Using a zoom-in tool is an important foundational part of modern visual agents, because it allows to efficiently handle tasks involving high-resolution images. Most previous methods need an extensive warm-start supervised fine-tuning phase for teaching models zoom-in. We show that this is not necessary by proposing a new intrinsic reward for learning tool use in MLLMs without the need for additional labels or warm-start SFT. Our InfoNCE-style reward uses a curriculum of increasingly hard negative tool calls as a contrastive training signal. Empirical experiments on $V^*$, HRBench and MME-RealWorld show that our approach is competitive while being more efficient. When used as a drop-in replacement for SFT, we even outperform all baselines. To directly measure the zoom-in ability of models, we further introduce the scalable synthetic Muffin&Chihuahua (M&C) dataset. Each image consists of a grid with every cell either showing a muffin or chihuahua. Leveraging the M&C dataset's unique region of interest labels, we find that recall is the metric that most strongly correlates the zoom-in region with final task performance. Our model and code for reproduction is publicly available under https://github.com/UKPLab/emnlp2026-zoom-in
Evaluation is shifting from static QA toward agentic settings where models act through external tools. We identify a critical yet underexplored capability within this space - dexterous visual tool use: fine-grained, closed-loop parameterized visual action in which models infer tool parameters from visual evidence, and those parameters directly govern the final result. Existing benchmarks cover web navigation, GUI operation, and software engineering, but rarely target this coupling between visual evidence and execution precision. We propose EASEL, a benchmark evaluating a controlled instance of dexterous visual tool use that adopts reference-guided visual reconstruction as its primary proxy task: the agent incrementally paints a canvas to match a reference image. EASEL additionally includes semantic tasks spanning region annotation, handwriting, and path planning. We further provide EASEL-Data, a 440k-sample two-stage curriculum dataset for trajectory supervision, and EASEL-9B to investigate its effect on this capability. Evaluation of 25 models reveals that current multimodal agents systematically struggle on EASEL. Reconstruction similarity bottlenecks at low levels (0.40-0.54), while trajectory diagnostics expose severe closed-loop instability - models typically saturate early or degrade post-peak. Semantic tasks reveal sharp capability boundaries in precision annotation and path planning. EASEL-9B, trained on EASEL-Data, surpasses the base model by a relative 6.3%, ranking third among all evaluated models.
Although multimodal large language models (MLLMs) have shown substantial potential in visual understanding and graphic code generation, editing scientific figures through code presents a greater challenge: a model must jointly recover visual structure, ground the requested change, generate compilable code, and preserve all unrelated content. While existing TikZ benchmarks mainly focus on figure reconstruction and generation, few systematically evaluate instruction-guided scientific figure editing with compilable code. We introduce Edit2TikZ, a comprehensive benchmark for scientific figure editing tasks, featuring 1,548 diverse and high-quality samples. Edit2TikZ combines real-world and controlled synthetic edit cases, supports both textual and visual localization request, and contains multi-step editing, each with step-level annotations. We further construct a human-aligned evaluation framework to measure whether a requested edit is completed while irrelevant content is preserved. Utilizing Edit2TikZ, we evaluate 14 mainstream MLLMs and find that current systems remain unreliable: on average, proprietary models achieve a compilation success rate of merely 75% and remain limited in both figure restoration and edit correctness, while compact models below 9B struggle further with instruction following and complete figure generation. Therefore, we build a mixed training set TikZEditMix and adopt reconstruction-then-editing curriculum learning for compact models. On Qwen3.5-4B, this training improves the compilation success rate from 45.35% to 83.40% and yields an average improvement of 18.7 points across our proposed evaluation metrics. The code and data will be released at https://github.com/Solunny/Edit2TikZ.
Understanding dynamic sound sources requires jointly determining what produces a sound, where the source is located, and how it moves over time. Yet existing audio-language models often represent clips as global acoustic events, while vision-language models lack the spatial audio cues needed to localize and track individual sources. To evaluate this missing capability, we introduce ST-OmniQA, a spatio-temporal audio-visual question-answering benchmark built from panoramic videos paired with synchronized first-order Ambisonics (FOA) audio of moving sound sources. It contains 40K videos and 400K question-answer pairs organized into four capability levels covering sound-event recognition, direction of arrival, source distance, motion trajectories, and temporally grounded audio-visual reasoning. Building on this benchmark, we propose ST-Omni-R1, which integrates FOA-derived semantic and trajectory representations with panoramic visual context and is trained through progressive curriculum learning and reasoning-tree reinforcement learning. ST-Omni-R1 achieves 77.83\% average semantic accuracy across the four levels, compared with 37.28\% for the best evaluated baseline. Results on three public spatial-audio benchmarks further indicate that its learned spatial and motion representations transfer beyond ST-OmniQA.
Yexing Du, Kaiyuan Liu, Youcheng Pan +4cs.CL cs.AI
Multimodal large language models (MLLMs) have achieved significant success in speech-to-text translation (S2TT). However, when processing multilingual speech inputs, a single speech encoder shared across all languages suffers from the curse of multilinguality: languages at different resource levels compete for limited representation capacity, leading to strong high-resource performance but substantial degradation on low-resource speech. To address this problem and improve multilingual consistency, we propose MSRT, a novel framework built around a resource-aware Mixture of Speech Encoders (MoSE). MoSE uses an explicit language router to assign each utterance to an appropriate expert encoder. A frozen expert preserves high-resource language capabilities, while a trainable expert adapts to and specializes in medium- and low-resource languages. We further introduce a five-stage curriculum learning strategy that substantially reduces data dependence, requiring only 10 hours of paired S2TT data per language for effective alignment. We conduct extensive experiments on 45 languages, systematically evaluating all $45 \times 44$ translation directions. Our 4B-parameter model achieves state-of-the-art performance, outperforming substantially larger baselines. Empirical analyses show that MoSE improves high-, medium-, and low-resource languages simultaneously, with the largest gains on low-resource speech, thereby breaking the curse of multilinguality without compromising high-resource performance. To support future multilingual S2TT research, we release our code and models.
Chart question answering (CQA) requires multimodal large language models (MLLMs) to integrate visual comprehension with logical reasoning, yet current models struggle with accurate visual grounding and coherent reasoning chains. While extrinsic chain-of-thought prompting and visual cues significantly improve performance, current MLLMs lack intrinsic visual grounded reasoning capabilities, leading to inaccurate perception and reasoning disconnected from visual evidence. To address these limitations, we propose CURV, a curriculum learning framework that develops intrinsic visual reasoning capabilities by reformulating CQA as multi-step visual grounded reasoning, where each step coordinates logical reasoning with dynamic visual grounding through spatial attention concentration. To assist model learning, we further introduce CCQA, a three-level curriculum dataset with scalable synthetic generation across diverse chart types and reasoning patterns. Our curriculum systematically progresses from basic single-operation reasoning to complex multi-chart compositional tasks. Experiments demonstrate that CURV achieves up to $\uparrow20.50\%$ improvements over baselines and is generalizable to real-world benchmarks (up to $\uparrow12.30\%$) and out-of-domain multimodal reasoning tasks (up to $\uparrow10.20\%$), validating the effectiveness of internalizing visual reasoning with dynamic grounding for enhanced chart understanding capabilities. Code is available at: https://xhguo7.github.io/CURV/.
Multimodal large language models (MLLMs) still struggle with spatial reasoning that requires perspective transformation. In particular, they often rely on camera-centric cues rather than reasoning from the reference object's viewpoint, leading to systematic errors in non-camera reference settings. In this paper, we first analyze this failure mode and show that object orientation is a key factor underlying such camera-centric shortcut behavior. To address this issue, we propose OrientSAM, an orientation-aware spatial alignment framework for multimodal models. OrientSAM injects explicit orientation information into multimodal representations through orientation-aware tokens and Fourier-based angle encoding, and further adopts a curriculum learning strategy to progressively improve perspective-aware reasoning. In addition, we build a spatial data construction pipeline to generate orientation-aware spatial supervision from large-scale images. Experiments on Spatial-MM, ViewSpatial, and 3DSRBench show that OrientSAM consistently outperforms strong baselines, especially on non-camera-view, person-centric, and orientation-sensitive tasks. The results further demonstrate that explicit orientation modeling is important for mitigating camera-centric shortcut behavior and enabling more robust allocentric spatial reasoning in multimodal models.
We present Audio-Visual Flamingo (AV-Flamingo), a fully open state-of-the-art audio-visual large language model (AV-LLM) for joint understanding and reasoning over audio, images, and long-form videos. Unlike prior AV-LLMs that primarily focus on short clips, AV-Flamingo is designed for understanding and reasoning over long and complex real-world (audio-visual) videos. To support this, we make three key contributions: (i) Audio-Visual-Skills, a large-scale collection of real-world videos with ~7M caption and question-answer training instances designed to emphasize temporal, compositional, and cross-modal audio-visual reasoning; (ii) a novel three-stage curriculum that progressively trains the model from short-range perception to long-horizon multi-event reasoning; and (iii) Temporal Audio-Visual Interleaved Chain-of-Thought, a reasoning framework that explicitly grounds intermediate reasoning steps to timestamps in long audio-visual streams, improving temporal alignment and interpretability. Extensive experiments across 15+ audio-visual, omni-modal, audio, and vision benchmarks show that AV-Flamingo outperforms similarly sized open models by clear margins and remains highly competitive with, and in some cases surpasses, much larger open-weight and closed models, particularly on long and complex real-world audio-visual understanding and reasoning tasks. Beyond benchmark performance, AV-Flamingo exhibits strong real-world utility and transfers well to unseen tasks, highlighting its robustness and generalization ability.
Music-driven dance video generation aims to synthesize expressive human motion that is temporally aligned with music while maintaining high visual fidelity. Despite recent progress, existing methods still face two key limitations: the lack of large-scale, high-quality dance video datasets, and the absence of principled frameworks for integrating music as a complementary conditioning signal into Video Generation Foundation Models. To address these limitations, we introduce CIPE-Dance, a large-scale Internet-sourced dance video dataset with choreography-informed text annotations, constructed via a progressive expert pipeline. To the best of our knowledge, CIPE-Dance is the largest dataset for dance video generation to date, comprising 300k high-quality clips over 400 hours and covering diverse dancers, environments, and dance genres. We further propose OmniDance, a framework-level recipe for integrating music into a TI2V foundation model without sacrificing its original controllability or visual fidelity. Motivated by the complementary roles of text as low-frequency semantics and music as high-frequency temporal dynamics, OmniDance co-designs a depth-aware specialization architecture, an anchored easy-to-hard curriculum learning strategy, and a modality-specialized time-dependent CFG strategy, enabling unified TI2V, MI2V, and MTI2V generation. Extensive experiments on CIPE-Dance demonstrate that OmniDance achieves state-of-the-art performance across all three tasks and exhibits robust multimodal integration capability. Project is available at https://github.com/AMAP-ML/OmniDance.
Achieving human-like reasoning in Vision-Language Models (VLMs) remains a long-standing challenge. Recent approaches leverage Chain-of-Thought (CoT) rationales generated by human annotators or proprietary models, which are costly and difficult to scale. Self-training offers a promising alternative but often suffers from visual hallucinations and language shortcuts because rationales are filtered only by answer correctness without verifying visual perception. We propose a perception-verified self-training framework that enforces visually grounded reasoning. Our method employs a CoT template (caption-reasoning-conclusion) that disentangles perception from reasoning, enabling independent verification of visual understanding. To compensate for the absence of ground-truth captions, we introduce PerceptEval, an unsupervised method that evaluates caption quality based on its alignment with visual and textual elements in the image. Using caption verification together with answer correctness, we partition the data into easy, medium, and hard subsets and design a two-stage curriculum learning strategy. Stage 1 trains on easy samples, while Stage 2 enhances medium samples by regenerating reasoning conditioned on verified captions and retaining only those with correct conclusions. This ensures training is performed exclusively on perceptually grounded reasoning, reducing hallucinations and language shortcuts. Extensive experiments across diverse domains and models demonstrate improvements of up to 16% over standard self-training baselines, showing that our framework provides a scalable and cost-effective solution for advancing multimodal reasoning without manually annotated CoT rationales.
Multimodal learning exploits complementary information across heterogeneous modalities. The informativeness of each modality can vary widely across samples and training stages. Existing multimodal curriculum learning strategies often assume that the relative complexity of samples remains unchanged throughout training and therefore cannot adapt to model evolution. We propose SPICE (Synergy and Partial Information based Curriculum Evolution), a novel progressive curriculum framework for multimodal interaction learning. Guided by Partial Information Decomposition (PID) theory, our approach decomposes multimodal interactions into redundant, unique, and synergistic information components, enabling an interpretable and dynamic characterization of sample complexity. Building on this decomposition, we design a progressive curriculum that evolves throughout training, allowing the model to transition from learning shared cross-modal cues to modality-specific patterns and, finally, to complex synergistic interactions. Adapting to model evolution, sample ordering is refined in real-time using PID information estimates derived from unimodal and multimodal predictions. Experiments across multiple multimodal benchmarks demonstrate consistent improvements over conventional training and state-of-the-art baselines, highlighting the effectiveness of PID information decomposition and adaptive sample ordering for multimodal curriculum learning.
Recent advances in self-evolution video understanding frameworks have demonstrated the potential of autonomous learning without human annotations. However, existing methods often suffer from weakly controlled optimization and uncontrolled difficulty progression, as they lack structured guidance throughout the iterative learning process. To address these limitations, we propose CurEvo, a curriculum-guided self-evolution framework that introduces curriculum learning into self-evolution to achieve more structured and progressive model improvement. CurEvo dynamically regulates task difficulty, refines evaluation criteria, and balances data diversity according to model competence, forming a curriculum-guided feedback loop that aligns learning complexity with model capability. Built upon this principle, we develop a multi-dimensional adaptive QA framework that jointly evolves question generation and answer evaluation across perception, recognition, and understanding dimensions, ensuring coherent and measurable curriculum progression. Through this integration, CurEvo transforms weakly controlled self-evolution into a more structured learning process for autonomous video understanding. Across seven backbones, CurEvo consistently improves both benchmark accuracy and evaluator-based semantic score on four VideoQA benchmarks, validating the effectiveness of curriculum-guided self-evolution for video understanding.