Anand Rao

Gemma Team, Morgane Riviere, Shreya Pathak et al. · deepmind

In this work, we introduce Gemma 2, a new addition to the Gemma family of lightweight, state-of-the-art open models, ranging in scale from 2 billion to 27 billion parameters. In this new version, we apply several known technical modifications to the Transformer architecture, such as interleaving local-global attentions (Beltagy et al., 2020a) and group-query attention (Ainslie et al., 2023). We also train the 2B and 9B models with knowledge distillation (Hinton et al., 2015) instead of next token prediction. The resulting models deliver the best performance for their size, and even offer competitive alternatives to models that are 2-3 times bigger. We release all our models to the community.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Gemma Team, Aishwarya Kamath, Johan Ferret et al. · deepmind, mit

We introduce Gemma 3, a multimodal addition to the Gemma family of lightweight open models, ranging in scale from 1 to 27 billion parameters. This version introduces vision understanding abilities, a wider coverage of languages and longer context - at least 128K tokens. We also change the architecture of the model to reduce the KV-cache memory that tends to explode with long context. This is achieved by increasing the ratio of local to global attention layers, and keeping the span on local attention short. The Gemma 3 models are trained with distillation and achieve superior performance to Gemma 2 for both pre-trained and instruction finetuned versions. In particular, our novel post-training recipe significantly improves the math, chat, instruction-following and multilingual abilities, making Gemma3-4B-IT competitive with Gemma2-27B-IT and Gemma3-27B-IT comparable to Gemini-1.5-Pro across benchmarks. We release all our models to the community.

Andrew Sellergren, Sahar Kazemzadeh, Tiam Jaroensri et al.

Artificial intelligence (AI) has significant potential in healthcare applications, but its training and deployment faces challenges due to healthcare's diverse data, complex tasks, and the need to preserve privacy. Foundation models that perform well on medical tasks and require less task-specific tuning data are critical to accelerate the development of healthcare AI applications. We introduce MedGemma, a collection of medical vision-language foundation models based on Gemma 3 4B and 27B. MedGemma demonstrates advanced medical understanding and reasoning on images and text, significantly exceeding the performance of similar-sized generative models and approaching the performance of task-specific models, while maintaining the general capabilities of the Gemma 3 base models. For out-of-distribution tasks, MedGemma achieves 2.6-10% improvement on medical multimodal question answering, 15.5-18.1% improvement on chest X-ray finding classification, and 10.8% improvement on agentic evaluations compared to the base models. Fine-tuning MedGemma further improves performance in subdomains, reducing errors in electronic health record information retrieval by 50% and reaching comparable performance to existing specialized state-of-the-art methods for pneumothorax classification and histopathology patch classification. We additionally introduce MedSigLIP, a medically-tuned vision encoder derived from SigLIP. MedSigLIP powers the visual understanding capabilities of MedGemma and as an encoder achieves comparable or better performance than specialized medical image encoders. Taken together, the MedGemma collection provides a strong foundation of medical image and text capabilities, with potential to significantly accelerate medical research and development of downstream applications. The MedGemma collection, including tutorials and model weights, can be found at https://goo.gle/medgemma.

Yee-Yin Choong, Kristen Greene, Alice Qian et al.

AI measurement science has a wide variety of methodologies and measurements for comparing AI systems, resulting in what often appear to be "apples-to-oranges" comparisons across AI evaluations. To move toward "apples-to-apples" comparisons in real-world AI evaluations, this work advocates for methodological transparency in evaluation scenarios, operational grounding, and human-centered design (HCD) principles. We propose a repeatable process for transforming high-level use cases to detailed scenarios by eliciting use cases from subject matter experts (SMEs) via a structured AI Use Case Worksheet with six key elements: use case, sector, user (direct and indirect), intended outcomes, expected impacts (positive and negative), and KPIs and metrics. We demonstrate utility of the worksheet and process in the U.S. financial services sector. This paper reports on example high-level AI use cases identified by financial services sector SMEs: cyber defense enablement, developer productivity, financial crime aggregation, suspicious activity report (SAR) filing, credit memo generation, and internal call center support. These AI use cases provided are illustrative of the process and not exhaustive. Central to our work is a three-stage expansion pipeline combining LLM prompting with human reviews to generate 107 scenarios from those use cases elicited from SMEs. This process integrates iterative human reviews at every juncture to ensure operational grounding: for scenario titles and descriptions; for core scenario elements like users, benefits and risks, and metrics; and for scenario narratives and evaluation objectives. Human checkpoints ensure scenarios remain reflective of real-world usage and human needs. We describe a validation rubric to assess scenario quality. By defining key scenario components, this work supports a more consistent and meaningful paradigm for human-centered AI evaluations.

Prasang Gupta, Shaz Hoda, Anand Rao

Machine learning driven trading strategies have garnered a lot of interest over the past few years. There is, however, limited consensus on the ideal approach for the development of such trading strategies. Further, most literature has focused on trading strategies for short-term trading, with little or no focus on strategies that attempt to build long-term wealth. Our paper proposes a new approach for developing long-term investment strategies using an ensemble of evolutionary algorithms and a deep learning model by taking a series of short-term purchase decisions. Our methodology focuses on building long-term wealth by improving systematic investment planning (SIP) decisions on Exchange Traded Funds (ETF) over a period of time. We provide empirical evidence of superior performance (around 1% higher returns) using our ensemble approach as compared to the traditional daily systematic investment practice on a given ETF. Our results are based on live trading decisions made by our algorithm and executed on the Robinhood trading platform.

Saurabh Mishra, Mahendra Shinde, Aniket Yadav et al.

Infrastructure construction, often dubbed an "industry of industries," is closely linked with government spending and public procurement, offering significant opportunities for improved efficiency and productivity through better transparency and information access. By leveraging these opportunities, we can achieve notable gains in productivity, cost savings, and broader economic benefits. Our approach introduces an integrated software ecosystem utilizing Data Mesh and Service Mesh architectures. This system includes the largest training dataset for infrastructure and procurement, encompassing over 100 billion tokens, scientific publications, activities, and risk data, all structured by a systematic AI framework. Supported by a Knowledge Graph linked to domain-specific multi-agent tasks and Q&A capabilities, our platform standardizes and ingests diverse data sources, transforming them into structured knowledge. Leveraging large language models (LLMs) and automation, our system revolutionizes data structuring and knowledge creation, aiding decision-making in early-stage project planning, detailed research, market trend analysis, and qualitative assessments. Its web-scalable architecture delivers domain-curated information, enabling AI agents to facilitate reasoning and manage uncertainties, while preparing for future expansions with specialized agents targeting particular challenges. This integration of AI with domain expertise not only boosts efficiency and decision-making in construction and infrastructure but also establishes a framework for enhancing government efficiency and accelerating the transition of traditional industries to digital workflows. This work is poised to significantly influence AI-driven initiatives in this sector and guide best practices in AI Operations.

Saurabh Mishra, Anand Rao, Ramayya Krishnan et al.

As AI systems become integral to critical operations across industries and services, ensuring their reliability and safety is essential. We offer a framework that integrates established reliability and resilience engineering principles into AI systems. By applying traditional metrics such as failure rate and Mean Time Between Failures (MTBF) along with resilience engineering and human reliability analysis, we propose an integrate framework to manage AI system performance, and prevent or efficiently recover from failures. Our work adapts classical engineering methods to AI systems and outlines a research agenda for future technical studies. We apply our framework to a real-world AI system, using system status data from platforms such as openAI, to demonstrate its practical applicability. This framework aligns with emerging global standards and regulatory frameworks, providing a methodology to enhance the trustworthiness of AI systems. Our aim is to guide policy, regulation, and the development of reliable, safe, and adaptable AI technologies capable of consistent performance in real-world environments.

Amitoj Singh, Jingshu Chen, Lihao Zhang et al.

An independent ethical assessment of an artificial intelligence system is an impartial examination of the system's development, deployment, and use in alignment with ethical values. System-level qualitative frameworks that describe high-level requirements and component-level quantitative metrics that measure individual ethical dimensions have been developed over the past few years. However, there exists a gap between the two, which hinders the execution of independent ethical assessments in practice. This study bridges this gap and designs a holistic independent ethical assessment process for a text classification model with a special focus on the task of hate speech detection. The assessment is further augmented with protected attributes mining and counterfactual-based analysis to enhance bias assessment. It covers assessments of technical performance, data bias, embedding bias, classification bias, and interpretability. The proposed process is demonstrated through an assessment of a deep hate speech detection model.

Shaz Hoda, Amitoj Singh, Anand Rao et al.

The current pandemic has introduced substantial uncertainty to traditional methods for demand planning. These uncertainties stem from the disease progression, government interventions, economy and consumer behavior. While most of the emerging literature on the pandemic has focused on disease progression, a few have focused on consequent regulations and their impact on individual behavior. The contributions of this paper include a quantitative behavior model of fear of COVID-19, impact of government interventions on consumer behavior, and impact of consumer behavior on consumer choice and hence demand for goods. It brings together multiple models for disease progression, consumer behavior and demand estimation-thus bridging the gap between disease progression and consumer demand. We use panel regression to understand the drivers of demand during the pandemic and Bayesian inference to simplify the regulation landscape that can help build scenarios for resilient demand planning. We illustrate this resilient demand planning model using a specific example of gas retailing. We find that demand is sensitive to fear of COVID-19: as the number of COVID-19 cases increase over the previous week, the demand for gas decreases -- though this dissipates over time. Further, government regulations restrict access to different services, thereby reducing mobility, which in itself reduces demand.