A new preprint audits data leakage in whole-slide image (WSI) multimodal benchmarks and finds that patient-level and institutional-level overlaps significantly compromise reported zero-shot performance of vision-language models. The authors document case-level train-test overlaps of 92.3–100% on TCGA-derived benchmarks and show that this leakage is linearly decodable from foundation-model features. They recommend concrete steps for contamination-free evaluation practices.
Why it matters: This work reveals that widely-used WSI VQA benchmarks may not accurately measure genuine multimodal reasoning, calling into question reported advances in computational pathology.
A new framework called DM-KG introduces a structured knowledge graph to enhance spatial reasoning in vision-language models (VLMs) applied to street view imagery. By extracting and encoding directional and metric relationships between entities from 2D images, DM-KG significantly reduces distance estimation and direction judgment errors on spatial question-answering benchmarks.
Why it matters: This work offers a notable advance in improving the spatial cognition of VLMs, addressing a key limitation for their deployment in geospatial and geographic visual question answering tasks.
A new evaluation framework uses four vision-language models (VLMs) to assess EEG-to-image reconstructions, introducing Tolerant Perceptual Alignment Scores (T-PAS) and Tolerant Semantic Alignment Scores (T-SAS). The distilled BCI-Coherence Score (BCS) demonstrates lower mean absolute error and higher correlation with human judgments than traditional pixel or representation-based metrics, addressing the challenge of distinguishing visual fidelity from semantic recoverability.
Why it matters: This framework offers a more meaningful way to evaluate EEG-to-image reconstructions, potentially advancing brain-computer interface research by better capturing both perceptual and semantic accuracy.
A new preprint introduces DeGuNet, an ultra-compact image backbone designed for LiDAR-camera 3D detection in autonomous driving. DeGuNet uses depth-guided representation learning to address parameter redundancy in current multi-modal frameworks, reducing GPU memory usage by up to 66.5% and achieving a 1.16x speedup. The method also improves mean average precision (mAP) by up to 6.20 points on the nuScenes dataset, demonstrating both efficiency and accuracy gains.
Why it matters: DeGuNet could enable more efficient and accurate 3D perception for autonomous vehicles by significantly reducing computational demands.
A new preprint systematically evaluates continual learning (CL) methods for medical visual question answering (MedVQA) across a range of clinical tasks, including classification, detection, cell counting, and report generation. The study finds that current CL methods have difficulty maintaining a balance between retaining old knowledge and learning new tasks when faced with diverse objectives and supervision formats. The authors also analyze the impact of task ordering and the evolution of model parameters during continual learning. Code and experimental setup will be made publicly available.
Why it matters: This work exposes key limitations in current continual learning approaches for medical VQA, highlighting challenges that must be addressed for robust real-world deployment.
A recent audit of inference-time defenses for multimodal large language models (MLLMs) found that three benchmark branches failed provenance checks, restricting reliable comparative results to only two datasets. The study also discovered that a legacy keyword-based safety protocol incorrectly counted empty strings as safe, with raw model responses unavailable to reassess this effect. Contrary to earlier claims, the archive does not support widespread model refusal, with a pooled refusal rate of just 0.52%.
Why it matters: This work exposes critical methodological issues in evaluating MLLM safety defenses, emphasizing the importance of traceable audits and provenance standards for trustworthy comparisons.
KnowAct-GUIClaw is a new framework for GUI automation that addresses limitations in cross-platform support and self-evolution found in previous systems like OpenClaw. The framework introduces a Know-Route-Act-Reflect paradigm, enabling the agent to accumulate user interaction experience and improve execution accuracy over time. In experiments, KnowAct-GUIClaw using the open-source Kimi-2.6 model achieved 64.1% on the MobileWorld benchmark, outperforming both open and closed-source alternatives such as Seed-2.0-Pro and GPT-5.5.
Why it matters: This work represents a notable advance in personal GUI assistants by enabling self-evolving memory and skill transferability across platforms and models, potentially improving automation efficiency and adaptability.
CityBehavEx is an interactive urban simulation platform that integrates large language models (LLMs) with traditional human mobility models to efficiently simulate city-scale populations. The system can run simulations of 100,000 agents over 75 days in under one hour on a single consumer GPU, and supports empirical validation by comparing generated mobility patterns to real-world spatial, temporal, and semantic distributions. CityBehavEx also provides tools for inspecting agent behavior, debugging, and validating routines against real-world data, addressing scalability and validation challenges in prior LLM-based simulators.
Why it matters: This work offers a scalable and empirically validated approach to LLM-driven urban simulation, potentially advancing applications in urban planning and policy analysis.
A new neuro-symbolic approach integrates a Maximum Satisfiability (MaxSAT) oracle with Vision-Language Models (VLMs) to improve their performance on Sudoku puzzles. The MaxSAT solver identifies consistent subsets of VLM-generated assignments and provides structured feedback, guiding the models to refine their solutions. Experiments across multiple VLMs show that this method enhances logical consistency and increases the number of correctly solved Sudoku instances.
Why it matters: This work shows that symbolic optimization can significantly improve the reliability of vision-language models on structured reasoning tasks.
Researchers introduce Light-MER, a multimodal emotion recognition model with fewer than 1 billion parameters, designed to achieve high performance through knowledge distillation from larger models. The framework incorporates optimal transport loss and a multi-reward optimization strategy to balance accuracy and efficiency. Experiments on nine benchmark datasets show that Light-MER achieves state-of-the-art results while enabling faster inference, making it suitable for deployment on resource-constrained devices.
Why it matters: This work demonstrates that high-quality multimodal emotion recognition is possible with much smaller models, enabling practical deployment on devices with limited computational resources.
A new preprint introduces Visual Access Sweep, a causal intervention technique to study how Vision-Language Models (VLMs) use image information during Chain-of-Thought (CoT) reasoning. The study finds that VLMs do not require ongoing access to image tokens throughout CoT reasoning; instead, CoT mainly extends language-side computation over previously extracted image-derived hidden states. The research also shows that the effectiveness of CoT is limited by the model's ability to reliably read out perceptual attributes, rather than by its counting ability.
Why it matters: This work challenges common assumptions about CoT in VLMs and suggests that improving perceptual readout, rather than extending visual access, may be key to further performance gains.
Researchers introduce FormalAnalyticGeo, a scalable framework that leverages formal languages and four specialized LLM components to automatically generate multimodal analytic geometry problems with precise diagrams. The system produces the AnalyticGeo7K dataset, comprising over 7,000 verified problems, and achieves a median ground-truth relative error of 0.70%.
Why it matters: By enabling fully automatic generation of high-quality multimodal analytic geometry problems, this framework addresses the scarcity of annotated data and could accelerate research in math reasoning for multimodal large language models.
Researchers introduced LMVQA, a large language model-based multimodal question-answering system designed for technical meeting videos containing diagram-rich content such as UML diagrams. LMVQA significantly improved answer accuracy over a state-of-the-art baseline, from 31% to 94% on an industrial dataset and from 21% to 88% on a public dataset. The system also reduced average response time and LLM API costs after a one-time video indexing process. Interviews with domain experts highlighted LMVQA's value for locating relevant information and tracing rationale in software engineering meetings.
Why it matters: LMVQA addresses the challenge of efficiently retrieving knowledge from complex, diagram-rich meeting recordings, which is crucial for software engineering teams relying on asynchronous communication.
Researchers introduce the CREATE-IF-LATE (CIL) algorithm to address 'Pac-Man' attacks in decentralized learning, where malicious nodes can terminate random walks and disrupt learning. The CIL algorithm enables self-creating random walks, ensuring that the random walk population does not go extinct, remains bounded, and that stochastic gradient descent converges with a quantifiable deviation even under attack. Empirical results on synthetic and benchmark datasets support the theoretical guarantees.
Why it matters: This work provides a novel, decentralized defense against a stealthy adversarial threat in distributed learning systems, helping to maintain robust learning progress without centralized control.
A new preprint introduces RMATS, a recursive multi-agent trading system composed of four specialized agents coordinated by a manager. In experiments spanning 561 trading days and multiple asset classes, RMATS achieved a maximum drawdown of 9.62%, outperforming traditional methods in downside protection during several geopolitical stress scenarios. Ablation studies indicate that each agent contributes to risk control, though RMATS underperforms return-maximizing baselines in bull markets.
Why it matters: This work presents a novel multi-agent architecture that advances risk-controlled portfolio management, particularly relevant for capital preservation during periods of geopolitical uncertainty.
Researchers have introduced an explicit multimodal routing framework for clinical prediction using electronic health record (EHR) data, allowing for interpretable reasoning across structured variables, clinical notes, and chest X-rays. The model constructs discrete unimodal, bimodal, and trimodal routes, and uses inference-time route masking to audit the contribution of each modality and assess robustness without retraining. Evaluated on phenotype and mortality prediction tasks with MIMIC-IV data, the framework reveals systematic differences in modality reliance across clinical conditions.
Why it matters: This work advances interpretability and robustness in AI-driven clinical decision support by providing a transparent method to understand how different data modalities influence predictions.
Researchers applied the MAPPO multi-agent reinforcement learning algorithm to the problem of selecting between cellular Uu, NR-V2X PC5 sidelink, or both for vehicle-to-everything (V2X) communication. In urban scenario simulations, this approach improved the on-time delivery ratio from 0.508 to 0.535 in single-vehicle settings and from 0.548 to 0.567 when all vehicles used the learned policy, while halving training time compared to a deep reinforcement learning baseline. The improvements were most pronounced for advanced V2X applications such as cooperative driving and shared perception.
Why it matters: This work shows that multi-agent reinforcement learning can effectively manage hybrid V2X communication to meet diverse latency and reliability needs, which is important for future autonomous and cooperative vehicle systems.
Researchers introduce SaMer, an object-aware token merging framework for multi-vector vision-language retrieval. SaMer compresses image-side tokens by over 93% while improving retrieval accuracy on benchmarks like Flickr30K and MSCOCO. The method preserves object-level evidence needed for effective retrieval, outperforming existing compression baselines and enhancing phrase-level grounding.
Why it matters: This work demonstrates that preserving object evidence, rather than merely reducing token count, is crucial for efficient and accurate multi-vector vision-language retrieval, enabling substantial storage and computation savings.
Researchers introduce MG²-RAG, a lightweight framework that constructs a hierarchical multimodal knowledge graph by combining textual parsing with entity-driven visual grounding, enabling unified multimodal nodes. The framework features a multi-granularity graph retrieval mechanism that supports structured multi-hop reasoning and aggregates dense similarities across the graph. MG²-RAG achieves state-of-the-art performance across four multimodal tasks, while reducing graph construction overhead with significant speedup and cost reduction compared to prior graph-based methods.
Why it matters: MG²-RAG advances multimodal retrieval-augmented generation by enabling efficient, fine-grained cross-modal reasoning without relying on costly translation-to-text pipelines.
Researchers have developed a foundation transformer model pretrained on multimodal sequences of user events for financial services. By unifying heterogeneous data sources into chronological event sequences and using next-event prediction, the model learns general-purpose representations that can be applied to multiple downstream tasks. The system outperformed traditional task-specific models and was deployed in production at a major Eastern European bank, where it led to measurable improvements in business metrics.
Why it matters: This work demonstrates a practical and effective foundation model for financial event sequences, showing real-world impact through improved predictive performance and reduced development overhead in a production banking environment.