Test-Time Scaling in Multimodal Foundation Models: A Comprehensive Survey of Generation and Reasoning

A team of researchers has proposed the first unified taxonomic framework for test-time scaling in multimodal foundation models, organizing a rapidly expanding field into three distinct methodological categories. The survey, submitted in 2026, aims to provide a systematic roadmap for a domain that has until now lacked a comprehensive theoretical structure [1]. Test-time scaling (TTS) refers to the practice of dynamically allocating computational resources during a model's inference phase to improve performance, rather than solely relying on resources expended during initial training [1][6]. While the technique has been widely explored in large language models, its application to multimodal foundation models (MFMs) — systems that process text, images, audio, and other data types — has recently accelerated, unlocking new capabilities in reasoning and generation [1][7]. The new survey, published on arXiv, categorizes these TTS methodologies for MFMs into three strategies: sampling-based, feedback-based, and search-based approaches [1][3]. Sampling-based methods generate multiple outputs and select the best candidate. Feedback-based approaches use internal or external critics to iteratively refine results. Search-based strategies systematically explore a space of possible reasoning steps, often using tree-search or beam-search techniques [3][5]. The framework distinguishes TTS from related concepts like model adaptation or memory augmentation, highlighting the unique challenges posed by multimodal data, such as coordinating reasoning across visual and textual modalities [3]. Prior surveys have focused predominantly on TTS for text-only large language models, leaving a gap in the literature for multimodal systems [3]. The authors of the survey summarize representative applications and benchmarks used to evaluate these capabilities. Benchmarks are standardized tests that measure model performance on specific tasks, such as visual question answering or instruction following, and are critical for comparing different approaches [8]. One recent framework, UniT, demonstrated the efficiency of sequential, chain-of-thought scaling over parallel sampling, achieving comparable performance with 2.5 times less computational cost [4]. UniT also showed that models trained on shorter reasoning trajectories could generalize to longer inference chains, a behavior previously observed only in text-only models [4]. The survey further identifies open challenges for the field, including the mitigation of hallucination — where models generate plausible but incorrect information — and the problem of error accumulation across multiple reasoning steps [3]. The authors outline future research directions such as hybrid scaling strategies that combine multiple TTS methods, aiming to build more efficient and robust multimodal systems [3].