Vision-language models, speech and cross-modal systems that read, look and listen in the same forward pass.
We introduce VGA-BenchV2, an extended human-aligned benchmark and optimization framework for jointly evaluating and improving video generation quality and aesthetic value. Built upon VGA-Bench, VGA-Be...
Multi-vector representations have emerged as an effective paradigm for multimodal retrieval, representing each sample with multiple complementary embeddings to capture fine-grained cross-modal informa...
Multimodal Large Language Models (MLLMs) have demonstrated strong abilities in solving diverse visual and textual reasoning tasks. However, their development in the physics domain is significantly hin...
Vision-language models (VLMs) have made substantial progress in long-video understanding, with standard backbone models typically answering questions from frames sampled across the full video. However...
The paper introduces DoublesEval, a diagnostic framework that uses professional doubles badminton to test visual‑language models’ ability to reason about dynamic multi‑agent interactions. It decomposes rallies into key moments and evaluates models across four dimensions—atomic recognition, intra‑segment composite understanding, cross‑segment causal reasoning, and high‑level tactical abstraction—highlighting specific reasoning failures. The authors also propose TacticCheck, a lightweight consistency checker that improves performance without retraining the models, yet significant gaps remain in tactical reasoning.
VideoHarness‑RSI explores how improving the executable context‑construction program alone can enhance long‑video understanding with frozen vision‑language models. By recursively searching for better harnesses—programs that select and structure video segments—using an outer‑loop proposer that learns from prior programs and execution traces, the method consistently outperforms weaker hand‑crafted baselines and further improves upon stronger ones. The resulting harnesses transfer to other long‑video benchmarks without additional search, demonstrating that executable context construction is a distinct, reusable optimization layer.
ForceFlow is a force-aware reactive framework that uses flow matching to improve contact-rich manipulation. It fuses force signals asymmetrically, treats force as a global regulator, and employs a joint prediction paradigm to couple force and motion. The approach splits tasks into a vision-dominant localization stage and a touch-dominant execution stage, using a Vision-to-Force handover to separate spatial generalization from contact regulation.
The paper introduces ICA, an evidence‑centric framework that represents information from web‑tool interactions as stable, rendered snapshots, enabling comparison across trajectories. It proposes Information‑Aware Credit Assignment, a post‑hoc reward propagation technique that estimates turn‑level utility from rollout success and assigns dense rewards to steps that provide high‑utility information. When combined with GSPO, ICA consistently improves performance on several web‑search benchmarks such as BrowseComp, GAIA, Xbench‑DS, and Seal‑0.
The paper introduces CAIT, a benchmark of 400 synthetic scenes featuring counter‑intuitive actions that challenge multimodal large language models (MLLMs). Human participants and proprietary models like Claude and Gemini perform well, but standard open‑source instruction‑tuned MLLMs fail, largely due to a strong language prior that overrides contradictory visual evidence. The study shows that Chain‑of‑Thought reasoning can help but introduces new issues, while targeted fine‑tuning and structured prompting can reduce reliance on language priors and improve visual grounding.
The paper introduces a new evaluation protocol for deep multivariate imputation models on wearable device data, addressing the issue of structured missingness where sensor features drop out together. Using a Garmin smartwatch dataset from an epilepsy patient, the authors generate realistic block-missing patterns from training data and show that matching the training protocol to this distribution reduces BRITS’ mean absolute error by 43%. They also extend BRITS with time‑of‑day encoding and compare it to linear interpolation and SAITS, finding that no single model dominates and that model rankings vary with evaluation design.
The paper tackles the challenge of predicting student engagement from online tutoring videos, noting that engagement is a complex, multidimensional construct influenced by behavioral, emotional, and cognitive states. By analyzing the CASED dataset, the authors highlight the difficulty posed by high inter‑person variability and subjective annotations. They propose a multimodal framework that fuses implicit spatiotemporal features from pretrained video, audio, and image encoders with structured behavioral cues such as head pose, gaze, facial action units, emotion, and wavelet‑based audio features, integrating them via a Perceiver IO bottleneck and modeling participant personalities with variational posteriors. The system employs evidential regression and spectral‑normalized Gaussian process classification heads to provide uncertainty‑aware predictions, achieving competitive performance on the CASED challenge test set while offering well‑calibrated uncertainty metrics.
PlaceSeek is a human‑centered geospatial retrieval framework that maps natural‑language queries to street‑view images by decomposing queries into functional and affective sub‑intents. It uses a Semantic Grounding Module to verify that candidate images contain the physical evidence needed for the intended activity, and an Affective Alignment Module to re‑rank these candidates based on human urban perception judgments. Evaluated on 31,956 Milan street‑view locations, PlaceSeek achieves high precision and ranking metrics, outperforming several vision‑language baselines and demonstrating the importance of both physical grounding and affective alignment for complex urban spatial queries.
The paper documents the migration of a live conversational recommendation system from a gradient‑boosted multiclass model to a pairwise‑binary deep recommender. It explains how reformulating the task, using negative sampling, noise injection, and attention pooling over transcript chunks enabled the new model to handle dynamic, multimodal data and long conversation context. The authors compare several architectures and loss functions, showing that the deep recommender matches or surpasses the CatBoost baseline, especially in later conversational stages.
NeuronGuard is a fine‑tuning defense for large language models that hardens them against both jailbreak and neuron‑level attacks. It redistributes safety signals across many neurons by identifying safety‑critical ones with per‑layer linear classifiers, enforcing refusal behavior when those neurons are ablated, and applying KL‑divergence regularization for consistency. A randomized gradient projection preserves task performance, and the authors provide a formal guarantee that NeuronGuard lowers the attack success rate upper bound, with experiments showing near‑zero success rates across multiple models and attack strategies.
arXiv:2602.19001v2 Announce Type: replace Abstract: As large language models increasingly power personal assistants, users expect them to reason over multimodal life histories, from recognizing peopl...
PhysMLLMs introduces physics-inspired spatial continuity priors into video multimodal large language models to address spatio‑temporal inconsistencies such as jitter, drift, and identity switches. The method, called Global Representation Prior Alignment (REPA‑Global), distills global visual representations from a frozen DINOv2 teacher during training, aligning student representations without affecting inference speed. Experiments on multiple video benchmarks show improved segmentation mask quality and cross‑frame consistency, especially for challenging scenarios involving small targets, fast motion, occlusion, and distractors, while maintaining comparable performance on single‑frame image segmentation and general VLM tasks.
HMGCLIP is a unified multimodal embedding framework that uses a heterogeneous hypergraph to capture both fine‑grained and coarse‑grained product attributes. By mining structure‑aware hard negatives and aligning multi‑granular semantics at relation and hyperedge levels, it enables a dual‑granularity inference mechanism that dynamically fuses attribute evidence. Experiments on a new fine‑grained e‑commerce dataset and the public MAVE benchmark show that HMGCLIP outperforms strong multimodal encoders, MLLMs, and e‑commerce baselines.
SonarLLM is a multimodal large language model that treats sonar as a native perceptual modality, combining a sonar‑specific encoder, physics‑aware feature enhancement, and reliability‑aware hierarchical fusion to align acoustic structure with optical semantics. The authors introduce SonarBench, a benchmark covering recognition, counting, visual question answering, and captioning across sonar‑only, optical‑only, and fusion settings, enabling controlled measurement of cross‑modal complementarity. SonarLLM achieves 72.0% macro accuracy on sonar‑only tasks and 68.7% under fusion, outperforming baselines by significant margins and demonstrating increasing fusion gains as optical visibility degrades.
The paper introduces KnowChange, a framework that uses pretrained vision‑language models to guide the synthesis of change data for remote sensing. By reasoning about plausible change locations and class transitions, KnowChange flexibly generates diverse change types within a unified pipeline. Experiments show that data produced by KnowChange outperforms existing synthetic datasets in both synthetic‑to‑real transfer and data augmentation scenarios, even at a compact scale.
The paper introduces D3-Omni, a balanced and decoupled benchmark designed to diagnose fine‑grained multimodal understanding in OmniJudges that evaluate text‑to‑image, text‑to‑video, and text‑to‑speech generation. D3-Omni covers 53 orthogonal binary dimensions across 10,671 samples, using fixed positive seeds and controlled prompt rewriting to generate negatives, thereby ensuring each error can be attributed to a single capability. The benchmark’s dual‑balanced, decoupled, and dynamic design achieves near 1:1 per‑dimension parity and a uniform total‑score distribution, revealing that strong OmniJudges often miss modality‑related failures and treat distinct attributes as a single decision, masking systematic blind spots.