Topics ›
50 articles RSS ↗
A new approach to composed video retrieval (CoVR) is presented, which leverages large multimodal models to infer causal and temporal consequences implied by an edit. The method aligns reasoned queries to candidate videos without task-specific finetuning. A new benchmark, CoVR-Reason, is introduced to evaluate reasoning in CoVR.
The paper introduces SectEval, a new benchmark to evaluate sectarian biases in LLMs concerning Sunni and Shia Islam, available in English and Hindi. Results show significant inconsistencies in LLM responses based on language, with some models favoring Shia responses in English but Sunni in Hindi. Location-based experiments further reveal that advanced models adapt their responses based on the user's claimed country, while smaller models exhibit a consistent Sunni-leaning bias.
MBZUAI researchers have developed an automatic interview system that uses LLMs to elicit nuanced, role-specific information from job candidates, improving early-stage hiring decisions. The system updates its belief about an applicant's rubric-oriented latent traits in a calibrated way based on their interview performance. Evaluation on simulated interviews showed the system's belief converges towards the simulated applicants' constructed ability levels.
The paper introduces Yet another Policy Optimization (YaPO), a reference-free method for learning sparse steering vectors in the latent space of a Sparse Autoencoder (SAE) to steer LLMs. By optimizing sparse codes, YaPO produces disentangled, interpretable, and efficient steering directions. Experiments show YaPO converges faster, achieves stronger performance, exhibits improved training stability and preserves general knowledge compared to dense steering baselines.
The paper introduces the Prism Hypothesis, which posits a correspondence between an encoder's feature spectrum and its functional role, with semantic encoders capturing low-frequency components and pixel encoders retaining high-frequency information. Based on this, the authors propose Unified Autoencoding (UAE), a model that harmonizes semantic structure and pixel details using a frequency-band modulator. Experiments on ImageNet and MS-COCO demonstrate that UAE effectively unifies semantic abstraction and pixel-level fidelity, achieving state-of-the-art performance.
A new benchmark, LongShOTBench, is introduced for evaluating multimodal reasoning and tool use in long videos, featuring open-ended questions and diagnostic rubrics. The benchmark addresses the limitations of existing datasets by combining temporal length and multimodal richness, using human-validated samples. LongShOTAgent, an agentic system, is also presented for analyzing long videos, with both the benchmark and agent demonstrating the challenges faced by state-of-the-art MLLMs.
A new study compares vision-language models (VLMs) to YOLOv8 for wastewater treatment plant (WWTP) identification in satellite imagery across the MENA region. VLMs like Gemma-3 demonstrate superior zero-shot performance compared to YOLOv8, trained on a dataset of 83,566 satellite images from Egypt, Saudi Arabia, and UAE. The research suggests VLMs offer a scalable, annotation-free alternative for remote sensing of WWTPs.
Researchers at MBZUAI introduce "Interactive Video Reasoning," a new paradigm enabling models to actively "think with videos" by performing iterative visual actions to gather and refine evidence. They developed Video CoM, which reasons through a Chain of Manipulations (CoM), and constructed Video CoM Instruct, an 18K instruction tuning dataset for multi-step manipulation reasoning. The model is further optimized via reinforcement learning with reasoning aware Group Relative Policy Optimization (GRPO), achieving strong results across nine video reasoning benchmarks.
Researchers at MBZUAI have introduced Video-R2, a reinforcement learning approach to improve the consistency and visual grounding of reasoning in multimodal language models. Video-R2 combines timestamp-aware supervised fine-tuning with Group Relative Policy Optimization (GRPO) guided by a Temporal Alignment Reward (TAR). The model demonstrates higher Think Answer Consistency (TAC), Video Attention Score (VAS), and accuracy across multiple benchmarks, showing improved temporal alignment and reasoning coherence for video understanding.
This paper analyzes the energy consumption and carbon footprint of LLM inference in the UAE compared to Iceland, Germany, and the USA. The study uses DeepSeek Coder 1.3B and the HumanEval dataset to evaluate code generation. It provides a comparative analysis of geographical trade-offs for climate-aware AI deployment, specifically addressing the challenges and potential of datacenters in desert regions.
Researchers at MBZUAI have introduced EvoLMM, a self-evolving framework for large multimodal models that enhances reasoning capabilities without human-annotated data or reward distillation. EvoLMM uses two cooperative agents, a Proposer and a Solver, which generate image-grounded questions and solve them through internal consistency, using a continuous self-rewarding process. Evaluations using Qwen2.5-VL as the base model showed performance gains of up to 3% on multimodal math-reasoning benchmarks like ChartQA, MathVista, and MathVision using only raw training images.
This paper proposes a framework for understanding AI sovereignty as a balance between autonomy and interdependence, considering global data, supply chains, and standards. It introduces a planner's model with policy heuristics for equalizing marginal returns across sovereignty pillars and setting openness. The model is applied to India and the Middle East (Saudi Arabia and UAE), finding that managed interdependence, rather than isolation, is key for AI sovereignty.
arXiv · · Research Healthcare
Researchers from MBZUAI have developed MMRINet, a Mamba-based neural network for efficient brain tumor segmentation in MRI scans. The model uses Dual-Path Feature Refinement and Progressive Feature Aggregation to achieve high accuracy with only 2.5M parameters, making it suitable for low-resource clinical environments. MMRINet achieves a Dice score of 0.752 and HD95 of 12.23 on the BraTS-Lighthouse SSA 2025 benchmark.
This paper introduces Cross-Document Topic-Aligned (CDTA) chunking to address knowledge fragmentation in Retrieval-Augmented Generation (RAG) systems. CDTA identifies topics across documents, maps segments to topics, and synthesizes them into unified chunks. Experiments on HotpotQA and UAE legal texts show that CDTA improves faithfulness and citation accuracy compared to existing chunking methods, especially for complex queries requiring multi-hop reasoning.
A new method is proposed to reduce the verbosity of LLMs in step-by-step reasoning by retaining moderately easy problems during Reinforcement Learning with Verifiable Rewards (RLVR) training. This approach acts as an implicit length regularizer, preventing the model from excessively increasing output length on harder problems. Experiments using Qwen3-4B-Thinking-2507 show the model achieves baseline accuracy with nearly twice shorter solutions.
This paper introduces a framework that combines machine learning for multi-class attack detection in IoT/IIoT networks with large language models (LLMs) for attack behavior analysis and mitigation suggestion. The framework uses role-play prompt engineering with RAG to guide LLMs like ChatGPT-o3 and DeepSeek-R1, and introduces new evaluation metrics for quantitative assessment. Experiments using Edge-IIoTset and CICIoT2023 datasets showed Random Forest as the best detection model and ChatGPT-o3 outperforming DeepSeek-R1 in attack analysis and mitigation.
Researchers introduce MATRIX, a vision-centric agent tuning framework for robust tool-use reasoning in VLMs. The framework includes M-TRACE, a dataset of 28.5K multimodal tasks with 177K verified trajectories, and Pref-X, a set of 11K automatically generated preference pairs. Experiments show MATRIX consistently outperforms open- and closed-source VLMs across three benchmarks.
Researchers introduce ALARB, a new benchmark for evaluating reasoning in Arabic LLMs using 13K Saudi commercial court cases. The benchmark includes tasks like verdict prediction, reasoning chain completion, and identification of relevant regulations. Instruction-tuning a 12B parameter model on ALARB achieves performance comparable to GPT-4o in verdict prediction and generation.
Researchers from MBZUAI have introduced SPECS, a new reference-free evaluation metric for long image captions that modifies CLIP to emphasize specificity. SPECS aims to improve the correlation with human judgment while maintaining computational efficiency compared to LLM-based metrics. The proposed approach is intended for iterative use during image captioning model development, offering a practical alternative to existing methods.
This paper presents a UI-level evaluation of ALLaM-34B, an Arabic-centric LLM developed by SDAIA and deployed in the HUMAIN Chat service. The evaluation used a prompt pack spanning various Arabic dialects, code-switching, reasoning, and safety, with outputs scored by frontier LLM judges. Results indicate strong performance in generation, code-switching, MSA handling, reasoning, and improved dialect fidelity, positioning ALLaM-34B as a robust Arabic LLM suitable for real-world use.
This paper introduces Saudi-Dialect-ALLaM, a LoRA fine-tuned version of the Saudi Arabian foundation model ALLaM-7B-Instruct-preview, designed to improve the generation of Saudi dialects (Najdi and Hijazi). The model is trained on a private dataset of 5,466 synthetic instruction-response pairs, with two variants explored: Dialect-Token and No-Token training. Results indicate that the Dialect-Token model achieves superior dialect control and fidelity compared to generic instruction models, although the dataset and model weights are not released.
Researchers introduce UnsafeChain, a new safety alignment dataset designed to improve the safety of large reasoning models (LRMs) by focusing on 'hard prompts' that elicit harmful outputs. The dataset identifies and corrects unsafe completions into safe responses, exposing models to unsafe behaviors and guiding their correction. Fine-tuning LRMs on UnsafeChain demonstrates enhanced safety and preservation of general reasoning ability compared to existing datasets like SafeChain and STAR-1.
This paper introduces Absher, a new benchmark for evaluating LLMs' linguistic and cultural competence in Saudi dialects. The benchmark comprises over 18,000 multiple-choice questions spanning six categories, using dialectal words, phrases, and proverbs from various regions of Saudi Arabia. Evaluation of state-of-the-art LLMs reveals performance gaps, especially in cultural inference and contextual understanding, highlighting the need for dialect-aware training.
arXiv · · Research Healthcare
MBZUAI researchers have introduced MIRA, a novel framework for improving the factual accuracy of multimodal large language models in medical applications. MIRA uses calibrated retrieval to manage factual risk and integrates image embeddings with a medical knowledge base for efficient reasoning. Evaluated on medical VQA and report generation benchmarks, MIRA achieves state-of-the-art results, with code available on GitHub.
arXiv · · Healthcare Research
Researchers from MBZUAI introduce Forget-MI, a machine unlearning method tailored for multimodal medical data, enhancing privacy by removing specific patient data from AI models. Forget-MI utilizes loss functions and perturbation techniques to unlearn both unimodal and joint data representations. The method demonstrates superior performance in reducing Membership Inference Attacks and improving data removal compared to existing techniques, while preserving overall model performance and enabling data forgetting.
This paper introduces MOTOR, a multimodal retrieval and re-ranking approach for medical visual question answering (MedVQA) that uses grounded captions and optimal transport to capture relationships between queries and retrieved context, leveraging both textual and visual information. MOTOR identifies clinically relevant contexts to augment VLM input, achieving higher accuracy on MedVQA datasets. Empirical analysis shows MOTOR outperforms state-of-the-art methods by an average of 6.45%.
A new methodology emulating fact-checker criteria assesses news outlet factuality and bias using LLMs. The approach uses prompts based on fact-checking criteria to elicit and aggregate LLM responses for predictions. Experiments demonstrate improvements over baselines, with error analysis on media popularity and region, and a released dataset/code at https://github.com/mbzuai-nlp/llm-media-profiling.
A new benchmark, ViMUL-Bench, is introduced to evaluate video LLMs across 14 languages, including Arabic, with a focus on cultural inclusivity. The benchmark includes 8k manually verified samples across 15 categories and varying video durations. A multilingual video LLM, ViMUL, is also presented, along with a training set of 1.2 million samples, with both to be publicly released.
MBZUAI researchers introduce TerraFM, a scalable self-supervised learning model for Earth observation that uses Sentinel-1 and Sentinel-2 imagery. The model unifies radar and optical inputs through modality-specific patch embeddings and adaptive cross-attention fusion. TerraFM achieves strong generalization on classification and segmentation tasks, outperforming prior models on GEO-Bench and Copernicus-Bench.
Researchers from MBZUAI have introduced VideoMolmo, a large multimodal model for spatio-temporal pointing conditioned on textual descriptions. The model incorporates a temporal module with an attention mechanism and a temporal mask fusion pipeline using SAM2 for improved coherence across video sequences. They also curated a dataset of 72k video-caption pairs and introduced VPoS-Bench, a benchmark for evaluating generalization across real-world scenarios, with code and models publicly available.
MBZUAI researchers introduce VideoMathQA, a new benchmark for evaluating mathematical reasoning in videos, requiring models to interpret visual information, text, and spoken cues. The dataset spans 10 mathematical domains with videos ranging from 10 seconds to over 1 hour, and includes multi-step reasoning annotations. The benchmark aims to evaluate temporal cross-modal reasoning and highlights the limitations of existing approaches in complex video-based mathematical problem solving.
MBZUAI introduces Agent-X, a benchmark for evaluating multi-step reasoning in vision-centric agents across real-world, multimodal settings. Agent-X includes 828 tasks with diverse visual contexts and spans six environments, requiring tool use and stepwise decision-making. Experiments show that current LLMs struggle with multi-step vision tasks, achieving less than 50% success, highlighting areas for improvement in LMM reasoning and tool use.
MBZUAI researchers introduce SocialMaze, a new benchmark for evaluating social reasoning capabilities in large language models (LLMs). SocialMaze includes six diverse tasks across social reasoning games, daily-life interactions, and digital community platforms, emphasizing deep reasoning, dynamic interaction, and information uncertainty. Experiments show that LLMs vary in handling dynamic interactions, degrade under uncertainty, but can be improved via fine-tuning on curated reasoning examples.
MBZUAI researchers release 'Fann or Flop', a new benchmark for evaluating Arabic poetry understanding in LLMs. The benchmark covers 12 historical eras and 14 poetic genres, assessing semantic understanding, metaphor interpretation, and cultural context. Evaluation of state-of-the-art LLMs reveals challenges in poetic understanding despite strong performance on standard Arabic benchmarks.
MBZUAI researchers introduce ARB, the first comprehensive benchmark for evaluating step-by-step multimodal reasoning in Arabic across textual and visual modalities. The benchmark spans 11 diverse domains and includes 1,356 multimodal samples with 5,119 human-curated reasoning steps. Evaluations of 12 state-of-the-art LMMs revealed challenges in coherence, faithfulness, and cultural grounding, highlighting the need for culturally aware AI systems.
Researchers from MBZUAI have introduced UrduFactCheck, a new framework for fact-checking in Urdu, along with two datasets: UrduFactBench and UrduFactQA. The framework uses monolingual and translation-based evidence retrieval to address the lack of Urdu resources. Evaluations using twelve LLMs showed that translation-augmented methods improve performance, highlighting challenges for open-source LLMs in Urdu.
MBZUAI researchers introduce FAID, a fine-grained AI-generated text detection framework capable of classifying text as human-written, LLM-generated, or collaboratively written. FAID utilizes multi-level contrastive learning and multi-task auxiliary classification to capture authorship and model-specific characteristics, and can identify the underlying LLM family. The framework outperforms existing baselines, especially in generalizing to unseen domains and new LLMs, and includes a multilingual, multi-domain dataset called FAIDSet.
MBZUAI researchers introduce LLM-BabyBench, a benchmark suite for evaluating grounded planning and reasoning in LLMs. The suite, built on a textual adaptation of the BabyAI grid world, assesses LLMs on predicting action consequences, generating action sequences, and decomposing instructions. Datasets, evaluation harness, and metrics are publicly available to facilitate reproducible assessment.
This paper introduces SemDiff, a novel method for generating unrestricted adversarial examples (UAEs) by exploring the semantic latent space of diffusion models. SemDiff uses multi-attribute optimization to ensure attack success while preserving the naturalness and imperceptibility of generated UAEs. Experiments on high-resolution datasets demonstrate SemDiff's superior performance compared to state-of-the-art methods in attack success rate and imperceptibility, while also evading defenses.
The paper introduces SaudiCulture, a new benchmark for evaluating the cultural competence of LLMs within Saudi Arabia, covering five major geographical regions and diverse cultural domains. The benchmark includes questions of varying complexity and distinguishes between common and specialized regional knowledge. Evaluations of five LLMs (GPT-4, Llama 3.3, FANAR, Jais, and AceGPT) revealed performance declines on region-specific questions, highlighting the need for region-specific knowledge in LLM training.
arXiv · · Research Healthcare
Researchers introduce SALT, a parameter-efficient fine-tuning method for medical image segmentation that combines singular value adaptation with low-rank transformation. SALT selectively adapts influential singular values and complements this with a low-rank update for the remaining subspace. Experiments on five medical datasets show SALT outperforms state-of-the-art PEFT methods by 2-5% in Dice score with only 3.9% trainable parameters.
arXiv · · Research Healthcare
The paper introduces UAE-3D, a multi-modal VAE for 3D molecule generation that compresses molecules into a unified latent space, maintaining near-zero reconstruction error. This approach simplifies latent diffusion modeling by eliminating the need to handle multi-modality and equivariance separately. Experiments on GEOM-Drugs and QM9 datasets show UAE-3D establishes new benchmarks in de novo and conditional 3D molecule generation, with significant improvements in efficiency and quality.
Researchers at MBZUAI have introduced a novel approach to enhance Large Multimodal Models (LMMs) for autonomous driving by integrating 3D tracking information. This method uses a track encoder to embed spatial and temporal data, enriching visual queries and improving the LMM's understanding of driving scenarios. Experiments on DriveLM-nuScenes and DriveLM-CARLA benchmarks demonstrate significant improvements in perception, planning, and prediction tasks compared to baseline models.
MBZUAI researchers introduce LLMVoX, a 30M-parameter, LLM-agnostic, autoregressive streaming text-to-speech (TTS) system that generates high-quality speech with low latency. The system preserves the capabilities of the base LLM and achieves a lower Word Error Rate compared to speech-enabled LLMs. LLMVoX supports seamless, infinite-length dialogues and generalizes to new languages with dataset adaptation, including Arabic.
A new survey paper provides a deep dive into post-training methodologies for Large Language Models (LLMs), analyzing their role in refining LLMs beyond pretraining. It addresses key challenges such as catastrophic forgetting, reward hacking, and inference-time trade-offs, and highlights emerging directions in model alignment, scalable adaptation, and inference-time reasoning. The paper also provides a public repository to continually track developments in this fast-evolving field.
A new culturally inclusive and linguistically diverse dataset called Palm for Arabic LLMs is introduced, covering 22 Arab countries and featuring instructions in both Modern Standard Arabic (MSA) and dialectal Arabic (DA) across 20 topics. The dataset was built through a year-long community-driven project involving 44 researchers from across the Arab world. Evaluation of frontier LLMs using the dataset reveals limitations in cultural and dialectal understanding, with some countries being better represented than others.
Researchers introduce a benchmark to evaluate the factual recall and knowledge transferability of multilingual language models across 13 languages. The study reveals that language models often fail to transfer knowledge between languages, even when they possess the correct information in one language. The benchmark and evaluation framework are released to drive future research in multilingual knowledge transfer.
Researchers introduce TimeTravel, a benchmark dataset for evaluating large multimodal models (LMMs) on historical and cultural artifacts. The benchmark comprises 10,250 expert-verified samples across 266 cultures and 10 historical regions, designed to assess AI in tasks like classification and interpretation of manuscripts, artworks, inscriptions, and archaeological discoveries. The goal is to establish AI as a reliable partner in preserving cultural heritage and assisting researchers.
The paper introduces VENOM, a text-driven framework for generating high-quality unrestricted adversarial examples using diffusion models. VENOM unifies image content generation and adversarial synthesis into a single reverse diffusion process, enhancing both attack success rate and image quality. The framework incorporates an adaptive adversarial guidance strategy with momentum to ensure the generated adversarial examples align with the distribution of natural images.
Researchers at MBZUAI have demonstrated a method called "Data Laundering" to artificially boost language model benchmark scores using knowledge distillation. The technique covertly transfers benchmark-specific knowledge, leading to inflated accuracy without genuine improvements in reasoning. The study highlights a vulnerability in current AI evaluation practices and calls for more robust benchmarks.