Recent advances in video-language models have enabled applications like video retrieval, captioning, and assembly. However, executing such multi-stage pipelines efficiently on mobile devices remains challenging due to redundant model loads and fragmented execution. We introduce Atom, an on-device system that restructures video-language pipelines for fast and efficient execution. Atom decomposes a billion-parameter model into reusable modules, such as the visual encoder and language decoder, and parallely reuses them across subtasks like captioning, reasoning, and indexing. This reuse-centric design eliminates repeated model loading and enables parallel execution, reducing end-to-end latency without sacrificing performance. On commodity smartphones, Atom achieves 27-33% faster execution compared to non-reuse baselines, with only marginal performance drop (<= 2.3 Recall@1 in retrieval, <= 1.5 CIDEr in captioning). These results position Atom as a practical, scalable approach for efficient video-language understanding on edge devices.

Recent work shows that pre-trained, instruction-tuned vision-language models (VLMs) perform well at mapping from instructions and observations to high-level plans, but struggle to realize such plans as reliable low-latency action sequences in unfamiliar environments. At the same time, world-model controllers excel at fast observation-to-action control, but lack open-ended task guidance. In this work, we combine these strengths into a single system, Instruct-to-Act, where we train a world-model controller to act autonomously at high frequency when conditioned on sparse, higher-latency, and high-level text instructions generated by a VLM planner. To train controllers to be language-instructable, we relabel segments of controller policy rollouts with synthetic instructions and jointly optimize a behavior-cloning objective along with existing reward-maximizing and world-modeling objectives. We evaluate our proposed approach across seven embodied environments, including three multi-agent environments where VLM planners coordinate through language while trained controllers serve as their actuators. Under matched observation and action spaces, our decoupled approach consistently outperforms controller-only and direct VLM action-generation variants, preserves fast control, and lets us swap in different pretrained VLM planners without fine-tuning, while remaining competitive with strong vision-language-action and multi-agent RL baselines on six of seven tasks.

Language models (LMs) are used for a diverse range of tasks, from question answering to writing fantastical stories. In order to reliably accomplish these tasks, LMs must be able to discern the modal category of a sentence (i.e., whether it describes something that is possible, impossible, completely nonsensical, etc.). However, recent studies have called into question the ability of LMs to categorize sentences according to modality (Michaelov et al., 2025; Kauf et al., 2023). In this work, we identify linear representations that discriminate between modal categories within a variety of LMs, or modal difference vectors. Analysis of modal difference vectors reveals that LMs have access to more reliable modal categorization judgments than previously reported. Furthermore, we find that modal difference vectors emerge in a consistent order as models become more competent (i.e., through training steps, layers, and parameter count). Notably, we find that modal difference vectors identified within LM activations can be used to model fine-grained human categorization behavior. This potentially provides a novel view into how human participants distinguish between modal categories, which we explore by correlating projections along modal difference vectors with human participants' ratings of interpretable features. In summary, we derive new insights into LM modal categorization using techniques from mechanistic interpretability, with the potential to inform our understanding of modal categorization in humans.

A fundamental part of human cognition is the ability to not only recall previous memories, but also reason across them to draw conclusions. In cognitive science and psychology, this is termed relational reasoning and a number of effects and biases have been observed in human cognition. Designing experiments to measure these reasoning effects is effortful and does not transfer easily to analyzing language model reasoning patterns. To make exploring language models on relational reasoning easier, we introduce ReCogLab – a generative framework for constructing reasoning examples. Unlike static datasets, our framework has a number of benefits that help us in our goal of flexible evaluation of LLMs. First, our framework allows us to control the difficulty and context-length of the problem, allowing us to scale with model capability and evaluate LLMs at a variety of scales. Second, the ability to change the configuration of a dataset dynamically allows us to probe models on different aspects and capabilities. Finally, the flexibility of our approach enables the recreation of classic cognitive science experiments and the systematic study of relational reasoning biases in language models. We demonstrate several such experiments and present our findings on a wide variety of open and closed-source language models. We release all data and code at https://github.com/google-deepmind/recoglab.

Inductive biases are what allow learners to make guesses in the absence of conclusive evidence. These biases have often been studied in cognitive science using concepts or categories – e.g. by testing how humans generalize a new category from a few examples that leave the category boundary ambiguous. We use these approaches to study generalization in foundation models during in-context learning. Modern foundation models can condition on both vision and text, and differences in how they interpret and learn from these different modalities is an emerging area of study. Here, we study how their generalizations vary by the modality in which stimuli are presented, and the way the stimuli are described in text. We study these biases with three different experimental paradigms, across three different vision-language models. We find that the models generally show some bias towards generalizing according to shape over color. This shape bias tends to be amplified when the examples are presented visually. By contrast, when examples are presented in text, the ordering of adjectives affects generalization. However, the extent of these effects vary across models and paradigms. These results help to reveal how vision-language models represent different types of inputs in context, and may have practical implications for the use of vision-language models.

We present SKALD, a multi-shot video assembly method that constructs coherent video sequences from candidate shots with minimal reliance on text. Central to our approach is the Learned Clip Assembly (LCA) score, a learning-based metric that measures temporal and semantic relationships between shots to quantify narrative coherence. We tackle the exponential complexity of combining multiple shots with an efficient beam-search algorithm guided by the LCA score. To train our model effectively with limited human annotations, we propose two tasks for the LCA encoder: Shot Coherence Learning, which uses contrastive learning to distinguish coherent and incoherent sequences, and Feature Regression, which converts these learned representations into a real-valued coherence score. We develop two variants: a base SKALD model that relies solely on visual coherence and SKALD-text, which integrates auxiliary text information when available. Experiments on the VSPD and our curated MSV3C datasets show that SKALD achieves an improvement of up to 48.6

Understanding geometry relies heavily on vision. In this work, we evaluate whether state-of-the-art vision language models (VLMs) can understand simple geometric concepts. We use a paradigm from cognitive science that isolates visual understanding of simple geometry from the many other capabilities it is often conflated with such as reasoning and world knowledge. We compare model performance with human adults from the USA, as well as with prior research on human adults without formal education from an Amazonian indigenous group. We find that VLMs consistently underperform both groups of human adults, although they succeed with some concepts more than others. We also find that VLM geometric understanding is more brittle than human understanding, and is not robust when tasks require mental rotation. This work highlights interesting differences in the origin of geometric understanding in humans and machines – e.g. from printed materials used in formal education vs. interactions with the physical world or a combination of the two – and a small step toward understanding these differences.

Today, short-videos dominate social media, yet short-video creators lack systematic tools to predict engagement and refine content before uploading. Existing approaches focus primarily on post-publication metrics, failing to address engagement prediction via video editing elements. To address this, we curate VidES, a novel dataset linking short-vdieo engagement to specific editing elements, referred to as Edit Signals (e.g.: narration, text overlays). VidES contains overall engagement scores and detailed human evaluations of Edit Signals. Building on VidES, we propose SmartEdit, a multi-stage framework that predicts short-video engagement and generates actionable improvement suggestions based on Edit Signals. SmartEdit decomposes engagement prediction into interpretable components and uses few-shot Multi-modal Large Language Models to provide specific feedback for content refinement. Experimental results validate SmartEdit’s effectiveness, demonstrating its superiority over zero-shot methods. This work bridges the gap between intuitive and data-driven short-video editing, offering the first dataset and tool tailored for Edit Signal-based short-video refinement.

Synthetic text generated by Large Language Models (LLMs) is increasingly used for further training and improvement of LLMs. Diversity is crucial for the effectiveness of synthetic data, and researchers rely on prompt engineering to improve diversity. However, the impact of prompt variations on response text length, and, more importantly, the consequential effect on lexical diversity measurements, remain underexplored. In this work, we propose Penalty-Adjusted Type-Token Ratio (PATTR), a diversity metric robust to length variations. We generate a large synthetic corpus of over 20M words using seven models from the LLaMA, OLMo, and Phi families, focusing on a creative writing task of video script generation, where diversity is crucial. We evaluate per-response lexical diversity using PATTR and compare it against existing metrics of Moving-Average TTR (MATTR) and Compression Ratio (CR). Our analysis highlights how text length variations introduce biases favoring shorter responses. Unlike existing metrics, PATTR explicitly considers the task-specific target response length (L_T) to effectively mitigate length biases. We further demonstrate the utility of PATTR in filtering the top-10/100/1,000 most lexically diverse responses, showing that it consistently outperforms MATTR and CR by yielding on par or better diversity with high adherence to L_T.

Data attribution methods quantify the influence of training data on model outputs and are becoming increasingly relevant for a wide range of LLM research and applications, including dataset curation, model interpretability, data valuation. However, there remain critical gaps in systematic LLM-centric evaluation of data attribution methods. To this end, we introduce DATE-LM (Data Attribution Evaluation in Language Models), a unified benchmark for evaluating data attribution methods through real-world LLM applications. DATE-LM measures attribution quality through three key tasks – training data selection, toxicity/bias filtering, and factual attribution. Our benchmark is designed for ease of use, enabling researchers to configure and run large-scale evaluations across diverse tasks and LLM architectures. Furthermore, we use DATE-LM to conduct a large-scale evaluation of existing data attribution methods. Our findings show that no single method dominates across all tasks, data attribution methods have trade-offs with simpler baselines, and method performance is sensitive to task-specific evaluation design. Finally, we release a public leaderboard for quick comparison of methods and to facilitate community engagement. We hope DATE-LM serves as a foundation for future data attribution research in LLMs.

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99 as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75 categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

reasoning is a key ability for an intelligent system. Large language models (LMs) achieve above-chance performance on abstract reasoning tasks but exhibit many imperfections. However, human abstract reasoning is also imperfect. Human reasoning is affected by our real-world knowledge and beliefs, and shows notable "content effects"; humans reason more reliably when the semantic content of a problem supports the correct logical inferences. These content-entangled reasoning patterns are central to debates about the fundamental nature of human intelligence. Here, we investigate whether language models-whose prior expectations capture some aspects of human knowledge-similarly mix content into their answers to logic problems. We explored this question across three logical reasoning tasks: natural language inference, judging the logical validity of syllogisms, and the Wason selection task. We evaluate state of the art LMs, as well as humans, and find that the LMs reflect many of the same qualitative human patterns on these tasks-like humans, models answer more accurately when the semantic content of a task supports the logical inferences. These parallels are reflected in accuracy patterns, and in some lower-level features like the relationship between LM confidence over possible answers and human response times. However, in some cases the humans and models behave differently-particularly on the Wason task, where humans perform much worse than large models, and exhibit a distinct error pattern. Our findings have implications for understanding possible contributors to these human cognitive effects, as well as the factors that influence language model performance.

With the continuous advancement of large language models (LLMs), it is essential to create new benchmarks to effectively evaluate their expanding capabilities and identify areas for improvement. This work focuses on multi-image reasoning, an emerging capability in state-of-the-art LLMs. We introduce ReMI, a dataset designed to assess LLMs' ability to Reason with Multiple Images. This dataset encompasses a diverse range of tasks, spanning various reasoning domains such as math, physics, logic, code, table/chart understanding, and spatial and temporal reasoning. It also covers a broad spectrum of characteristics found in multi-image reasoning scenarios. We have benchmarked several cutting-edge LLMs using ReMI and found a substantial gap between their performance and human-level proficiency. This highlights the challenges in multi-image reasoning and the need for further research. Our analysis also reveals the strengths and weaknesses of different models, shedding light on the types of reasoning that are currently attainable and areas where future models require improvement. To foster further research in this area, we are releasing ReMI publicly: https://huggingface.co/datasets/mehrankazemi/ReMI.

Shannon and Weaver's seminal information theory divides communication into three levels: technical, semantic, and effectiveness. While the technical level deals with the accurate reconstruction of transmitted symbols, the semantic and effectiveness levels deal with the inferred meaning and its effect on the receiver. Large Language Models (LLMs), with their wide generalizability, make some progress towards the second level. However, LLMs and other communication models are not conventionally designed for predicting and optimizing communication for desired receiver behaviors and intents. As a result, the effectiveness level remains largely untouched by modern communication systems. In this paper, we introduce the receivers' "behavior tokens," such as shares, likes, clicks, purchases, and retweets, in the LLM's training corpora to optimize content for the receivers and predict their behaviors. Other than showing similar performance to LLMs on content understanding tasks, our trained models show generalization capabilities on the behavior dimension for behavior simulation, content simulation, behavior understanding, and behavior domain adaptation. We show results on all these capabilities using a wide range of tasks on three corpora. We call these models Large Content and Behavior Models (LCBMs). Further, to spur more research on LCBMs, we release our new Content Behavior Corpus (CBC), a repository containing communicator, message, and corresponding receiver behavior (https://behavior-in-the-wild.github.io/LCBM).

To process novel sentences, language models (LMs) must generalize compositionally -- combine familiar elements in new ways. What aspects of a model's structure promote compositional generalization? Focusing on transformers, we test the hypothesis, motivated by theoretical and empirical work, that deeper transformers generalize more compositionally. Simply adding layers increases the total number of parameters; to address this confound between depth and size, we construct three classes of models which trade off depth for width such that the total number of parameters is kept constant (41M, 134M and 374M parameters). We pretrain all models as LMs and fine-tune them on tasks that test for compositional generalization. We report three main conclusions: (1) after fine-tuning, deeper models generalize more compositionally than shallower models do, but the benefit of additional layers diminishes rapidly; (2) within each family, deeper models show better language modeling performance, but returns are similarly diminishing; (3) the benefits of depth for compositional generalization cannot be attributed solely to better performance on language modeling. Because model latency is approximately linear in the number of layers, these results lead us to the recommendation that, with a given total parameter budget, transformers can be made shallower than is typical without sacrificing performance.

Vision language models (VLMs) have shown impressive capabilities across a variety of tasks, from logical reasoning to visual understanding. This opens the door to richer interaction with the world, for example robotic control. However, VLMs produce only textual outputs, while robotic control and other spatial tasks require outputting continuous coordinates, actions, or trajectories. How can we enable VLMs to handle such settings without fine-tuning on task-specific data? In this paper, we propose a novel visual prompting approach for VLMs that we call Prompting with Iterative Visual Optimization (PIVOT), which casts tasks as iterative visual question answering. In each iteration, the image is annotated with a visual representation of proposals that the VLM can refer to (e.g., candidate robot actions, localizations, or trajectories). The VLM then selects the best ones for the task. These proposals are iteratively refined, allowing the VLM to eventually zero in on the best available answer. We investigate PIVOT on real-world robotic navigation, real-world manipulation from images, instruction following in simulation, and additional spatial inference tasks such as localization. We find, perhaps surprisingly, that our approach enables zero-shot control of robotic systems without any robot training data, navigation in a variety of environments, and other capabilities. Although current performance is far from perfect, our work highlights potentials and limitations of this new regime and shows a promising approach for Internet-Scale VLMs in robotic and spatial reasoning domains. Website: pivot-prompt.github.io and HuggingFace: https://huggingface.co/spaces/pivot-prompt/pivot-prompt-demo.

Recent advances in multimodal models show remarkable performance in real-world benchmarks for chart and figure understanding like ChartQA that involve interpreting trends, comparing data points, and extracting insights from visuals.In this paper, we investigate the extent to which these models truly comprehend the underlying information in charts by posing direct, elementary questions about simple features such as axes ranges and values to examine their fundamental visual understanding abilities in the context of charts.Our questions are applied to two sets of figures: synthetic and real-world.The empirical evaluation of 5 popular multimodal models on our dataset reveals shortfalls in understanding charts and figures, contrary to what their performance on complex benchmarks might suggest.For instance, Gemini Pro Vision only achieves 57.9% accuracy on our elementary set of questions on real-world plots, while other popular multimodal models showed similar or less performance.This work highlights an important limitation of current multimodal models, and cautions against overly optimistic interpretations of their abilities based on results of canonical evaluations.

We are encouraged by the many positive commentaries on our target article. In this response, we recapitulate some of the points raised and identify synergies between them. We have arranged our response based on the tension between data and architecture that arises in the meta-learning framework. We additionally provide a short discussion that touches upon connections to foundation models.

Text-to-image generative models can generate high-quality humans, but realism is lost when generating hands. Common artifacts include irregular hand poses, shapes, incorrect numbers of fingers, and physically implausible finger orientations. To generate images with realistic hands, we propose a novel diffusion-based architecture called HanDiffuser that achieves realism by injecting hand embeddings in the generative process. HanDiffuser consists of two components: a Text-to-Hand-Params diffusion model to generate SMPL-Body and MANO-Hand parameters from input text prompts, and a Text-Guided Hand-Params-to-Image diffusion model to synthesize images by conditioning on the prompts and hand parameters generated by the previous component. We incorporate multiple aspects of hand representation, including 3D shapes and joint-level finger positions, orientations and articulations, for robust learning and reliable performance during inference. We conduct extensive quantitative and qualitative experiments and perform user studies to demonstrate the efficacy of our method in generating images with high-quality hands.