google gemini

To address the dual challenge of achieving strong multimodal reasoning capabilities while enabling efficient on-device deployment, this paper introduces the Gemini family of large multimodal models—Gemini Ultra, Pro, and Nano. We propose a novel unified multimodal sequence modeling framework that integrates cross-modal joint representation learning, a hierarchical and scalable architecture, and responsible alignment techniques. Our models achieve human-expert-level performance on the MMLU benchmark—the first such result—and establish new state-of-the-art (SOTA) results on 30 out of 32 comprehensive evaluations, leading across 20 major multimodal benchmarks. Gemini Ultra significantly surpasses prior best methods on critical tasks including MMLU. The entire Gemini series has been deployed across Google AI Studio, Vertex AI, and Gemini Advanced, enabling seamless end-to-end deployment—from cloud-based complex reasoning to resource-constrained edge devices.

This study addresses the lack of structured, reproducible, and high-precision prompt engineering methodologies for Google Gemini 3 Pro Image, which leads to insufficient consistency and compliance in professionally generated images. To resolve this, we propose SCHEMA, a systematic prompting framework featuring a three-tier control system (BASE/MEDIO/AVANZATO), a modular tag architecture, and a compliance decision tree that enables fine-grained regulation of the image generation process. As the first domain-specific prompting framework tailored to this multimodal large language model, SCHEMA supports progressive controllability from 5% to 95% and includes strategies to mitigate model limitations with alternative pathways. Evaluated on 621 test cases using 850 API calls and 4,800 generated images, SCHEMA achieves 91% compliance on mandatory requirements and 94% on prohibited constraints, with over 95% first-generation compliance in information design tasks and significantly improved cross-domain image consistency in independent validation (n=40).

This work investigates how large language models—such as Gemini Deep Think—can effectively augment expert-driven creative research in mathematics and theoretical sciences. Through a series of interdisciplinary case studies, it introduces a collaborative framework centered on iterative refinement, problem decomposition, and knowledge transfer. The study pioneers a novel paradigm that positions the model as an “adversarial reviewer” and integrates it into a neuro-symbolic reasoning loop. By combining code generation, automated execution, and adversarial validation, the approach successfully resolves multiple open problems across theoretical computer science, economics, optimization, and physics. This represents the first systematic demonstration of the feasibility and practical efficacy of AI as a creative partner in expert-level scientific discovery.

This work addresses the joint advancement of multimodal understanding, long-context reasoning, and intelligent agent capabilities. Method: We propose the Gemini 2.X family of large language models, featuring the first unified integration of multimodal fusion, million-token-scale context modeling, and deep symbolic-neural hybrid reasoning—enabled by hierarchical compute scheduling and lightweight hardware adaptation for cross-platform efficiency. Contributions/Results: (1) We release three model variants—Pro, Flash, and Lite—spanning use cases from three-hour video comprehension to sub-millisecond response latency; (2) Gemini 2.5 Pro achieves state-of-the-art performance on MMLU, HumanEval, and other major benchmarks; (3) The Flash series attains >90% of Pro’s capability with <10% of its parameter count, substantially expanding the Pareto-optimal frontier. These advances accelerate the practical deployment of next-generation embodied intelligent agents.

Deploying physical agents for embodied intelligence remains challenged by multimodal understanding, environmental generalization, and task transfer. This paper introduces a family of Vision-Language-Action (VLA) foundation models specifically designed for robotics, proposing the first robot-native architecture—Gemini Robotics-ER—an embodied reasoning model that tightly integrates 3D perception, spatiotemporal modeling, and cross-view correspondence learning. The method enables end-to-end, reactive robotic manipulation, supporting open-vocabulary instruction following, zero-shot generalization to unseen environments and objects, embodiment-agnostic transfer, and rapid adaptation to new tasks with only ~100 demonstrations. Evaluated on diverse real-world robotic platforms, it achieves high success rates in complex, long-horizon dexterous manipulation, exhibits strong robustness in previously unobserved scenes, enables fast short-horizon task acquisition, and incorporates safety mechanisms to ensure reliable physical interaction.

This study addresses the lack of systematic evaluation criteria for large language model (LLM) selection by proposing a three-dimensional comparative framework encompassing performance, ethical behavior, and engineering usability. We conduct a qualitative and quantitative co-evaluation of five mainstream models—Claude, Gemini, DeepSeek, LLaMA, and ChatGPT. Innovatively, our unified benchmark integrates moral reasoning capability, factual accuracy, multimodal understanding, bias robustness, and API integration maturity. Results indicate that Claude achieves superior ethical reasoning; Gemini excels in multimodal processing and features a structured ethical framework; DeepSeek demonstrates exceptional factual consistency; LLaMA exhibits strong adaptability within open-source ecosystems; and ChatGPT attains the optimal balance between overall performance and user experience. The framework provides a reproducible, multi-dimensional empirical benchmark to guide principled LLM selection in both research and practice.

Large language models (LLMs) frequently fail in multimodal engineering tasks—such as circuit analysis—due to source polarity misidentification and reasoning hallucinations. Method: This paper introduces the first end-to-end AI solver system tailored for circuit analysis. It innovatively integrates YOLOv8 with OpenCV for high-accuracy component detection, and establishes an ngspice simulation–driven iterative LLM reasoning correction mechanism enabling human-in-the-loop verification. The system leverages Gemini 2.5 Pro for multimodal understanding and automated .cir netlist generation. Contribution/Results: Evaluated on 83 undergraduate-level circuit problems, the system achieves 97.59% accuracy—outperforming baseline methods by 18 percentage points. It significantly enhances the reliability and practical utility of AI tools in engineering education, addressing critical limitations in multimodal reasoning and domain-specific verification.

This work proposes a unified embedding model natively supporting arbitrary interleaved combinations of video, audio, image, and text modalities to construct a general-purpose multimodal representation for cross-modal and multimodal tasks. Built upon the Gemini architecture, the approach leverages large-scale contrastive learning, multi-task multi-stage training, and a unified embedding space to achieve zero-shot cross-domain generalization without fine-tuning. The model surpasses specialized architectures across diverse benchmarks, including MSCOCO (R@1=62.9), Vatex (NDCG@10=68.8), MTEB multilingual (69.9), and code retrieval (84.0), significantly enhancing downstream performance in retrieval, recommendation, and retrieval-augmented generation (RAG) tasks.

Large language models (LLMs) exhibit insufficient reasoning capability on International Mathematical Olympiad (IMO)-level problems. Method: This work proposes a fine-grained prompting framework tailored for formal mathematical reasoning, integrating chain-of-thought design, task-decomposition pipelines, and domain-adaptive prompting—implemented end-to-end on the unadulterated-training-data Gemini 2.5 Pro model without fine-tuning on competition problems or external tools. Contribution/Results: The core innovation lies in enhancing deep-reasoning stability solely through structural optimization of the reasoning process. Evaluated on all six problems from IMO 2025, the method achieves a 5/6 correctness rate—the first demonstration of a pure LLM approaching human gold-medalist performance without data contamination. This establishes a reproducible, interpretable paradigm for higher-order mathematical reasoning.

This work proposes a methodology for customizing large language models (LLMs) tailored to enterprise software engineering contexts, aiming to enhance developer productivity and code quality. By constructing a trillion-token-scale dataset comprising proprietary enterprise code and engineering artifacts, the approach integrates continual pretraining, post-training optimization, and an intermediate training strategy designed to mitigate catastrophic forgetting, thereby enabling deep adaptation to enterprise development environments. The end-to-end pipeline encompasses high-value signal extraction, data curation, and deployment. In a blind evaluation involving 29,000 developers, the customized model significantly reduced interactive iteration counts by 23% and improved code survival rates by approximately 17%, demonstrating its effectiveness and generalizability.