchers propose i2L (image-to-LoRA), a framework that amortizes style LoRA training into a single forward pass, eliminating per-style optimization. Using an image encoder, learnable LoRA queries, and compressed decoding heads, i2L predicts LoRA weights and outperforms baselines on Z-Image, FLUX.2, and Hidream-O1 in style fidelity, prompt alignment, and perceptual quality.

CineOrchestra is a unified video diffusion model that simultaneously controls subjects, events, cameras, and shot transitions via entity-centric conditioning primitives and two parameter-free coordinated rotary embeddings. It outperforms six per-axis specialists on two new benchmarks.

This paper introduces learnable pairwise connections between filters in CNNs, moving beyond traditional pointwise nonlinearities. The proposed method allows the network to adapt connection functions per layer, improving accuracy.

IoU is a key metric for evaluating spatial alignment between proposals and ground truth, but it has an insensitive region where samples with distinct geometric overlaps yield similar scores. This work introduces morphological similarity metrics (area, shape, aspect ratio) to refine positive sample assignment, producing a supplementary matching score via mean aggregation. Experiments on YOLOv9 show consistent gains on NEUDET and GC10-DET datasets with zero extra inference overhead.

A new study explores how post-training methods like supervised fine-tuning and reinforcement learning can significantly improve generative LLMs for ICD coding, challenging the notion that LLMs are weak medical coders when evaluated solely via prompting.

DLawBench is a new diagnostic benchmark for evaluating LLMs in multi-turn legal consultation. It includes 461 cases from Chinese and U.S. law, categorizing interactions into Cooperative, Dependent, Withdrawn, and Adversarial types. Experiments show the best model, GPT-5.5, achieves only 0.562 on consultation-grounded legal reasoning, exposing sycophancy and a paradox where models perform worse when clients need guidance most.

This paper presents SANA (Search Agent Navigation Ablation framework), a diagnostic ablation framework that decomposes failures in exploratory question answering (EQA) over data lakes. By transforming EQA tasks into runtime profiles with gold source sequence, sanitized subquestions, and execution records, SANA constructs idealized search, planning, and data-analysis tools for component ablation, with residual gaps indicating policy failures. Experiments on LakeQA and KramaBench show data analysis as a consistent bottleneck, search as a major limitation in large data lakes, and planning as less critical.

Neural topic models enable scalable semantic discovery, but their integration with quantum hardware remains largely unexplored. This paper presents a proof-of-concept hybrid classical-quantum variational autoencoder (VAE) for topic modeling, embedding parameterized quantum circuits within the VAE inference network while retaining a classical topic-word decoder. To address quantum hardware resource constraints, a modified Gaussian Softmax posterior is proposed that decouples latent space dimensionality from the number of topics, enabling operation on a 10-qubit quantum device. On the AgNews dataset, the hybrid VAE outperforms state-of-the-art neural topic models, achieving a C_v coherence score of 0.71 and NPMI of 0.20 while preserving high topic diversity. A fully classical variant also outperforms SOTA models, showing clear class separation in latent space. These results demonstrate the computational viability of hybrid VAEs on NISQ-era devices and point to a promising direction for quantum-enhanced topic modeling.

This paper introduces a validation framework to evaluate whether LLM-based urban simulators produce empirically realistic human mobility patterns or merely plausible narratives. Testing AgentSociety and CitySim on datasets from Paris and Shanghai reveals a substantial gap between narrative plausibility and empirical realism, especially in spatial and temporal constraints. The authors also provide open-source tools for scalable evaluation.

Current cultural alignment approaches focus on inference-time interventions, assuming models already contain sufficient cultural knowledge. This paper argues modern LLM pipelines suffer from a 'cultural data funnel,' where explicit cultural signals decline sharply during post-training, while geographically concentrated, task-specialized data dominates. Multilinguality enhances geographic diversity but does not ensure balanced representation. The authors release a culturally tagged dataset with 5.6M samples to facilitate further research.

This paper overviews the QIAS 2026 shared task on Islamic inheritance reasoning, using the MAWARITH benchmark of 12,500 Arabic cases. 16 teams participated, and results show LLMs struggle with precise legal and numerical reasoning.

This paper presents the participation of team PSL in the QIAS 2026 Shared Task on Arabic Islamic inheritance reasoning. The task evaluates large language models on solving inheritance cases requiring legal interpretation, multi-step reasoning, and precise numerical computation. Results show commercial models like Gemini 2.5 Flash outperform open-source models in identifying heirs, applying exclusion rules, and maintaining consistency, while open-source models exhibit instability.

A new study introduces WebDecept, a framework for injecting deceptive interfaces to test web agent safety. Evaluations show current agents are highly susceptible to manipulations like targeted ads and domain redirection, and prompt-based defenses often fail.

A study on LLM-as-a-Judge reveals significant instability: pairwise preference flips 13.6% on average, with 28% of questions exceeding a 20% flip rate. GPT-4o-mini shows a first-position bias (72% A-majority). Cross-judge agreement is only 76%. The authors recommend multi-trial aggregation, position randomization, and explicit uncertainty reporting.

Research shows that trial-to-trial variability (structured noise) in cortical neural responses can significantly improve the robustness of artificial neural networks against adversarial attacks and naturalistic image modifications. The benefit is greatest for naturalistic modifications but transfers poorly across types, while noise structure from adversarial attacks can generalize to other attack types.

This paper identifies the double-drift phenomenon in diffusion policy optimization and proposes DiPOD, a framework that interleaves self-distillation with policy-improving gradient updates to maintain tight-bound behavior, stabilizing training and achieving higher rewards.

A new paper proposes Cheap LoRA (cLA) and a chained circulant variant c³LA, inducing sparsity in existing LoRA variants for parameter-efficient fine-tuning. Theoretical generalization error bounds are derived, and experiments across 11 fine-tuning methods, 10 pretrained models, and 14 datasets show that sparse structured column-space restriction remains competitive while reducing training time by up to 10% and peak GPU memory by up to 15%.

This paper presents D2H-AD, a novel anomaly detection framework based on Hyperdimensional Computing (HDC). It integrates distance-based similarity and density-aware encoding, outperforming existing baselines on five benchmark datasets with improved accuracy, interpretability, and computational efficiency, making it suitable for TinyML and edge AI deployments.

A new paper resolves the debate over whether weight norm causes grokking in neural networks by directly intervening on the norm during training. Under free training, grokking occurs when the norm reaches a critical value Wc that is highly consistent across seeds and learning rates. When the norm is clamped, delay follows an exponential law T_grok ∝ exp(αρ) with α≈7.5. LayerNorm eliminates this dependence.

Introducing a fully GPU workflow that integrates accelerated data generation with training of neural emulators augmented by uncertainty quantification and physics-aware refinement, leveraging differentiable solver JAX-Fluids for enhanced physical consistency in hypersonic flow prediction.

A new algorithmic pricing tool for fashion e-commerce sales campaigns uses daily-resolution demand forecasting and multi-objective optimization, achieving 6% higher profit in A/B tests at Zalando.

This paper presents llada.cpp, the first NPU-aware inference framework for accelerating diffusion LLMs on smartphones. It introduces three techniques—Multi-Block Speculative Decoding, Dual-Path Progressive Revision, and Swap-Optimized Memory Runtime—to overcome challenges like shrinking workloads, KV cache reuse, and memory overhead. llada.cpp reduces LLaDA-8B generation latency by 17x-42x over CPU baseline with no quality loss.

Yes, editing a single neuron can eliminate repetition loops in Gemma 4 models, but doom loops during extended reasoning persist due to missing facts, highlighting a knowledge precision problem.

In low-resource data silos, using local reference distributions for data selection can accelerate model collapse instead of preventing it, leading to power-law diversity decay. Collaborative Wasserstein proxy references, constructed without sharing raw data, mitigate diversity degradation.