Large language models can exhibit emergent reasoning behaviors, often manifested as recurring lexical patterns (e.g., “wait,” indicating verification). However, complex reasoning trajectories remain sparse in unconstrained sampling, and standard RL often fails to guarantee the acquisition of diverse reasoning behaviors. We propose a systematic discovery and reinforcement of diverse reasoning patterns through structured reasoning, a paradigm that requires targeted exploration of specific reasoning patterns during the RL process. To this end, we propose Ctrl-R, a framework for learning structured reasoning via tractable trajectory control that actively guides the rollout process, incentivizing the exploration of diverse reasoning patterns that are critical for complex problem-solving. The resulting behavior policy enables accurate importance-sampling estimation, supporting unbiased on-policy optimization. We further introduce a power-scaling factor on the importance-sampling weights, allowing the policy to selectively learn from exploratory, out-of-distribution trajectories while maintaining stable optimization. Experiments demonstrate that Ctrl-R enables effective exploration and internalization of previously unattainable reasoning patterns, yielding consistent improvements across language and vision–language models on mathematical reasoning tasks.

Unified multimodal Large Language Models (LLMs) that can both understand and generate visual content hold immense potential. However, existing open-source models often suffer from a performance trade-off between these capabilities. We present Manzano, a simple and scalable unified framework that substantially reduces this tension by coupling a hybrid image tokenizer with a well-curated training recipe. A single shared vision encoder feeds two lightweight adapters that produce continuous embeddings for image-to-text understanding and discrete tokens for text-to-image generation within a common semantic space. A unified autoregressive LLM predicts high-level semantics in the form of text and image tokens, with an auxiliary diffusion decoder subsequently translating the image tokens into pixels. The architecture, together with a unified training recipe over understanding and generation data, enables scalable joint learning of both capabilities. Manzano achieves state-of-the-art results among unified models, and is competitive with specialist models, particularly on text-rich evaluation. Our studies show minimal task conflicts and consistent gains from scaling model size, validating our design choice of a hybrid tokenizer.

Preference finetuning methods like Direct Preference Optimization (DPO) with AI-generated feedback have shown promise in aligning Large Vision-Language Models (LVLMs) with human preferences. However, existing techniques overlook the prevalence of noise in synthetic preference annotations in the form of stylistic and length biases. To this end, we introduce a hard-negative response generation framework based on LLM-guided response editing, that produces rejected responses with targeted errors, maintaining stylistic and length similarity to the accepted ones. Using this framework, we develop the VaPR dataset, comprising 30K high-quality samples, to finetune three LVLM families: LLaVA-V1.5, Qwen2VL & Qwen2.5VL (2B-13B sizes). Our VaPR models deliver significant performance improvements across ten benchmarks, achieving average gains of 6.5% (LLaVA), 4.0% (Qwen2VL), and 1.5% (Qwen2.5VL), with notable improvements on reasoning tasks. A scaling analysis shows that performance consistently improves with data size, with LLaVA models benefiting even at smaller scales. Moreover, VaPR reduces the tendency to answer "Yes" in binary questions - addressing a common failure mode in LVLMs like LLaVA. Lastly, we show that the framework generalizes to open-source LLMs as editors, with models trained on VaPR-OS achieving ~99% of the performance of models trained on \name, which is synthesized using GPT-4o. Our data, models, and code can be found on the project page https://vap-r.github.io

Multimodal agents for computer use rely exclusively on primitive actions (click, type, scroll) that require accurate visual grounding and lengthy execution chains, leading to cascading failures and performance bottlenecks. While other agents leverage rich programmatic interfaces (APIs, MCP servers, tools), computer-use agents (CUAs) remain isolated from these capabilities. We present UltraCUA, a foundation model that bridges this gap through hybrid action – seamlessly integrating GUI primitives with high-level programmatic tool calls. To achieve this, our approach comprises four key components: (1) an automated pipeline that scales programmatic tools from software documentation, open-source repositories, and code generation; (2) a synthetic data engine producing over 17,000 verifiable tasks spanning real-world computer-use scenarios; (3) a large-scale high-quality hybrid action trajectory collection with both low-level GUI actions and high-level programmatic tool calls; and (4) a two-stage training pipeline combining supervised fine-tuning with online reinforcement learning, enabling strategic alternation between low-level and high-level actions. Experiments with our 7B and 32B models demonstrate substantial improvements over state-of-the-art agents. On OSWorld, UltraCUA models achieve an average 22

We introduce two multilingual, multimodal foundation language models that power Apple Intelligence features across Apple devices and services: i a 3B-parameter on-device model optimized for Apple silicon through architectural innovations such as KV-cache sharing and 2-bit quantization-aware training; and ii a scalable server model built on a novel Parallel-Track Mixture-of-Experts PT-MoE transformer that combines track parallelism, mixture-of-experts sparse computation, and interleaved global-local attention to deliver high quality with competitive cost on Apple's Private Cloud Compute platform. Both models are trained on large-scale multilingual and multimodal datasets sourced via responsible web crawling, licensed corpora, and high-quality synthetic data, then further refined with supervised fine-tuning and reinforcement learning on a new asynchronous platform. The resulting models support several additional languages while understanding images and executing tool calls. In public benchmarks and human evaluations, both the server model and the on-device model match or surpass comparably sized open baselines. A new Swift-centric Foundation Models framework exposes guided generation, constrained tool calling, and LoRA adapter fine-tuning, allowing developers to integrate these capabilities with a few lines of code. The latest advancements in Apple Intelligence models are grounded in our Responsible AI approach with safeguards like content filtering and locale-specific evaluation, as well as our commitment to protecting our users' privacy with innovations like Private Cloud Compute.

Existing multimodal retrieval benchmarks primarily focus on evaluating whether models can retrieve and utilize external textual knowledge for question answering. However, there are scenarios where retrieving visual information is either more beneficial or easier to access than textual data. In this paper, we introduce a multimodal retrieval-augmented generation benchmark, MRAG-Bench, in which we systematically identify and categorize scenarios where visually augmented knowledge is better than textual knowledge, for instance, more images from varying viewpoints.MRAG-Bench consists of 16,130 images and 1,353 human-annotated multiple-choice questions across 9 distinct scenarios. With MRAG-Bench, we conduct an evaluation of 10 open-source and 4 proprietary large vision-language models (LVLMs). Our results show that all LVLMs exhibit greater improvements when augmented with images compared to textual knowledge, confirming that MRAG-Bench is vision-centric. Additionally, we conduct extensive analysis with MRAG-Bench, which offers valuable insights into retrieval-augmented LVLMs. Notably, the top-performing model, GPT-4o, faces challenges in effectively leveraging retrieved knowledge, achieving only a 5.82\% improvement with ground-truth information, in contrast to a 33.16\% improvement observed in human participants. These findings highlight the importance of MRAG-Bench in encouraging the community to enhance LVLMs' ability to utilize retrieved visual knowledge more effectively.

Developing autonomous agents that effectively interact with Graphic User Interfaces (GUIs) remains a challenging open problem, especially for small on-device models. In this paper, we present Ferret-UI Lite, a compact, end-to-end GUI agent that operates across diverse platforms, including mobile, web, and desktop. Utilizing techniques optimized for developing small models, we build our 3B Ferret-UI Lite agent through curating a diverse GUI data mixture from real and synthetic sources, strengthening inference-time performance through chain-of-thought reasoning and visual tool-use, and reinforcement learning with designed rewards. Ferret-UI Lite achieves competitive performance with other small-scale GUI agents. In GUI grounding, Ferret-UI Lite attains scores of 91.6%, 53.3%, and 61.2% on the ScreenSpot-V2, ScreenSpot-Pro, and OSWorld-G benchmarks, respectively. For GUI navigation, Ferret-UI Lite achieves success rates of 28.0% on AndroidWorld and 19.8% on OSWorld. We share our methods and lessons learned from developing compact, on-device GUI agents.

There is increasing interest in equipping language models with the ability to leverage external tools for complex, goal-oriented tasks. However, interacting with external tools introduces inherent uncertainties due to imperfections and misalignments between the tools’ outputs and the agents’ internal models, often leading to suboptimal outcomes. We thus study the problem of tool-use calibration in language agents, and identify prompt design and execution trace selection as two primary areas that suffer from miscalibration. We then propose ProbeCal, which recalibrates the internal probabilities of tool-using language agents to better reflect the actual effectiveness of tool, and enables a more appropriate selection of prompts and execution paths. We empirically show that ProbeCal can significantly and consistently improve off-the-shelf language models in tool-using applications.

Large Vision-Language Models (LVLMs) typically encode an image into a fixed number of visual tokens (e.g., 576) and process these tokens with a language model. Despite their strong performance, LVLMs face challenges in adapting to varying computational constraints. This raises the question: can we achieve flexibility in the number of visual tokens to suit different tasks and computational resources? We answer this with an emphatic yes. Inspired by Matryoshka Representation Learning, we introduce the Matryoshka Query Transformer (MQT), capable of encoding an image into $m$ visual tokens during inference, where $m$ can be any number up to a predefined maximum. This is achieved by employing a query transformer with $M$ latent query tokens to compress the visual embeddings. During each training step, we randomly select $m \leq M$ latent query tokens and train the model using only these first $m$ tokens, discarding the rest. Combining MQT with LLaVA, we train a single model once, and flexibly and drastically reduce the number of inference-time visual tokens while maintaining similar or better performance compared to training independent models for each number of tokens. Our model, MQT-LLaVA, matches LLaVA-1.5 performance across 11 benchmarks using a maximum of 256 tokens instead of LLaVA’s fixed 576. Reducing to 16 tokens (8x less TFLOPs) only sacrifices the performance by 2.4 points on MMBench. On certain tasks such as ScienceQA and MMMU, we can even go down to only 2 visual tokens with performance drops of just 3\% and 6\% each. Our exploration of the trade-off between the accuracy and computational cost brought about by the number of visual tokens facilitates future research to achieve the best of both worlds.

Large Vision-Language Models (LVLMs) suffer from hallucination issues, wherein the models generate plausible-sounding but factually incorrect outputs, undermining their reliability. A comprehensive quantitative evaluation is necessary to identify and understand the extent of hallucinations in these models. However, existing benchmarks are often limited in scope, focusing mainly on object hallucinations. Furthermore, current evaluation methods struggle to effectively address the subtle semantic distinctions between model outputs and reference data, as well as the balance between hallucination and informativeness. To address these issues, we introduce a multi-dimensional benchmark covering objects, attributes, and relations, with challenging images selected based on associative biases. Moreover, we propose a large language model (LLM)-based two-stage evaluation framework that generalizes the popular CHAIR metric and incorporates both faithfulness and coverage into the evaluation. Experiments on 10 established LVLMs demonstrate that our evaluation metric is more comprehensive and better correlated with humans than existing work when evaluating on our challenging human-annotated benchmark dataset. Our work also highlights the critical balance between faithfulness and coverage of model outputs, and encourages future works to address hallucinations in LVLMs while keeping their outputs informative.

Finetuning language agents with reasoning-action trajectories is effective, but obtaining these trajectories from human annotations or stronger models is costly and sometimes impractical. In this paper, we investigate the use of self-training in language agents, which can generate supervision from the agent itself, offering a promising alternative without relying on human or stronger model demonstrations. Self-training, however, requires high-quality model-generated samples, which are hard to obtain for challenging language agent tasks. To address this, we present Reflection-Reinforced Self-Training (Re-ReST), which uses a reflector to refine low-quality generated samples during self-training. The reflector takes the agent’s output and feedback from an external environment (e.g., unit test results in code generation) to produce improved samples. This technique enhances the quality of inferior samples and efficiently enriches the self-training dataset with higher-quality samples. We conduct extensive experiments on open-source language agents across tasks, including multi-hop question answering, sequential decision-making, code generation, visual question answering, and text-to-image generation. The results demonstrate the effectiveness of self-training and Re-ReST in language agent tasks, with self-training improving baselines by 7.6% on HotpotQA and 28.4% on AlfWorld, and Re-ReST further boosting performance by 2.0% and 14.1%, respectively. Our studies also confirm the efficiency of using a reflector to generate high-quality samples for self-training. Moreover, we demonstrate a method to employ reflection during inference without ground-truth feedback, addressing the limitation of previous reflection work.

Vision-language models have become increasingly powerful for tasks that require an understanding of both visual and linguistic elements, bridging the gap between these modalities. In the context of multimodal clinical AI, there is a growing need for models that possess domain-specific knowledge, as existing models often lack the expertise required for medical applications. In this paper, we take brain abnormalities as an example to demonstrate how to automatically collect medical image-text aligned data for pretraining from public resources such as PubMed. In particular, we present a pipeline that streamlines the pre-training process by initially collecting a large brain image-text dataset from case reports and published journals and subsequently constructing a high-performance vision-language model tailored to specific medical tasks. We also investigate the unique challenge of mapping subfigures to subcaptions in the medical domain. We evaluated the resulting model with quantitative and qualitative intrinsic evaluations. The resulting dataset and our code can be found here https://github.com/masoud-monajati/MedVL_pretraining_pipeline

We present EMBED (Egocentric Models Built with Exocentric Data), a method designed to transform exocentric video-language data for egocentric video representation learning. Large-scale exocentric data covers diverse activities with significant potential for egocentric learning, but inherent disparities between egocentric and exocentric data pose challenges in utilizing one view for the other seamlessly. Egocentric videos predominantly feature close-up hand-object interactions, whereas exocentric videos offer a broader perspective on human activities. Additionally, narratives in egocentric datasets are typically more action-centric and closely linked with the visual content, in contrast to the narrative styles found in exocentric datasets. To address these challenges, we employ a data transformation framework to adapt exocentric data for egocentric training, focusing on identifying specific video clips that emphasize hand-object interactions and transforming narration styles to align with egocentric perspectives. By applying both vision and language style transfer, our framework creates a new egocentric dataset derived from exocentric video-language data. Through extensive evaluations, we demonstrate the effectiveness of EMBED, achieving state-of-the-art results across various egocentric downstream tasks, including an absolute improvement of 4.7 Epic-Kitchens-100 multi-instance retrieval and 6.2 benchmarks in zero-shot settings. Furthermore, EMBED enables egocentric video-language models to perform competitively in exocentric tasks. Finally, we showcase EMBED's application across various exocentric datasets, exhibiting strong generalization capabilities when applied to different exocentric datasets.

We present X-Decoder, a generalized decoding model that can predict pixel-level segmentation and language tokens seamlessly. X-Decoder takes as input two types of queries: (i) generic non-semantic queries and (ii) semantic queries induced from text inputs, to decode different pixel-level and token-level outputs in the same semantic space. With such a novel design, X-Decoder is the first work that provides a unified way to support all types of image segmentation and a variety of vision-language (VL) tasks. Without any pseudo-labeling, our design enables seamless interactions across tasks at different granularities and brings mutual benefits by learning a common and rich pixel-level understanding. After pretraining on a mixed set of a limited amount of segmentation data and millions of image-text pairs, X-Decoder exhibits strong transferability to a wide range of downstream tasks in both zero-shot and finetuning settings. Notably, it achieves (1) state-of-the-art results on open-vocabulary segmentation and referring segmentation on seven datasets; (2) better or competitive finetuned performance to other generalist and specialist models on segmentation and VL tasks; and (3) flexibility for efficient fine-tuning and novel task composition (e.g., referring captioning and image editing shown in Fig. 1). Code, demo, video and visualization are available at: https://x-decoder-vl.github.io.

Model-based evaluation metrics (e.g., CLIPScore and GPTScore) have demonstrated decent correlations with human judgments in various language generation tasks. However, their impact on fairness remains largely unexplored. It is widely recognized that pretrained models can inadvertently encode societal biases, thus employing these models for evaluation purposes may inadvertently perpetuate and amplify biases. For example, an evaluation metric may favor the caption "a woman is calculating an account book" over "a man is calculating an account book," even if the image only shows male accountants. In this paper, we conduct a systematic study of gender biases in model-based automatic evaluation metrics for image captioning tasks. We start by curating a dataset comprising profession, activity, and object concepts associated with stereotypical gender associations. Then, we demonstrate the negative consequences of using these biased metrics, including the inability to differentiate between biased and unbiased generations, as well as the propagation of biases to generation models through reinforcement learning. Finally, we present a simple and effective way to mitigate the metric bias without hurting the correlations with human judgments. Our dataset and framework lay the foundation for understanding the potential harm of model-based evaluation metrics, and facilitate future works to develop more inclusive evaluation metrics.

Multimodal counterfactual reasoning is a vital yet challenging ability for AI systems. It involves predicting the outcomes of hypothetical circumstances based on vision and language inputs, which enables AI models to learn from failures and explore hypothetical scenarios. Despite its importance, there are only a few datasets targeting the counterfactual reasoning abilities of multimodal models. Among them, they only cover reasoning over synthetic environments or specific types of events (e.g. traffic collisions), making them hard to reliably benchmark the model generalization ability in diverse real-world scenarios and reasoning dimensions. To overcome these limitations, we develop a video question answering dataset, ACQUIRED: it consists of 3.9K annotated videos, encompassing a wide range of event types and incorporating both first and third-person viewpoints, which ensures a focus on real-world diversity. In addition, each video is annotated with questions that span three distinct dimensions of reasoning, including physical, social, and temporal, which can comprehensively evaluate the model counterfactual abilities along multiple aspects. We benchmark our dataset against several state-of-the-art language-only and multimodal models and experimental results demonstrate a significant performance gap (>13%) between models and humans. The findings suggest that multimodal counterfactual reasoning remains an open challenge and ACQUIRED is a comprehensive and reliable benchmark for inspiring future research in this direction.

Recent development in vision-language approaches has instigated a paradigm shift in learning visual recognition models from language supervision. These approaches align objects with language queries (e.g. "a photo of a cat") and improve the models' adaptability to identify novel objects and domains. Recently, several studies have attempted to query these models with complex language expressions that include specifications of fine-grained semantic details, such as attributes, shapes, textures, and relations. However, simply incorporating language descriptions as queries does not guarantee accurate interpretation by the models. In fact, our experiments show that GLIP, the state-of-the-art vision-language model for object detection, often disregards contextual information in the language descriptions and instead relies heavily on detecting objects solely by their names. To tackle the challenges, we propose a new description-conditioned (DesCo) paradigm of learning object recognition models with rich language descriptions consisting of two major innovations: 1) we employ a large language model as a commonsense knowledge engine to generate rich language descriptions of objects based on object names and the raw image-text caption; 2) we design context-sensitive queries to improve the model's ability in deciphering intricate nuances embedded within descriptions and enforce the model to focus on context rather than object names alone. On two novel object detection benchmarks, LVIS and OminiLabel, under the zero-shot detection setting, our approach achieves 34.8 APr minival (+9.1) and 29.3 AP (+3.6), respectively, surpassing the prior state-of-the-art models, GLIP and FIBER, by a large margin.

Vision-and-language navigation (VLN) agents are trained to navigate in real-world environments by following natural language instructions. A major challenge in VLN is the limited availability of training data, which hinders the models' ability to generalize effectively. Previous approaches have attempted to address this issue by introducing additional supervision during training, often requiring costly human-annotated data that restricts scalability. In this paper, we introduce a masked path modeling (MPM) objective, which pretrains an agent using self-collected data for downstream navigation tasks. Our proposed method involves allowing the agent to actively explore navigation environments without a specific goal and collect the paths it traverses. Subsequently, we train the agent on this collected data to reconstruct the original path given a randomly masked subpath. This way, the agent can actively accumulate a diverse and substantial amount of data while learning conditional action generation. To evaluate the effectiveness of our technique, we conduct experiments on various VLN datasets and demonstrate the versatility of MPM across different levels of instruction complexity. Our results exhibit significant improvements in success rates, with enhancements of 1.32\%, 1.05\%, and 1.19\% on the val-unseen split of the Room-to-Room, Room-for-Room, and Room-across-Room datasets, respectively. Furthermore, we conduct an analysis that highlights the potential for additional improvements when the agent is allowed to explore unseen environments prior to testing.

Vision-and-language (VL) pre-training has proven to be highly effective on various VL downstream tasks. While recent work has shown that fully transformer-based VL models can be more efficient than previous region-feature-based methods, their performance on downstream tasks often degrades significantly. In this paper, we present METER, a Multimodal End-to-end TransformER framework, through which we investigate how to design and pre-train a fully transformer-based VL model in an end-to-end manner. Specifically, we dissect the model designs along multiple dimensions: vision encoders (e.g., CLIP-ViT, Swin transformer), text encoders (e.g., RoBERTa, DeBERTa), multimodal fusion module (e.g., merged attention vs. co-attention), architectural design (e.g., encoder-only vs. encoder-decoder), and pre-training objectives (e.g., masked image modeling). We conduct comprehensive experiments and provide insights on how to train a performant VL transformer. METER achieves an accuracy of 77.64% on the VQAv2 test-std set using only 4M images for pre-training, surpassing the state-of-the-art region-feature-based model by 1.04%, and outperforming the previous best fully transformer-based model by 1.6%. Notably, when further scaled up, our best VQA model achieves an accuracy of 80.54%. Code and pre-trained models are released at https://github.com/zdou0830/METER.