Interleaves chain-of-thought reasoning with tool-using actions in one LLM loop: the model plans, queries a source like Wikipedia, then revises from results. Cuts hallucination versus reasoning-only prompting and beats trained agents on interactive tasks.
First model to learn control policies straight from raw Atari pixels, pairing a convolutional net with Q-learning and experience replay. One unchanged architecture played seven games, beating prior methods on six and a human expert on three.
Combines a policy network (to narrow move choices) and a value network (to score board positions) with Monte Carlo tree search, cutting Go's vast search space enough to beat top programs 99.8% of the time and the European champion 5-0.
AI Infra2017
Scales any Python or ML workload across CPUs and GPUs with a few decorators, instead of rewriting code for Spark or MPI. Bundles libraries for distributed training, hyperparameter tuning, RL, batch inference, and online model serving on one cluster.
Provides 150+ executed Jupyter notebooks and code that reproduce the book 'Machine Learning for Algorithmic Trading (2nd ed.)' — covers feature engineering, alternative-data signal extraction, backtesting, NLP, deep learning and reinforcement learning for trading; best for quant researchers and practitioners.
AI Train2019
Implements deep RL algorithms (PPO, DQN, SAC, TD3, DDPG, C51, PPG) as standalone single-file scripts — the PPO Atari variant is ~340 readable lines. Built for research debugging and reproducibility, with W&B and TensorBoard tracking.
AI Train2020
thu-ml (Tsinghua University Machine Learning Group)Tsinghua University, appliedAI Institute for Europe
Deep reinforcement learning library on pure PyTorch and Gymnasium, with 30+ algorithms across on-policy, off-policy, and offline RL. Exposes both a one-call high-level interface and a procedural API, plus vectorized envs and reproducible MuJoCo benchmarks.
AI Infra2020
An AI-native, weight-centric infrastructure for quantitative trading that produces target portfolio weight vectors to unify data ingestion, strategy composition, backtesting, and live/broker execution. Modular pipeline supports ML/DRL allocators, LLM-ready preprocessing, multi-source data, and Alpaca integration for paper/live trading.
AI Infra2020
Covers the full AI quant pipeline — point-in-time data, model training, backtesting, portfolio optimization, and order execution. Supports supervised learning, market dynamics, and RL on 20+ models, plus an LLM-based RD-Agent for factor mining.
Made reinforcement learning from human feedback (RLHF) the standard alignment recipe: collect demonstrations and preference rankings, train a reward model, then optimize with PPO. A 1.3B aligned model was preferred over the 175B GPT-3 by human raters.
Embodied AI2022
GPU-accelerated robot-learning framework on NVIDIA Isaac Sim, running thousands of parallel environments on one GPU for reinforcement and imitation learning. Ships 30+ ready-to-train tasks and 16+ robot models wired to RSL RL, SKRL, and RL Games.
AI Dataset2022
Provides human preference comparison pairs and red-team conversation transcripts collected by Anthropic for training preference/reward models and studying harmful model behaviors; intended for RLHF and safety research, not for supervised fine-tuning of dialogue agents.