Today 1

• Microsoft Research·4hResearch

  Simulating student learning for better AI tutoring↗

  Microsoft Research introduced StudentSim, a system that simulates student behavior to evaluate and improve AI tutors before real-learner testing.

  #ai-tutoring#student-simulation#education

This week 11

• The Verge AI·3dResearch

  Pure insanity’: Mathematicians will need years to make sense of OpenAI’s latest drop↗

  Mathematicians describe OpenAI's sudden release of a large batch of mathematical results as overwhelming, and say understanding it could take years.

  #openai#mathematics#ai-for-science
• Hugging Face·3dResearch

  Impactful scheduling for GPU clusters↗

  A Hugging Face post discusses approaches to scheduling workloads on GPU clusters for greater impact, though the body text is missing.

  #gpu#scheduling#clusters
• MIT Technology Review AI·4dResearch

  AI breakthroughs in robotics won’t change your life any time soon↗

  An MIT Technology Review feature argues that recent robotics advances, including viral humanoid robot demos, will not quickly change everyday life.

  #robotics#humanoid-robots#ai-hype
• Google Research·5dResearch

  Does better work always mean better workers?↗

  Google Research asks whether improved work outcomes necessarily translate into better outcomes for workers, though the item provides no body text.

  #google-research#workforce#productivity
• Hugging Face·5dResearch

  Multimodal open d1 decision models for the edge↗

  Hugging Face describes a set of open multimodal decision models designed to run on edge devices.

  #multimodal#edge-ai#decision-models
• Hugging Face·5dResearch

  One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO↗

  A Hugging Face post describes fine-tuning one Nemotron model family to reach gold-level results on both the IOI and IMO competitions.

  #nemotron#fine-tuning#ioi
• Microsoft Research·6dResearch

  What AI gets wrong and what failure teaches us↗

  Microsoft Research interview with Jennifer Neville on her career path and her work identifying surprising ways AI systems fail when handling complexity.

  #ai-failures#microsoft-research#complexity
• Google Research·6dResearch

  Unlocking Earth AI’s planetary geospatial foundation models for global public health↗

  Google Research describes how its Earth AI planetary geospatial foundation models could be applied to global public health challenges.

  #geospatial#foundation-models#public-health
• OpenAI·6dResearch●

  Sharing AI progress in mathematics↗

  OpenAI reports new results on open mathematical problems produced by an internal frontier model, with Lean formalizations and research details released on GitHub.

  #openai#mathematics#frontier-model
• OpenAI·6dResearch

  Advancing computer use with Ironclad↗

  OpenAI and Ironclad are training and evaluating AI agents on complex contracting workflows to improve computer use for professional tasks.

  #computer-use#ai-agents#openai
• Google Research·7dResearch

  Open and Emergent Problems in Agentic Privacy and Security: A Contextual Angle↗

  Google Research discusses open and unresolved problems in privacy and security for agentic AI systems, framed through a contextual perspective.

  #agentic-ai#privacy#security

Earlier 45

• MIT Technology Review AI·Oct 5Research

  Connecting AI agents to enterprise knowledge↗

  Argues that enterprise AI agents underperform because they lack organization-specific knowledge, not just data, and need context to reason and decide.

  #enterprise-ai#ai-agents#knowledge-management
• Hugging Face·Oct 4Research

  The Agent Said It Was Done. The Database Disagreed.↗

  A Hugging Face post reportedly examines cases where an AI agent claimed a task was complete while the underlying database showed otherwise.

  #ai-agents#verification#evaluation
• Google Research·Oct 2Research

  Toward provably private learning from federated data↗

  Google Research describes work toward provably private machine learning on federated data, which is training models across decentralized devices without centralizing raw user data.

  #federated-learning#differential-privacy#privacy
• MIT Technology Review AI·Oct 2Research

  Don’t be fooled—LLMs don’t reason↗

  An essay recalling AlphaGo's Move 37 argues that large language models do not truly reason, questioning how their abilities are characterized.

  #llm-reasoning#alphago#ai-capabilities
• OpenAI·Oct 1Research

  The eternal complement↗

  An OpenAI essay argues advanced AI may mainly accelerate routine execution work that supports breakthrough ideas, shaping the economy and pace of progress.

  #openai#economics#future-of-work
• Microsoft Research·Sep 30Research

  Forecasting space weather risks on power grids↗

  Microsoft Research describes a machine learning system that predicts where power grid damage from extreme space weather is likely, 30 to 60 minutes before a storm arrives.

  #space-weather#power-grids#forecasting
• Google DeepMind·Sep 30Research

  Introducing SynthID Bio↗

  Google DeepMind unveils SynthID Bio, a proof-of-concept method for watermarking AI-generated proteins without disrupting their biological function.

  #watermarking#synthid#protein-design
• Hugging Face·Sep 30Research

  Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning↗

  Hugging Face introduces an open leaderboard for evaluating multilingual text-to-speech and voice cloning systems at scale.

  #text-to-speech#voice-cloning#benchmark
• Google Research·Sep 29Research

  How Diffusion Controller unifies and simplifies AI image generation↗

  Google Research describes Diffusion Controller, a method intended to unify and simplify how diffusion-based AI image generation is controlled.

  #diffusion-models#image-generation#google-research
• Microsoft Research·Sep 29Research

  Introducing Quine: An AI research system designed for the complexity of biology↗

  Microsoft Research introduces Quine, an early-stage multimodal world model of biology designed to help scientists prioritize hypotheses before lab experiments.

  #biology#multimodal#world-model
• Hugging Face·Sep 29Research

  Getting the Source Right, Not Just the Fact: Source-Aware Verification for MCP Agents↗

  A paper proposing source-aware verification for agents built on the Model Context Protocol, checking whether claims are grounded in the correct sources.

  #mcp#agents#verification
• Google Research·Sep 24Research

  Automating coherent long-form video generation↗

  Google Research describes a technique for automatically generating coherent long-form videos, addressing consistency across extended sequences.

  #video-generation#long-form-video#google-research
• Microsoft Research·Sep 23Research

  Offloaded inference for real-world physical AI robotics↗

  Microsoft Research reports findings that running AI inference off the robot, rather than onboard, can improve task success and efficiency for physical AI systems.

  #robotics#inference-offloading#physical-ai
• Anthropic·Sep 23Research

  Claude discovers a novel enzyme system with CRISPR-like repeats↗

  Anthropic reports that its Claude model helped identify a previously unknown enzyme system featuring CRISPR-like repeat sequences.

  #anthropic#claude#biology
• Hugging Face·Sep 22Research

  How UK AISI and EvalEval Are Making Benchmark Results Reproducible↗

  UK AISI and EvalEval describe efforts to make AI benchmark results more reproducible, though the accompanying text was not provided.

  #benchmarks#reproducibility#evaluation
• Microsoft Research·Sep 21Research

  Improving synthesis prediction of small molecules at scale with RetroChimera↗

  Microsoft Research describes RetroChimera, a predictive model published in Nature that aims to speed up small-molecule chemical synthesis planning.

  #chemistry#synthesis-prediction#drug-discovery
• Hugging Face·Sep 21Research

  tokenizers v1: encode, decode and scaling, measured↗

  Hugging Face publishes measurements of encoding and decoding speed and scaling for its tokenizers library v1.

  #tokenizers#hugging-face#benchmarks
• Google Research·Sep 18Research

  MilleMiglia: A realistic instance generator for middle-mile logistics↗

  Google Research introduces MilleMiglia, a generator of realistic problem instances for middle-mile logistics, aimed at algorithm development and evaluation.

  #logistics#optimization#benchmarks
• Google Research·Sep 17Research

  The future of practice: Enabling teachers to create learning interactives with generative UI↗

  Google Research describes a approach that lets teachers generate interactive learning activities using generative UI, aimed at education practice.

  #generative-ui#education#teachers
• Google Research·Sep 15Research

  Bypassing inference bottlenecks: Accelerating complex AI search with Retrieve-for-Train↗

  Google Research describes Retrieve-for-Train, a technique aimed at reducing inference bottlenecks in complex AI search systems.

  #inference#retrieval#search
• Hugging Face·Sep 15Research

  Your Agent Aced the Task. Will It Do It Again?↗

  A Hugging Face post questions whether AI agents that succeed at a task can reliably repeat that performance on later attempts.

  #ai-agents#evaluation#reliability
• Google AI (Keyword)·Sep 15Research

  AI for everyone in every language↗

  Google AI outlines a shift from text-only translation toward models that interpret languages as people actually speak and write them, across many languages.

  #multilingual#translation#google
• Google AI (Keyword)·Sep 15Research

  New insights from Google’s AI & Economy ATLAS↗

  Google released an interactive, open-access tool presenting data on AI's economic impact and adoption worldwide.

  #google#economics#ai-adoption
• Google Research·Sep 11Research

  ToolGrad: Efficient tool-use dataset generation with textual "gradients"↗

  Google Research introduces ToolGrad, a method that uses textual gradients to efficiently generate datasets for training models on tool use.

  #tool-use#dataset-generation#llm-agents
• Hugging Face·Sep 10Research

  Async GRPO with LoRA across HF Jobs: a bucket, a proxy, and no NCCL↗

  A technical write-up describes running asynchronous GRPO reinforcement learning with LoRA on Hugging Face Jobs, using storage buckets and a proxy instead of NCCL.

  #grpo#lora#reinforcement-learning
• Google DeepMind·Sep 8Research

  AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome↗

  DeepMind released a predictive atlas estimating molecular effects of about 9 billion possible single-letter DNA variants across the human genome.

  #alphagenome#genomics#dna-variants
• Google Research·Sep 3Research

  Transfer learning for genomic prediction in underrepresented populations↗

  Google Research describes applying transfer learning to improve genomic prediction accuracy for populations underrepresented in existing genetic datasets.

  #genomics#transfer-learning#health-ai
• Google Research·Sep 3Research

  A connectomics milestone: Mapping the complete male fruit fly brain↗

  Google Research reports mapping the complete connectome of a male fruit fly brain, a milestone in neuroscience and connectomics.

  #connectomics#neuroscience#google-research
• Hugging Face·Sep 3Research

  NeoMME: an efficient Multimodal-native and Multilingual Encoder↗

  NeoMME is introduced as an efficient encoder designed natively for multimodal and multilingual inputs, announced via a Hugging Face post.

  #multimodal#multilingual#encoder
• Hugging Face·Sep 3Research

  Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps↗

  A tutorial shows how to fine-tune a 350M-parameter model with GRPO reinforcement learning in 100 steps to produce more reliable structured outputs.

  #fine-tuning#grpo#structured-outputs
• Hugging Face·Sep 3Research

  Training a coding model to paint watercolours with TRL and OpenEnv↗

  A Hugging Face post describes training a coding-focused model to generate watercolour paintings using the TRL and OpenEnv frameworks.

  #reinforcement-learning#trl#openenv
• Hugging Face·Sep 2Research

  BenchMIRT: What are LLM benchmarks actually measuring?↗

  BenchMIRT examines what large language model benchmarks actually measure, questioning whether scores reflect the capabilities they are intended to assess.

  #benchmarks#evaluation#llm
• Google Research·Sep 1Research

  Mapping global methane emissions from space with deep learning↗

  Google Research describes a deep learning approach that uses satellite data to map methane emissions globally.

  #methane#climate#deep-learning
• Microsoft Research·Aug 31Research

  GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models↗

  Microsoft Research introduces GigaPath-Flash and GigaTIME-Flash, pathology foundation models designed to reduce computational cost while keeping strong performance for large-scale studies.

  #pathology#foundation-models#efficiency
• Hugging Face·Aug 28Research

  The Open ASR Leaderboard Adds Its First Global South Language↗

  Hugging Face's Open ASR Leaderboard has added its first language from the Global South, expanding speech recognition benchmarking beyond predominantly high-resource languages.

  #speech-recognition#asr#benchmarks
• Google Research·Aug 27Research

  Planetary prediction engine: Automating global models via Earth AI↗

  Google Research describes a planetary prediction engine that automates building global Earth models using its Earth AI approach.

  #earth-ai#climate-modeling#google-research
• Google DeepMind·Aug 27Research

  Piloting the world's first double-blind AI evaluations↗

  Google DeepMind is running a pilot of double-blind evaluations, where neither the model developers nor the evaluators know the other's identity, to test AI systems.

  #ai-evaluation#double-blind-testing#benchmarking
• Google Research·Aug 26Research

  GlucoFM: Foundation model for continuous glucose monitoring↗

  Google Research introduced GlucoFM, a foundation model designed for analyzing continuous glucose monitoring data in health and bioscience applications.

  #foundation-model#health#glucose-monitoring
• Google Research·Aug 25Research

  AgentHands: Generating interactive hand gestures for spatially grounded agent conversations in XR↗

  Google Research presents AgentHands, a system that generates interactive hand gestures to ground conversations with AI agents in extended reality spaces.

  #xr#hand-gestures#embodied-agents
• Anthropic·Aug 25Research

  Funding better evaluations of AI’s impact on wellbeing↗

  Anthropic announces funding for research aimed at developing better evaluations of how AI systems affect human wellbeing.

  #ai-wellbeing#evaluations#funding
• Google Research·Aug 21Research

  An AI tool for prioritizing candidate biomarkers from wearable sensor data↗

  Google Research describes an AI tool that helps prioritize candidate biomarkers derived from wearable sensor data.

  #wearables#biomarkers#sensor-data
• Google DeepMind·Aug 21Research

  From Atari to EVE Online: Building on 15 Years of AI Research in Games↗

  Google DeepMind reflects on 15 years of game AI research, from Atari to EVE Online, and is partnering with game studios to prototype new AI gameplay.

  #google-deepmind#game-ai#gaming
• Google Research·Aug 21Research

  How mobility gives language models a deeper understanding of place↗

  Google Research describes using human mobility data to improve how language models interpret and reason about geographic places. The item provides little detail beyond its title.

  #google-research#mobility-data#geospatial
• Google Research·Aug 17Research

  Seeing beyond BMI: Estimating cardiometabolic risk with smartphone imagery↗

  Google Research describes a method that estimates cardiometabolic risk from smartphone images, aiming to go beyond body mass index as a risk measure.

  #health#computer-vision#cardiometabolic-risk
• Anthropic·Aug 14Research

  How Claude’s text watermark works↗

  Anthropic explains the technical method it uses to embed watermarks in text generated by Claude, so the output can be identified later.

  #watermarking#claude#anthropic