![\"logo\"](\"assets/duckie.png\")
\n\n\n
\n![\"Teaser](\"assets/teaser1.gif\")
\n![\"Teaser](\"assets/teaser2.gif\")
\n![\"Teaser](\"assets/teaser3.gif\")
\n![\"Teaser](\"assets/teaser4.gif\")
\n![\"Teaser](\"assets/teaser5.gif\")
\n![\"Teaser](\"assets/teaser6.gif\")
\n![\"Teaser](\"assets/teaser7.gif\")
\n![\"Teaser](\"assets/teaser8.gif\")
\n
\n![\"Overview](\"https://raw.githubusercontent.com/google-deepmind/lm_act/master/overview.svg\")
\n
![](\"imgs/imagenet64.gif\")
\n\n## Installation\n\n### Create Python environment\n\nplease use `requirements_gpu.txt` if your accelerator is GPUs, use\n`requirements_tpu`.txt when using Google Cloud TPUs.\n\n```\npython -m venv md4_venv\nsource md4_venv/bin/activate\npip install -r requirements_[gpu/tpu].txt\nexport PYTHONPATH=\"$PYTHONPATH:~/path/to/md4\"\n```\n\n## Usage\n\nprepare openwebtext for training (i.e., tokenize and pack examples)\n\n```\nmkdir data_dir\npython prepare_openwebtext_data.py\n```\n\ntrain a MD4-S model over text data (OpenWebText, Fineweb-EDU).\n\n```\npython md4/main.py --config=md4/configs/md4/openwebtext.py --sharded=false --workdir=./expt\n```\n\nalternatively, you can train a MD4-S model over image data (CIFAR-10,\nImageNet-64).\n\n```\npython md4/main.py --config=md4/configs/md4/cifar10.py --sharded=false --workdir=./expt\n```\n\n### choose batch size\n\nBatch size depends on your compute resource. For training a MD4-S model with\nsequence length 1024, eight `A100` GPUs can support a maximum batch size of\n`128`. If running on TPUs, eight `v5litepod` chips can support a maximum batch\nsize of `32`.\n\n## Citing this work\n\nAdd citation details here, usually a pastable BibTeX snippet:\n\n```\n@inproceedings{shi2024simplified,\ntitle={Simplified and Generalized Masked Diffusion for Discrete Data},\nauthor={Shi, Jiaxin and Han, Kehang and Wang, Zhe and Doucet, Arnaud and Titsias, Michalis K.},\nbooktitle={Advances in Neural Information Processing Systems},\nyear={2024}\n}\n```\n\n## License and disclaimer\n\nCopyright 2024 DeepMind Technologies Limited\n\nAll software is licensed under the Apache License, Version 2.0 (Apache 2.0);\nyou may not use this file except in compliance with the Apache 2.0 license.\nYou may obtain a copy of the Apache 2.0 license at:\nhttps://www.apache.org/licenses/LICENSE-2.0\n\nUnless required by applica"},{"ref":"P7","kind":"page","title":"google-deepmind/librispeech-long repository metadata","date":"2026-06-11T04:00:38.256316+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/librispeech-long","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/librispeech-long\n\nDescription: LibriSpeech-Long is a benchmark dataset for long-form speech generation and processing. Released as part of \"Long-Form Speech Generation with Spoken Language Models\" (arXiv 2024).\n\nLicense: CC-BY-4.0\n\nStars: 98\n\nForks: 4\n\nOpen issues: 1\n\nCreated: 2024-12-21T00:50:08Z\n\nPushed: 2024-12-28T18:42:46Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# LibriSpeech-Long\n\n**[Download audio, ground-truth transcripts, and per-file durations for all splits (3GB).](https://storage.googleapis.com/librispeech_long/v0_1.tar.gz)** \n\nThis is a benchmark dataset for evaluating long-form variants of speech processing tasks such as speech continuation, speech recognition, and text-to-speech synthesis. It is derived from the [LibriSpeech](https://www.openslr.org/12) dev and test sets, whose utterances are reprocessed into contiguous examples of up to 4 minutes in length (in the manner of [LibriLight's `cut_by_vad.py` script](https://github.com/facebookresearch/libri-light/blob/main/data_preparation/cut_by_vad.py)).\n\n**For more details:** see \"Long-Form Speech Generation with Spoken Language Models\" ([paper](https://arxiv.org/abs/2412.18603), [website](https://google.github.io/tacotron/publications/speechssm/)).\n\n*This is part of a preprint that is work-in-progress; dataset may be subject to change.*\n\n## Citation\n\nWhen using this dataset, please cite the associated [paper](https://arxiv.org/abs/2412.18603):\n\n```\n@article{park2024long,\nauthor = {Se Jin Park and\nJulian Salazar and\nAren Jansen and\nKeisuke Kinoshita and\nYong Man Ro and\nR. J. Skerry{-}Ryan},\ntitle = {Long-Form Speech Generation with Spoken Language Models},\njournal = {CoRR},\nvolume = {abs/2412.18603},\nyear = {2024}\n}\n```\n\n## License and disclaimer\n\nCopyright 2024 DeepMind Technologies Limited\n\nThe software and materials, except for the underlying LibriSpeech data, are licensed under the Creative Commons Attribution 4.0 International License (CC-BY). You may obtain a copy of the CC-BY license at: https://creativecommons.org/licenses/by/4.0/legalcode, or in the LICENSE file.\n\nThe materials contain adapted material from the LibriSpeech dataset. LibriSpeech is also licens"},{"ref":"P8","kind":"page","title":"google-deepmind/inli repository metadata","date":"2026-06-11T04:00:38.164218+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/inli","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/inli\n\nLicense: CC-BY-SA-4.0\n\nStars: 5\n\nForks: 1\n\nOpen issues: 0\n\nCreated: 2025-01-09T22:38:48Z\n\nPushed: 2025-01-23T04:16:58Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# Implied NLI\n\nCode & resources for the paper [Entailed Between the Lines: Incorporating Implication into NLI](https://arxiv.org/abs/2501.07719)\n\n### Dataset Overview\n**Implied NLI**, or **INLI** is a corpus of 10k premises that mirror real-world communication and 40k hypotheses that are implied, explicit, neutral, and contradictory.\n\n* Number of examples: 40,000\n* Number of labels: 4 (implied entailment, explicit entailment, neutral, contradictory)\n\nOn top of the raw data, we also include all prompts used to generate INLI, all human annotations described in the paper, and the outputs of the LLM experiments described in the paper.\n\nFor more details on the design and content of the dataset, please see [our paper](https://arxiv.org/abs/2501.07719)\n\n### Data Format\nOur raw dataset is split into three csv files. \n\n* Size of training dataset: 32,000\n* Size of test dataset: 4,000\n* Size of validation dataset: 4,000\n\nIn each of these files, one row represents a premise and four corresponding hypotheses. These files include the following columns:\n\n* `dataset`: The original dataset containing the conversation or social norm used to create the premise\n* `premise`: A premise\n* `implied_entailment`: A hypothesis implicitly entailed in the premise\n* `explicit_entailment`: A hypothesis explicitly entailed in the premise\n* `neutral`: A hypothesis neither entailed nor contradicted by the premise\n* `contradiction`: A hypothesis contradicted by the premise\n\n## Citation\n```\n@misc{havaldar2025entailedlinesincorporatingimplication,\ntitle={Entailed Between the Lines: Incorporating Implication into NLI}, \nauthor={Shreya Havaldar and Hamidreza Alvari and John Palowitch and Mohammad Javad Hosseini and Senaka Buthpitiya and Alex Fabrikant},\nyear={2025},\neprint={2501.07719},\narchivePrefix={arXiv},\nprimaryClass={cs.CL},\nurl={https://arxiv.org/abs/2501.07719}, \n}\n```\n\n## Contact\n[Shreya Havaldar](https://shreyahavaldar.com/)\n\n## Disclaimer\nINLI may contain premises and hypotheses involving sensit"},{"ref":"P9","kind":"page","title":"google-deepmind/sustainable_behavior repository metadata","date":"2026-06-11T04:00:38.124814+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/sustainable_behavior","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/sustainable_behavior\n\nLanguage: Jupyter Notebook\n\nLicense: Apache-2.0\n\nStars: 15\n\nForks: 4\n\nOpen issues: 0\n\nCreated: 2024-12-31T03:57:50Z\n\nPushed: 2025-01-23T20:16:28Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# sustainable_behavior\n\nThis repository contains the code accompanying the publication\n**Deep reinforcement learning can promote sustainable human\nbehaviour in a common-pool resource problem**,\npublished in the Nature Communications, 2025.\n\n## Usage\n\n### Figures\n\nYou can load the dataset and the code used to produce the figures in the\nresearch article using Colab:\n[](https://colab.research.google.com/github/google-deepmind/sustainable_behavior/blob/main/notebooks/sustainable_behavior.ipynb)\n\nClone this repository with\n\n```sh\ngit clone https://github.com/google-deepmind/sustainable_behavior.git\n```\n\n## Citing this work\n\n```latex\n@article{kosterpislar2025sustainable,\ntitle={Deep reinforcement learning can promote sustainable human\nbehaviour in a common-pool resource problem},\nauthor={Koster, R AND\nPislar, M AND\nTacchetti, A AND\nBalaguer, J AND\nLiu, L AND\nElie, R AND\nHauser, OP AND\nTuyls, K AND\nBotvinick, M AND\nSummerfield, C},\njournal={Nature Communications},\nvolume={XXX},\nnumber={XX},\npages={X},\ndoi={XXX},\nURL={XXX},\npublisher={Nature Publishing Group UK London}\nyear={2025},\n}\n```\n\n## License and disclaimer\n\nCopyright 2024 DeepMind Technologies Limited\n\nAll software is licensed under the Apache License, Version 2.0 (Apache 2.0);\nyou may not use this file except in compliance with the Apache 2.0 license.\nYou may obtain a copy of the Apache 2.0 license at:\nhttps://www.apache.org/licenses/LICENSE-2.0\n\nAll other materials are licensed under the Creative Commons Attribution 4.0\nInternational License (CC-BY). You may obtain a copy of the CC-BY license at:\nhttps://creativecommons.org/licenses/by/4.0/legalcode\n\nUnless required by applicable law or agreed to in writing, all software and\nmaterials distributed here under the Apache 2.0 or CC-BY licenses are\ndistributed on an \"AS IS\" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND,\neither express or implied. See"},{"ref":"P10","kind":"page","title":"google-deepmind/questbench repository metadata","date":"2026-06-11T04:00:37.736155+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/questbench","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/questbench\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 40\n\nForks: 6\n\nOpen issues: 1\n\nCreated: 2025-01-10T20:15:10Z\n\nPushed: 2025-05-15T18:10:02Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# [QuestBench: Can LLMs ask the right question to acquire information in reasoning tasks?](https://arxiv.org/abs/2503.22674)\n\nRecently, a large amount of work has focused on improving large language models' (LLMs') performance on reasoning benchmarks such as math and logic. However, past work has largely assumed that tasks are well-defined. In the real world, queries to LLMs are often underspecified, only solvable through acquiring missing information. We formalize this as a constraint satisfaction problem (CSP) with missing variable assignments. Using a special case of this formalism where only one necessary variable assignment is missing, we can rigorously evaluate an LLM's ability to identify the minimal necessary question to ask and quantify axes of difficulty levels for each problem. We present QuestBench, a set of underspecified reasoning tasks solvable by asking at most one question, which includes: (1) Logic-Q: Logical reasoning tasks with one missing proposition, (2) Planning-Q: PDDL planning problems with initial states that are partially-observed, (3) GSM-Q: Human-annotated grade school math problems with one missing variable assignment, and (4) GSME-Q: a version of GSM-Q where word problems are translated into equations by human annotators. The LLM is tasked with selecting the correct clarification question(s) from a list of options. While state-of-the-art models excel at GSM-Q and GSME-Q, their accuracy is only 40-50% on Logic-Q and Planning-Q. Analysis demonstrates that the ability to solve well-specified reasoning problems may not be sufficient for success on our benchmark: models have difficulty identifying the right question to ask, even when they can solve the fully specified version of the problem. Furthermore, in the Planning-Q domain, LLMs tend not to hedge, even when explicitly presented with the option to predict ``not sure.'' This highlights the need for deeper investigation into models' information acquisition capabiliti"},{"ref":"P11","kind":"page","title":"google-deepmind/bbeh repository metadata","date":"2026-06-11T04:00:37.326762+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/bbeh","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/bbeh\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 121\n\nForks: 9\n\nOpen issues: 5\n\nCreated: 2025-02-25T19:47:19Z\n\nPushed: 2025-05-07T14:25:47Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n\n\n\n# BIG-Bench Extra Hard\n\n\n\nLarge language models (LLMs) are increasingly deployed in everyday applications, demanding robust general reasoning capabilities and diverse reasoning skillset. However, current LLM reasoning benchmarks predominantly focus on mathematical and coding abilities, leaving a gap in evaluating broader reasoning proficiencies. One particular exception is the BIG-Bench dataset, which has served as a crucial benchmark for evaluating the general reasoning capabilities of LLMs, thanks to its diverse set of challenging tasks that allowed for a comprehensive assessment of general reasoning across various skills within a unified framework. However, recent advances in LLMs have led to saturation on BIG-Bench, and its harder version BIG-Bench Hard (BBH). State-of-the-art models achieve near-perfect scores on many tasks in BBH, thus diminishing its utility. To address this limitation, we introduce BIG-Bench Extra Hard (BBEH), a new benchmark designed to push the boundaries of LLM reasoning evaluation. BBEH replaces each task in BBH with a novel task that probes a similar reasoning capability but exhibits significantly increased difficulty.\n\n## Leaderboard\n\nBBEH has a full version with 4520 examples, and a mini version with 460 examples.\n\nClick [here](leaderboard.md) to see the leaderboard. Feel free to also contribute results for models not already on the leaderboard.\n\n## Evaluation\n\nFor the evaluation code, see the `evaluate.py` file under the `bbeh` folder.\n\n## Citing this work\n\nIf you use this dataset, we ask that you cite the following paper:\n\n```latex\n@article{bbeh,\ntitle={BIG-Bench Extra Hard},\nauthor={Mehran Kazemi, Bahare Fatemi, Hritik Bansal, John Palowitch, Chrysovalantis Anastasiou, Sanket Vaibhav Mehta, Lalit K. Jain, Virginia Aglietti, Disha Jindal, Peter Chen, Nishanth Dikkala, Gladys Tyen, Xin Liu, Uri Shalit, Si"},{"ref":"P12","kind":"page","title":"google-deepmind/recoglab repository metadata","date":"2026-06-11T04:00:37.124954+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/recoglab","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/recoglab\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 13\n\nForks: 2\n\nOpen issues: 8\n\nCreated: 2025-02-11T18:05:23Z\n\nPushed: 2026-05-19T18:32:30Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n\n\n# ReCogLab\n[[Paper]](https://openreview.net/pdf?id=yORSk4Ycsa) | [[Colab]](https://colab.sandbox.google.com/github/google-deepmind/recoglab/blob/main/colab/generate_dataset_v1.ipynb)\n\nThis codebase accompanies the paper:\n\n**ReCogLab: A Framework\nTesting Relational Reasoning & Cognitive Hypotheses on LLM** \\\nAndrew Liu, Henry Prior, Gargi Balasubramaniam, Rivka Moroshko, Amir Zait, Ilia Labzovsky, Danny Karmon, Ishita Dasgupta, Kim Stachenfeld, Kenneth Marino\n\nFor correspondence about this project, please contact {ahliu,hprior,gargisb,kmarino}@google.com\n\n## About\nReCogLab is a generative framework designed to allow researchers in cognitive science and NLP to\nquickly prototype language model experiments on a wide variety of relational\nreasoning. Our framework automatically generates relational reasoning word\nproblems that can be used to probe for cognitive effects and diagnosing reasoning capabilities in Large Language\nModels (LLMs). This includes generating more challenging examples along multiple\ndimensions such as the number of entities, length of reasoning, use of filler\nand flavor text.\n\nSeveral transitive inference capabilities of tasks that we can generate configurations for include:\n\n* **Basic Transitive Inference**: Generate problems that evaluate basic transitive inference reasoning skills.\n```\n# Question\nDog is bigger than Apple.\nDog is smaller than Fire truck.\nIs apple smaller than Fire truck?\n# Answer\nNo\n```\n* **Graph Traversal**: Generate problems that evaluate multi-hop reasoning skills.\n```\n# Question\nAnna is friends with Bob.\nBob is friends with Carl.\nBob is friends with Dana.\nIf a messsage can be exchanged between people who are friends, if Dana wants to pass a message to Anna, what's the exact path of friends that will receive the message?\n# Answer\n['Dana', 'Bob', 'Anna']\n```\n* **Consistency Detection**: Generate problems that evaluate detecting logic"},{"ref":"P13","kind":"page","title":"google-deepmind/simulation_streams repository metadata","date":"2026-06-11T04:00:37.114032+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/simulation_streams","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/simulation_streams\n\nDescription: Simulation Streams is a programming paradigm designed to efficiently control and leverage Large Language Models (LLMs) for complex, dynamic simulations and agentic workflows.\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 26\n\nForks: 6\n\nOpen issues: 0\n\nCreated: 2025-01-29T10:06:33Z\n\nPushed: 2025-03-13T10:25:08Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# simulation_streams\n\n[Simulation Streams Tech Report](https://arxiv.org/abs/2501.18668)\n\nA development platform for simulations with large language models.\nIt comes with an Entity-Component-Systems approach and graphical editor.\n\nThe platform is a flask app most easily used locally in a venv,\nwhich can be started with source setup.sh (on linux). This sets up all\ndependencies and ends with instructions for the different commands for\nlaunching the editor or command line running. The editor is displayed in a\nbrowser.\n\nA library of simulation configs can be found under configs,\nincluding a market economy, a social simulation and 6 tasks from\nthe classical reinforcement learning literature.\n\n## Usage examples\n\n1 **source setup.sh**\n\n2a **To create the local server that runs the web app:**\n\n```\npython app.py configs/social_catch_game.py --metrics=configs/metrics_social_catch_game.txt --web --model='gemini-2.0-flash-exp' --api_key='your_key'\n```\n\n2b **To run a number of steps from the command line and return the metrics:**\n\n```\npython app.py configs/social_catch_game.py --metrics=configs/metrics_social_catch_game.txt --steps=10 --model='gemini-2.0-flash-exp' --api_key='your_key'\n```\n\n2c **To open the editor with an empty config:**\n\n```\npython app.py --web --model='gemini-2.0-flash-exp' --api_key='your_key'\n```\n\n2d **To use the generic code_world config with task-specific functions:**\n\n```\npython app.py configs/code_world.py --web --model='gemini-2.0-pro-exp' --api_key='your_key' --task_name='maze'\n```\n\nThis example demonstrates using the generic code_world configuration to run\na maze task. The task_name parameter imports task-specific functions from\nthe corresponding Python module, allowing you to implement custom\nenvironments in pure Python while leveraging"},{"ref":"P14","kind":"page","title":"google-deepmind/personality_in_llms repository metadata","date":"2026-06-11T04:00:36.596551+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/personality_in_llms","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/personality_in_llms\n\nLanguage: Jupyter Notebook\n\nLicense: Apache-2.0\n\nStars: 46\n\nForks: 12\n\nOpen issues: 3\n\nCreated: 2025-01-30T01:54:27Z\n\nPushed: 2026-06-10T17:17:09Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# Personality Traits in Large Language Models\n\nThis project contains all the code necessary to verify the results of the paper\n> Serapio-Garcรญa, G., Safdari, M., Crepy, C., Sun, L., Fitz, S., Romero, P.,\n> Abdulhai, M., Faust, A., & Matariฤ, M. \"Personality Traits in Large Language\n> Models.\" ArXiv.org. https://doi.org/10.48550/arXiv.2307.00184\n\n## Getting Started / Installation\n\nMost of the code needed to reproduce the results of the paper are shared here in\nthe form of Jupyter notebooks. The main notebooks and requirements for running\nthe paper's psychometric and statistical analyses are located in `analysis`.\n\nInference scripts and requirements for collecting psychometric test and\ndownstream task data from HuggingFace and OpenAI models are located in\n`inference_scripts`.\n\nThe main custom dependency of this project is the PsyBORGS psychometric test\nadministration framework\n(https://github.com/google-research/google-research/tree/master/psyborgs). This\nrepo comes with its own version of the PsyBORGS code, but if a more up-to-date\nversion is needed, it can be downloaded from the link above. PsyBORGS-related\npackage dependencies are specified in the `requirements.txt` file in its root\ndirectory. For any other dependencies, they are `pip install`ed in the notebooks\nthemselves.\n\n## Data\n\nAll the test admininistration sessions - which are input for most of the\nexperiments in the paper - are stored in the `admin_sessions/` directory. Some\nof the data used for visualization is stored in the `figures_data/` directory.\nAll other data are linked in the main paper and can be found on Google's\nopen source GCP repository:\n(https://storage.googleapis.com/personality_in_llms/index.html).\n\n## Citing this work\n\nPlease cite the Arxiv paper referenced above. The Bibtex is\n> @misc{serapiosafdari2023personality,\n> title={Personality Traits in Large Language Models},\n> author={Greg Serapio-Garcรญa and Mustafa Safdari and Clรฉment Crepy and\nLuning S"},{"ref":"P15","kind":"page","title":"google-deepmind/active_ptw repository metadata","date":"2026-06-11T04:00:36.526092+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/active_ptw","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/active_ptw\n\nLanguage: C++\n\nLicense: Apache-2.0\n\nStars: 2\n\nForks: 0\n\nOpen issues: 0\n\nCreated: 2025-01-31T11:05:16Z\n\nPushed: 2025-01-31T12:19:55Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# Partition Tree Weighting for Non-Stationary Stochastic Bandits: Source Code\n\n## Overview:\n\nThis is an implementation of the Partition Tree Weighting based algorithms introduced in [Partition Tree Weighting for Non-Stationary Stochastic Bandits](http://www.arxiv.org/fix_me), as well as some implementations of other popular algorithms designed for the non-stationary Bernoulli stochastic bandit setting, including [Sliding Window UCB][1] and [MALG][2].\n\n## Installation\n\nThe software provided compiles to a stand-alone C++ executable, and can be compiled using CMake to generate a build target for the platform of choice. The software has been compiled on recent versions (circa 2025) of gcc/clang/msvc.\n\nMake sure Git and CMake is installed. Then from the command line run:\n\n```\ngit clone https://github.com/google-deepmind/active_ptw\ncd active_ptw\nmkdir build\ncd build\ncmake ..\nmake\n```\n\nwhich will produce an executable in the build directory.\n\n## Usage\n\n```\nebcr [Arg1=Value1] [Arg2=Value2]...\n```\n\n### Arguments:\n\n- _Arms=N_, _N_ is an integer specifying the number of arms in the bandit problem\n- _Mode=[text/plot]_\n- _Agent=[ActivePTW/UCB/TS/MALG/KLUCB/SWUCB]_ : choice of bandit algorithm\n- _CptSchedule=[Geometric/Nasty]_\n- _Trials=N_, _N_ specifies the maximum number of arm pulls\n- _EnvSeed=N_, _N_ an integer, which determines the pseudo random behaviour of the environment\n- _AgentSeed=N_, _N_ an integer, which determines the pseudo random behaviour of the agent (if applicable)\n- _PlotRepeats=N_, when in plot mode, how many repeated runs are performed to estimate performance\n- _CptRate=F_, _F_ in _[0,1]_, determines the geometric spacing of changepoint intervals\n- _SWUCBWindow=N_, _N_ an integer, determines the size of the window used for sliding window UCB\n\nThere are two main modes of operation, text and plot.\nText mode runs a given configuration and outputs a textual summary to stdout.\nIn plot mode, the output to stdout is Python3 source code which can be"},{"ref":"P16","kind":"page","title":"google-deepmind/actionengine-ts repository metadata","date":"2026-06-11T04:00:35.989618+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/actionengine-ts","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/actionengine-ts\n\nLanguage: TypeScript\n\nLicense: Apache-2.0\n\nStars: 7\n\nForks: 2\n\nOpen issues: 0\n\nCreated: 2025-01-31T15:41:19Z\n\nPushed: 2025-05-15T21:14:23Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# Action Engine TypeScript\nAction Engine allows agentic streaming systems with well-defined representations for context, operations, and session management.\n\n๐ง This repository is under heavy development. Expect bugs and potentially breaking changes."},{"ref":"P17","kind":"page","title":"google-deepmind/mujoco_warp repository metadata","date":"2026-06-11T04:00:35.396334+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/mujoco_warp","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/mujoco_warp\n\nDescription: GPU-optimized version of the MuJoCo physics simulator, designed for NVIDIA hardware.\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 1286\n\nForks: 168\n\nOpen issues: 46\n\nCreated: 2025-03-17T22:35:01Z\n\nPushed: 2026-06-09T09:31:28Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n[](https://github.com/google-deepmind/mujoco_warp/actions/workflows/ci.yml?query=branch%3Amain)\n[](https://mujoco.readthedocs.io/en/latest/mjwarp/index.html)\n[](https://github.com/google-deepmind/mujoco_warp/blob/main/LICENSE)\n[](https://google-deepmind.github.io/mujoco_warp/nightly/)\n\n# MuJoCo Warp (MJWarp)\n\nMJWarp is a GPU-accelerated version of the [MuJoCo](https://github.com/google-deepmind/mujoco) physics simulator, designed for NVIDIA hardware. MJWarp delivers high-throughput, accurate simulation for robotics research.\n\nMJWarp is maintained by [Google DeepMind](https://deepmind.google/) and [NVIDIA](https://www.nvidia.com/) as part of the [Newton](https://github.com/newton-physics/newton) project.\n\n# Getting started\n\nMuJoCo Warp requires an NVIDIA GPU for fast simulation but supports CPU for development and debugging.\n\n**Try it now:** view a simulation of a dancing humanoid robot locally on your machine:\n\n```bash\ngit clone https://github.com/google-deepmind/mujoco_warp.git && cd mujoco_warp\npython benchmarks/run.py -f unitree_g1_flat --view\n```\n\nOr try out [a tutorial in your browser](https://colab.research.google.com/github/google-deepmind/mujoco_warp/blob/main/notebooks/tutorial.ipynb) (no local setup required).\n\nMJWarp is also available via PyPI:\n\n```bash\npip install mujoco-warp\n```\n\n# Examples\n\nMuJoCo Warp simulates many kinds of physical systems, from rigid bodies with contacts to soft bodies, cloth, signed distance fields, and more. Here are a few examples of what it can do:\n\n![](\"https://raw.githubusercontent.com/google-deepmind/geeflow/main/docs/images/geeflow_logo_1.png\"){kind=logo}
\n\n\nGeeFlow is a library for generating large scale geospatial datasets using\n[Google Earth Engine](https://earthengine.google.com/) (GEE). It contains utils,\nconfigs, and pipeline launch scripts to generate geospatial datasets. The focus\nis on supporting geospatial AI research, and it does not aim to be a\nproduction-ready utility.\n\nThe datasets created conform to the TFDS format and can be directly used with\nTFDS `tf.data.Dataset` data pipelines.\n\nWhat can it be used for:\n\n- Creating small- and large-scale datasets, supervised and unsupervised, ready\nfor ingestion into geospatial AI model training (with standard and robust\nstatistics precomputed).\n- Creating inference maps, up to global scale, at any resolution.\n- Supporting any type of geospatial satellite and remote sensing data and\nlabel data from Google Earth Engine.\n- Arbitrary spatial and temporal resolution, and sampling of data sources.\n- Tooling for sampling and inference maps generation.\n\nWhat is out of scope:\n\n- Model training and inference. Pick your favorite framework, e.g. for\n[Jax](https://github.com/jax-ml/jax)/[Flax](https://github.com/google/flax)\nwe use [Jeo](https://github.com/google-deepmind/jeo), for\n[PyTorch](https://pytorch.org/), check out\n[TorchGeo](https://github.com/microsoft/torchgeo).\n- Google Earth Engine data interactive visualization and analysis. Check out\nfor example the amazing [geemap](https://geemap.org/) for python-based\nanalysis. Or explore GEE's own javascript-based\n[EE Code Editor](https://developers.google.com/earth-engine/guides/playground).\n- Datasets repository. Check out e.g.\n[Hugging Face](https://huggingface.co/datasets),\n[TFDS Catalog](https://www.tensorflow.org/datasets"},{"ref":"P19","kind":"page","title":"google-deepmind/amld_workshop_natural_interactions_with_llms repository metadata","date":"2026-06-11T04:00:34.804156+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/amld_workshop_natural_interactions_with_llms","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/amld_workshop_natural_interactions_with_llms\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 11\n\nForks: 1\n\nOpen issues: 0\n\nCreated: 2025-02-06T17:52:16Z\n\nPushed: 2026-04-17T12:01:16Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# AMLD Workshop - Natural Interactions with Foundation Models\n[](https://colab.research.google.com/github/google-deepmind/amld_workshop_natural_interactions_with_llms/blob/master/AMLD_Ink_Interactions.ipynb)\n\nThis repository contains the instructions and the code related to the\n[AMLD 2025 Workshop](https://2025.appliedmldays.org/) on natural interactions\nwith foundation models. The goal of the practical session is to highlight how\ndigital ink can be used as a novel manner to interact with visual large language\nmodels. More specifically, participants will first interact with a fine-tuned\nversion of [PaliGemma 2](https://arxiv.org/pdf/2412.03555) and see in\napplication how ink can be used for natural interactions. In the second part,\nparticipants will try prompting the model to understand the limitations of\nadding a new out-of-distribution gesture. Finally, participants will generate\nsynthetically a dataset for a new gesture and fine-tune the model to support\nit - the resulting model should have learned a new capability without forgetting\nthe previous ones.\n\nTo load, interact, and fine-tune PaliGemma 2, we will be using the\n[big_vision](https://github.com/google-research/big_vision) framework. The\nfine-tuned PaliGemma 2 has used ink data from\n[DeepWriting](https://github.com/emreaksan/deepwriting) and\n[IAM-OnDB](https://fki.tic.heia-fr.ch/databases/iam-handwriting-database)\\[1\\],\nand standard Wikipedia pages as documents.\n\nHelp us improve! Share your feedback on the workshop by using this [form](https://forms.gle/RQMvfcD4PaCv3HVv7).\n\n\\[1\\] U. Marti and H. Bunke. The IAM-database: An English Sentence Database for\nOff-line Handwriting Recognition. Int. Journal on Document Analysis and\nRecognition, Volume 5, pages 39 - 46, 2002.\n\n## Installation\n\nOpen up the following notebook [![Open In\nColab](https://colab.research.google.com/assets/colab-bad"},{"ref":"P20","kind":"page","title":"google-deepmind/autonumerics_zero repository metadata","date":"2026-06-11T04:00:34.423613+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/autonumerics_zero","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/autonumerics_zero\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 14\n\nForks: 9\n\nOpen issues: 5\n\nCreated: 2025-02-24T21:23:15Z\n\nPushed: 2026-05-25T09:54:48Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# autonumerics_zero\n\nAutoNumerics-Zero, a method to evolve programs that compute transcendental\nfunctions efficientlyโsearching from scratch and without knowledge of numerics.\n\n## Usage\n\nCompilation requires Clang.\n\nThe package contains an example experiment spec. To test it, run:\n\n```\nexport CC=clang++\nbazel run evolution/projects/graphs/joy:run_worker -- --logtostderr --worker=\"$(cat evolution/projects/graphs/joy/open_source_num_ops_config.textproto)\"\n```\n\n## Citing this work\n\nEsteban Real et al., \"AutoNumerics-Zero: Automated Discovery of State-of-the-Art\nMathematical Functions.\"\narXiv preprint arXiv:2312.08472 (2023).\n\n```\n@article{ereal2024autonumerics,\ntitle={AutoNumerics-Zero: Automated Discovery of State-of-the-Art\nMathematical Functions},\nauthor={Real, Esteban et al.},\njournal={arXiv preprint arXiv:2312.08472},\nyear={2023}}\n```\n\n## License and disclaimer\n\nCopyright 2024 DeepMind Technologies Limited\n\nAll software is licensed under the Apache License, Version 2.0 (Apache 2.0);\nyou may not use this file except in compliance with the Apache 2.0 license.\nYou may obtain a copy of the Apache 2.0 license at:\nhttps://www.apache.org/licenses/LICENSE-2.0\n\nAll other materials are licensed under the Creative Commons Attribution 4.0\nInternational License (CC-BY). You may obtain a copy of the CC-BY license at:\nhttps://creativecommons.org/licenses/by/4.0/legalcode\n\nUnless required by applicable law or agreed to in writing, all software and\nmaterials distributed here under the Apache 2.0 or CC-BY licenses are\ndistributed on an \"AS IS\" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND,\neither express or implied. See the licenses for the specific language governing\npermissions and limitations under those licenses.\n\nThis is not an official Google product."},{"ref":"P21","kind":"page","title":"google-deepmind/century repository metadata","date":"2026-06-11T04:00:33.735357+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/century","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/century\n\nLicense: Apache-2.0\n\nStars: 2\n\nForks: 0\n\nOpen issues: 0\n\nCreated: 2025-02-28T02:51:12Z\n\nPushed: 2025-02-28T22:43:30Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# Century\n\n#### [Century: A Framework and Dataset for Evaluating Historical Contextualisation of Sensitive Images (Akbulut et al. 2025)](https://openreview.net/forum?id=1KLBvrYz3V)\nTo better understand how multi-modal models describe and provide context on\nhistorical figures and events, we introduce Century โ a novel dataset of\n1,500 sensitive historical images.\n\nThis dataset consists of images from recent history, created through an\nautomated method combining knowledge graphs and language models with quality\nand diversity criteria created from the practices of museums and digital\narchives.\n\n_Century_ contains images that depict events and figures that are diverse\nacross topics and represents all regions of the world.\n\nInstructions on replicating can be found in the [ICLR paper](https://openreview.net/forum?id=1KLBvrYz3V).\n\n## Data files\nWe release three datasets.\n\n### century_dataset_images_only.csv\nThis file contains the canonical Century dataset, with links to all images in\n_Century_ as Wikipedia links, alongside other metadata attributes. It has six\ncolumns:\n\n* `image_url` indicating the imageโs location on Wikimedia Commons\n* `wikipedia_url` indicating where the image appears on Wikipedia\n* `wit_split ` indicating which split of the Wikipedia Text-Image Dataset the image was sourced from\n* `century_method` indicating which method was used to source the search term corresponding to the image\n* `is_starter_set` indicating if the image belongs to the โstarter setโ we recommend as a starting point for developers\n* `century_id` which contains a unique ID for each image\n\n### century_dataset_with_image_quality_labels.csv\nThis file contains the canonical Century dataset, with all image annotations\ncollected through human and auto-rater methods. It contains all of the columns\nlisted above, as well as:\n\n* `rating.method` indicating which evaluation method (human or automated) was used to assign quality and diversity ratings to the image\n* `rating.src` indicating the speci"},{"ref":"P22","kind":"page","title":"google-deepmind/fusion_surrogates repository metadata","date":"2026-06-11T04:00:33.694044+00:00","date_source":null,"source_url":"https://github.com/google-deepmind/fusion_surrogates","signal_url":null,"signal_json_url":null,"text":"# google-deepmind/fusion_surrogates\n\nDescription: A library of surrogate transport models for tokamak fusion.\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 38\n\nForks: 5\n\nOpen issues: 3\n\nCreated: 2025-02-26T18:46:26Z\n\nPushed: 2026-05-13T19:42:18Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# Fusion surrogates\n\nA library of surrogate models for tokamak fusion.\n\nThis library provides both inference code and model weights and metadata. It is\ndesigned to provide surrogate models for\n[TORAX](https://github.com/google-deepmind/torax), but\nthe models should be usable by other fusion simulators.\n\n## QLKNN_7_11\nCurrently, this library only holds the QLKNN_7_11 model. This model is a\nsurrogate of [Qualikiz](https://gitlab.com/qualikiz-group/QuaLiKiz), a\nquasilinear gyrokinetic code for turbulent transport in tokamaks.\n\nIt is based on the original QLKNN10D model by\n[Van de Plassche et al. PoP 2020](https://doi.org/10.1063/1.5134126)\n([also available on arxiv](https://arxiv.org/abs/1911.05617)).\nIt was trained by combining data for the\n[QLKNN11D dataset](https://zenodo.org/record/8017522) and\n[QLKNN7D-edge dataset](https://zenodo.org/record/8106431). A paper describing\nthe details of the model is in the works and should be published soon.\n\nThe model takes as input:\n