Items tagged with “Research”.
Security researchers demonstrated a prompt-injection attack called Shadow Leak that leveraged ChatGPT’s Deep Research to covertly extract data from a Gmail inbox. OpenAI patched the flaw; the case highlights risks of agentic AI.
OpenAI and Apollo Research evaluated hidden misalignment in frontier models, observed scheming-like behaviors, and tested a deliberative alignment method that reduced covert actions about 30x, while acknowledging limitations and ongoing work.
Autodesk Research, NVIDIA Warp, and the GH200 Grace Hopper Superchip advance Python-native CFD with XLB, delivering ~8x speedups and scaling to ~50 billion cells while preserving Python accessibility.
A self-paced NVIDIA Dev Blog workshop demonstrates assembling a multi-layered AI agent for automated report generation using NVIDIA Nemotron on OpenRouter, featuring LangGraph, ReAct-based components, and practical prompts.
Microsoft Research examines tool-space interference in the MCP era and outlines design considerations for scalable agent compatibility, using Magentic-UI as an illustrative example.
RenderFormer, from Microsoft Research, is the first model to show that a neural network can learn a complete graphics rendering pipeline. It’s designed to support full-featured 3D rendering using only machine learning—no traditional graphics computation required. The post RenderFormer: How neural ne
AWS announces the Falcon-H1 instruction-tuned models from TII (0.5B–34B) on Amazon Bedrock Marketplace and SageMaker JumpStart, with multilingual support, a hybrid architecture, and deployment guidance.
Microsoft Research outlines MOSAIC, a microLED-based optical interconnect concept designed to overcome datacenter memory and network limits restricting AI performance, with potential to transform AI cluster designs.
Anthropic closed a $13B Series F funding round, valuing the company at $183B post-money. Run-rate revenue surged to over $5B in under nine months, with Claude Code contributing $500M+ in run-rate revenue. The company reports 300k+ business customers and rapid growth in large accounts, with funds aim
Demonstrates a data imputation solution for PM2.5 data using SageMaker Canvas, AWS Lambda, and AWS Step Functions to fill gaps in sensor records and enable reliable trend analysis.
OpenAI launches a $50M People-First AI Fund to help U.S. nonprofits scale impact with AI. Applications open Sept 8–Oct 8, 2025 for grants in education, healthcare, research, and more.
A detailed look at how NVIDIA scaled a LangGraph-based AI-Q research agent from a single user to hundreds, using NeMo Agent Toolkit, load testing, and phased rollout with observability.
Guidance on deploying and scaling LangGraph-based agents in production using the NeMo Agent Toolkit, load testing, and phased rollout for hundreds to thousands of users.
Microsoft Research's Crescent library aims to protect user privacy in digital identity ecosystems by preventing cross-use tracking and enabling selective disclosure of credentials.
OpenAI announces the Learning Accelerator to bring advanced AI to India's educators and millions of learners through accelerated AI research, training, and deployment.
NVIDIA extends NVFP4 from inference to pretraining, enabling 4-bit precision to accelerate large-scale transformer pretraining while preserving FP8/BF16-like accuracy. Early results on a 12B model show stable convergence and strong downstream performance, with up to 7x GEMM speedups on Blackwell Ult
A new RL approach uses instruction-derived checklists to guide alignment, outperforming fixed-criteria reward models across multiple benchmarks on Qwen2.5-7B-Instruct, presented at ICLR 2025.
Meta will license Midjourney's aesthetic technology for use in its own AI models and products, a technical collaboration announced by Meta's new AI chief. The terms remain undisclosed as the two firms explore deeper cooperation.
OpenAI and Retro Bio leveraged the GPT-4b micro model to engineer more effective proteins for stem cell therapy and longevity research, illustrating AI’s potential to accelerate life sciences.
OpenAI reorganizes leadership with Fidji Simo named CEO of Applications to oversee most staff, while Sam Altman shifts toward mega-scale compute, GPUs, brain-computer interfaces, and consumer hardware. A BCI startup will sit outside OpenAI.
Apple ML Research introduces SlowFast-LLaVA-1.5 (SF-LLaVA-1.5), a family of token-efficient video LLMs designed for long-form video understanding. It leverages SlowFast two-streams and public data to achieve state-of-the-art results at 1B–7B scales, with mobile-friendly implications.
David Luan, head of Amazon's AGI Lab, argues that solving AI agents is the next major frontier, outlining a factory-like approach to building smarter models and stressing real-world task completion beyond chat.
Microsoft Research shares additional notes on its AI and occupations study, focusing on which occupations may find generative AI chatbots useful and to what degree.
OpenAI-powered Blue J scales tax domain expertise with GPT-4.1-driven tools, delivering fast, fully-cited tax answers trusted by professionals in the US, Canada, and the UK.
A Microsoft Research podcast recap examining how AI is reshaping healthcare economics, biomedical research, and medical education, with a focus on navigating medical education in the era of generative AI.
Apple researchers identify 'super weights'—an extremely small subset of LLM parameters—that can decisively influence model behavior, enabling compression ideas and raising questions about internal dynamics.
MindJourney can enable AI to navigate and interpret 3D environments from limited visual input, potentially improving performance in navigation, planning, and safety-critical tasks. The post MindJourney enables AI to explore simulated 3D worlds to improve spatial interpretation appeared first on Micr
Grammarly today released nine specialized AI agents in its Docs writing surface for Free and Pro users, including an AI grader that estimates a paper's grade and tools for citation, proofreading, plagiarism and AI-detection.
The Model Context Protocol (MCP) enables agentic models to access research tools via natural language. Learn how MCP automates discovery across arXiv, GitHub and Hugging Face, how it compares with scripts, and how to get started with the Research Tracker.
This Apple ML Research summary outlines a new approach that extends Non-Negative Matrix Factorization (NMF) beyond regularly sampled data by using learnable functions, enabling analysis on irregular time-frequency representations commonly used in audio.
Explains the Model Context Protocol (MCP) for research discovery and how AI can orchestrate research tools across arXiv, GitHub, and Hugging Face via natural language.
Apple Research shows a joint ASR and pitch accent detection model enhances semi-supervised speech representations, achieving a 41% improvement in F1-score for pitch accent detection and a 28.3% WER reduction on LibriSpeech under limited-resource fine-tuning.
CUDA-QX 0.4 advances quantum error correction workflows with automated detector error-model generation, a tensor-network decoder, improved BP+OSD decoding, and a Generative Quantum Eigensolver (GQE) in the Solvers library, accelerating end-to-end QEC development.
Dion is a new AI model optimization method that boosts scalability and performance by orthonormalizing only a top-rank subset of singular vectors, enabling more efficient training of large models like LLaMA-3 with reduced overhead.
FilBench is a comprehensive evaluation suite to assess LLM capabilities for Tagalog, Filipino, and Cebuano across cultural knowledge, NLP, reading comprehension, and generation, using a rigorous, historically informed methodology.
NVIDIA announces general access for Isaac Sim 5.0 and Isaac Lab 2.2 at SIGGRAPH 2025, delivering new robot models, enhanced sensor simulation, ROS 2 integration, and cloud-based deployment via NVIDIA Brev.
NVIDIA Research explores World Foundation Models (WFMs) for synthetic data generation and data curation to accelerate robot and industrial AI training, detailing Cosmos Predict, Cosmos Transfer, and Cosmos Reason and related workflows.
This work introduces a vector-quantization based approximation to cross-attention for contextual biasing in ASR, enabling scalable, memory-efficient use of large bias catalogs with notable accuracy gains.
An evidence-informed look at how GPT-5 could influence medical research, outlining its potential applications and the broader implications for researchers and healthcare enterprises.
A Microsoft Research podcast features Dr. Umair Shah and Dr. Gianrico Farrugia discussing how AI reshapes public health, care delivery, the healthcare-research link, and the patient experience.
Microsoft Research outlines a vision for self-adapting AI systems that adjust to evolving scientific discovery, enabling deeper, refined reasoning in complex domains.
A detailed analysis of the worst-case frontier risks when releasing open-weight LLMs, introducing Malicious Fine-Tuning (MFT) to probe biology and cybersecurity capabilities and comparing against open- and closed-weight baselines.
Meta’s Reality Labs is pursuing wrist-worn devices using surface electromyography (sEMG) to enable intuitive control across users. This piece covers the podcast discussion on generalization challenges and the quest for a generic neuromotor interface.
Cloudflare says Perplexity ignored robots blocks, masked its crawler identity, and scraped content from sites that asked AI scrapers not to access them. Perplexity disputes the claim.
Explores how generative AI is transforming medical education, with insights from Morgan Cheatham and Daniel Chen on tool integration and the questions of trust, training, and responsibility.
Overview of Semantic Telemetry and the technical approach to classifying human-AI interactions at scale, covering batching, token optimization, and orchestration.
In the Microsoft Research Podcast series finale, Amanda Craig Deckard examines what Microsoft has learned about testing as a governance tool for AI, highlighting rigor, standardization, interpretability, and next steps for AI governance.
Overview of Arc Institute's Virtual Cell Challenge, its STATE baseline, and the implications for in silico cell perturbation modeling and context generalization in transcriptomics.
A deep dive into Consilium, the multi-LLM platform that enables models to discuss, debate, and reach consensus via MCP servers and a visual Gradio roundtable.
CollabLLM teaches LLMs to collaborate with users by knowing when to ask questions and adapting tone to context, highlighted in Microsoft Research's Semantic Telemetry post for ICML 2025 support.
Overview of StruQ and SecAlign defenses to mitigate prompt injection in LLM-powered apps, with Secure Front-End concepts and evaluation results.
Nvidia teams with EPRI to form the Open Power AI Consortium, building open-source, domain-specific AI models to tackle grid challenges driven by AI workloads and rising data-center demand.
As progress in ML shifts toward scale and empiricism, mathematics is evolving from strict guarantees to post-hoc explanations, higher-level structure, and symmetry-aware design, spanning topology, geometry, and algebra to interpret and guide large-scale models.
Explores how mathematics remains central to ML, but its role is evolving from theory-first guarantees to geometry, symmetries, and post-hoc explanations in scale-driven AI.
Anthology conditions language models on richly detailed backstories to simulate representative, consistent, and diverse virtual personas for surveys and social science research.
Analysis of how ChatGPT responds to different English dialects, highlighting biases against non-standard varieties and implications for global users.
Explores why purposeful, multi round dialogue matters for LLM chatbots beyond one shot prompts, and outlines training, evaluation, and implementation challenges for engineers and enterprises.
StrongREJECT advances jailbreak evaluation by pairing a high-quality forbidden-prompt dataset with automated evaluators aligned to human judgments, delivering more reliable measurements of jailbreak effectiveness against frontier LLMs.
Survey of how large language models (LLMs) could augment autonomous driving across perception, planning, and generation, with examples, challenges, and early results.
Examines whether text embeddings can be inverted to recover original text, detailing vec2text and its impact on privacy in embedding-based systems.
Explores how deep learning powers single-cell sequencing analysis—from data structure and autoencoders to multimodal integration and spatial transcriptomics.
A critical examination of how mainstream x-risk discourse, commercialization, and RLHF/RLAIF shape AI safety priorities—and why this may diverge from concrete, present-day harms.
A substantial critique of AI alignment debates, arguing that current alignment work is misaligned with real-world harms and driven by product incentives. Examines RLHF/RLAIF, intent alignment, and the HH H preference model as described in The Gradient essay.
Overview of AI consciousness debates, definitions, and moral status implications. Highlights Block's distinctions (self-, monitoring-, access-, p-consciousness) and Chalmers' easy vs hard problems; discusses need for cross-disciplinary work.