Phi-4
Phi-4 is Microsoft Research’s small-model masterpiece: 14 billion parameters yet competitive with much larger models on math and reasoning, trained largely on synthetic data.
Released
December 12, 2024
Type
llm
License
open-weight
Context
16,000 tokens
Capabilities
Architecture
Parameters: 14B
Links
Phi-4 in the news
OpenAI · Sep 8, 2026
On the Navier–Stokes Millennium Prize Problem
An artificial intelligence system has generated a solution to the Navier-Stokes Millennium Prize Problem, a long-standing challenge in mathematics and physics. The AI's output includes a detailed writeup of the solution and a formal proof formalized in the Lean theorem prover. This development marks a significant moment in the application of AI to complex scientific and mathematical problems, potentially accelerating research and discovery in fields reliant on fluid dynamics. The Navier-Stokes equations describe the motion of viscous fluid substances, such as air, water, and blood. They are fundamental to understanding a vast array of phenomena, from weather patterns and ocean currents to the aerodynamics of aircraft and the flow of blood in arteries. Despite their importance, proving the existence and smoothness of solutions to these equations in three dimensions has remained an unsolved problem for over a century. The Clay Mathematics Institute designated it as one of seven Millennium Prize Problems in 2000, offering a $1 million prize for a correct solution. The AI's contribution, as detailed in its accompanying writeup, aims to address the existence and smoothness of solutions. Formalizing the proof in Lean is a critical step, as it allows for rigorous verification by human mathematicians. Lean is an interactive theorem prover and programming language developed by Leonardo de Moura at Microsoft Research, designed to assist in the formal verification of mathematical proofs and software. Its use signifies a commitment to mathematical rigor and allows the broader scientific community to scrutinize the AI's findings. While the AI has produced the solution and proof, its validity is now subject to review by experts in the field. The process of peer review for such a complex mathematical problem, especially one generated by an AI, is expected to be thorough and potentially lengthy. The implications of a confirmed AI-generated solution could be profound, not only for fluid dynamics but also for the broader landscape of AI-assisted scientific research. It could pave the way for AI to tackle other grand challenges in mathematics and physics, potentially leading to breakthroughs that have eluded human researchers for decades. The Clay Mathematics Institute will ultimately determine if the provided solution meets the criteria for the Millennium Prize.
Neil Patel · Aug 12, 2026
Large Language Model (LLM) Editing Quietly Corrupts Documents. Here’s What the Research Says
A Microsoft Research study published on April 17, 2026, has revealed that large language models (LLMs) can subtly and systematically corrupt documents during extended editing workflows. The DELEGATE-52 study tested 19 LLMs across 52 professional domains over 20 editing interactions, simulating real-world document work. The research found that even advanced frontier LLMs, including Gemini 3.1 Pro, Claude 4.6 Opus, and GPT 5.4, corrupted an average of 25 percent of document content by the end of these long editing sessions. Across all 19 LLMs tested, the average degradation reached a significant 50 percent. The errors introduced by LLMs are described as sparse but severe, consisting of a small number of consequential changes that are grammatically correct, making them difficult to detect during a casual review. This contrasts with typical hallucinations or minor typos, as these errors are subtle enough to pass initial scrutiny but damaging enough to impact the document's integrity and can compound over multiple editing sessions. The study's methodology involved providing LLMs with professional documents from diverse fields such as coding, crystallography, music notation, accounting records, and recipes. These domains encompassed both highly structured formats like code and database schemas, and natural language writing, demonstrating that corruption occurred across various types of content. The DELEGATE-52 study specifically focused on multi-session workflows, where an LLM handles a continuous sequence of revisions and refinements, rather than isolated, one-off edits. This approach aimed to mimic how users might employ LLMs for ongoing document development. The researchers also investigated the impact of providing LLMs with basic agentic harnesses and file tools. Counterintuitively, this slightly worsened performance, leading to approximately 6 percent more degradation while consuming two to five times more input tokens. Python emerged as the only domain where most LLMs managed to clear the study's stringent 98 percent accuracy threshold. However, even the best-performing model achieved this benchmark in only 11 out of the 52 domains tested, highlighting the widespread nature of the corruption issue. These findings suggest a need for content teams to re-evaluate the integration of AI tools into their editing processes, particularly for tasks involving iterative document refinement.
Search Engine Land · Jul 1, 2026
GraphRAG: What entity-first retrieval means for SEO
GraphRAG, a retrieval method developed by Microsoft Research in 2024, enhances traditional retrieval-augmented generation (RAG) by incorporating knowledge graphs. This approach allows AI systems to better identify entities, understand their interconnections, and cite sources more reliably. Unlike standard RAG, which processes information as flat text, GraphRAG constructs a "map" where entities like companies, products, and individuals are nodes, and their relationships (e.g., "offers," "is certified by") are edges. This graph-based retrieval enables AI models to navigate and connect facts with greater confidence, leading to more complete and grounded answers while reducing hallucinations. The system follows defined paths within the knowledge graph to retrieve information, rather than inferring it from unstructured text. This structured approach is crucial for AI systems to accurately answer complex queries and understand nuanced relationships between different pieces of information. Microsoft's patent application, "Knowledge Graph Extraction" (US20250131289A1), details the limitations of naive RAG, specifically the "recall problem" where less prominent entities can be lost in text embeddings. GraphRAG addresses this through entity resolution, a process that merges duplicate spellings or variations of the same entity, ensuring they are treated as a single unit. This foundational capability is key to the effectiveness of graph-based retrieval systems, improving their ability to recall and synthesize information accurately.