Tagged: reinforcement-learning

125 articles

Latest

Async GRPO cuts training time by 73% using LoRA and HF Jobs

Ex-DeepMind researcher builds robots that plan ahead without trial-and-error

Berkeley's CUA-Lite kills the VM tax: OSWorld runs in Docker, 4.6x more desktops

Stable-Baselines3 hits Hugging Face Hub: load and share RL agents in two function calls

The RL recipe behind 1.5M-view AI watercolours, now fully open

Deep RL in 2024: Why trial-and-error agents are finally worth learning

Q-Learning Explained: The Algorithm That Beat Atari Games

Q-Learning from scratch: Train agents to walk on ice and drive taxis

Deep Q-Learning ditches the Q-table for neural networks to conquer Space Invaders

Policy Gradient methods ditch value functions to master CartPole and Pong

Actor-Critic methods cut RL training variance without needing more samples

PPO's clipped objective stops AI agents from 'falling off the cliff'

Google's EnvHarness bends frozen benchmarks to each AI's flaws, boosting scores 9 points

This $399 Bipedal Robot Ships Its Entire Training Stack, Not Just Demos

RLHF: The 3-step recipe behind ChatGPT's uncanny alignment

Hugging Face's New AI vs AI Tool Ranks Bots With ELO, Starting at 1200

Hugging Face's $399 duck robot ships before Christmas

Arga Labs raises $10M to build digital twins for AI agent training

IBM's Granite 4.2 trains on 15T tokens and learns to code inside real sandboxes

Harvey's Tenet model doubles legal task completion but ships no weights—just a recipe

DeepMind Alumni's Tiny 27B Model Beats GPT-5.5 at Reproducing Science

Stable Diffusion gets RLHF-style fine-tuning with TRL's new DDPOTrainer

RLHF Reproduction Nails OpenAI's Learning Curves—With a Crucial Adam Catch

RL pioneer Richard Sutton calls synthetic data a 'big mistake' for scaling LLMs

OpenAI Halts 'Astra' Training Over Fears It's Close to Real Cyberattack Power

OpenAI adds 30-minute alerts and network isolation after model escape

OpenAI admits AI hacked Hugging Face, pauses RL runs to lock down labs

ByteDance's CUDA Agent trains AI to write GPU kernels 2.11x faster than torch.compile

AI models can't beat 70 handcrafted puzzle games humans solved 100%

xAI Co-Founder's Startup River AI Lands $1.1B to Rebuild AI Training From Scratch

Liquid AI's new 2.6B model runs AI agents at 220 tok/s on a MacBook

OpenAI's escape-artist models expose why AI reward hacking is getting harder to stop

NVIDIA’s Molt packs RL training into 8.6K lines—any agent, no fork required

Moonshot open-sources the microVM engine that trained its 2.8T-param model

Kuaishou's KAT-Coder cracks 100K repo tasks by fixing a 16% infrastructure error

Induction Labs’ 106B model masters tasks from raw video, no action labels needed

Laguna S 2.1 found a proof for a 1975 math problem — and it's an 8B model

Simulation Isn't a Game Anymore—It's How Robots Get Their Smarts

OpenAI just loaded up on some of the sharpest minds in AI

OpenAI reveals 3 moonshot projects to measure 'true' AI

Deep inverse models close the sim-to-real gap for robots

This RNN learns new tasks in one shot—no fine-tuning needed

Why counting states still beats complex exploration in deep RL

OpenAI's Universe lets AI train on GTA V, Portal, and 1,000 apps

OpenAI Agent Hacks CoastRunners, Scores 20% Higher Than Humans

Reinforcement learning policies crumble under simple adversarial attacks

Why 'optical illusions for machines' remain alarmingly easy to pull off

AI Learns by Watching You, Not Just Your Data

OpenAI taught AI agents to invent their own language

80 CPUs, 10 minutes: Evolution beats backprop for AI training

Soft Q-learning and policy gradients are secretly the same

Stochastic networks crack sparse RL rewards via skill reuse

OpenAI's Q* Strikes Again: UCB Trick Supercharges Atari Scores

MADDPG: The Algorithm Where AI Agents Learn to Cooperate and Compete

OpenAI's algorithm learns backflip from under 900 bits of human feedback

OpenAI drops MuJoCo Python library with 400% speed boost

AI Teachers That Pick Your Homework Could Solve Unsolvable Problems

OpenAI scraps A3C's async trick, says ACKTR is 10-25% pricier but far better

How remembering failure made robots learn 8x faster

OpenAI makes PPO its default RL algorithm for good reason

Parameter noise cuts RL training time in half on HalfCheetah

OpenAI drops RL-Teacher: train AI with human nods, not code

OpenAI's Dota 2 bot went from 1.5k to crushing pros in 30 days

OpenAI's LOLA teaches AI the art of tit-for-tat

OpenAI's Self-Play Robots Learn to Tackle, Duck, and Fake

Asymmetric RL uses god-mode sim data to train real-world robots

OpenAI's sim-trained robot arm pushes into reality without real-world reps

OpenAI's robots learn real-world tricks without real-world training

RL agents learn to walk, then run through mazes

OpenAI drops 8 robot training grounds and a method that learns from failure

OpenAI's Gym gets a reality check with tricky robot tasks

OpenAI's first hackathon draws 500 RSVPs for 100 seats

Why your AI’s reward signal needs to look at its own hands

OpenAI launches retro contest to break RL's memorization habit

Evolved Policy Gradients: Teaching RL Agents How to Learn

OpenAI Unleashes 1,000+ Retro Games to Break RL's Single-Task Obsession

Feyn's SQRL-35B model edges Claude Opus by inspecting databases before writing SQL

AI Learns to Predict Opponents From Just a Glimpse

OpenAI's Retro Contest ends: 923 teams, one clear lesson

OpenAI bots now beat humans at Dota 2 after 180 years of daily practice

OpenAI cracks Montezuma's Revenge with 74,500 score

VAE-like option discovery VALOR unlocks 4x more robot skills

OpenAI's Robot Hand Learned Dexterity From Scratch in a Glitchy Simulation

OpenAI Five Crushes Dota Pros, Then Gets Humbled by the Crowd

Curiosity alone beats hand-crafted rewards in 54 games

OpenAI Five Stumbles at The International, Loses Twice

Curiosity finally cracks Montezuma's Revenge, a game that broke AI

POLO combines real-time planning with offline learning for fast robot skill acquisition

OpenAI's CoinRun Exposes a Deep Overfitting Crisis in RL

OpenAI open-sources a bootcamp for deep reinforcement learning

OpenAI's first Fellows class goes from zero to core contributor in 6 months

OpenAI’s Neural MMO Trains 100 Million AI Agents in a Battle Royale

OpenAI's 8 New Scholars Are Rethinking Who Gets to Do AI

OpenAI Five faces world champs OG in final showdown April 13

OpenAI Five crushes Dota 2 champs, then retires

Why your next RL agent might need a safety net

AI Agents Invented Tool Use Just to Win Hide-and-Seek

OpenAI's robot hand solves the Rubik's Cube, but only 20% of the time

OpenAI's Safety Gym lets robots learn without the bruises

OpenAI Five crushed Dota 2 champs using 2M frames per second

OpenAI's 16-Game Gauntlet Exposes RL's Memorization Problem

OpenAI's Two NeurIPS Competitions Target RL's Biggest Weaknesses

OpenAI's RLHF Summarization Model Beats 10x Larger Systems

OpenAI taught GPT‑3 to browse the web, and it still screws up

AI Learns Minecraft by Binging 70,000 Hours of YouTube

OpenAI's o1 model beats PhDs and cracks top 500 in Math Olympiad

OpenAI's GPT-Red jailbreaks its own models 84% of the time

OpenAI drops o3 and o4-mini: Reasoning models that actually use tools

OpenAI's GPT-4o turned sycophantic. Here's what broke and why.

OpenAI slips a code agent called Codex into o3 system card

AI agents leak corporate secrets through web searches, and polite prompts can't stop them

Meta, Nvidia, and Hugging Face are building a universal socket for AI agents

Four V1 bugs nearly sank our RL training — here’s what vLLM 0.18.1 actually fixed

Qwen 3 8B Learns to Shop: RL Training Produces AI That Can Handle Returns, Bundles, and Multi-Step Orders

TRL v1.0: The 3M-Download Library Built to Survive AI's Chaotic Evolution

Your GPUs Are Idle 60% of the Time—Here’s How Async RL Fixes That

GPT-OSS training blew up until LinkedIn engineers fixed a sneaky MoE routing bug

Meta and Hugging Face drop OpenEnv: a shared sandbox that could standardize how AI agents use tools

A tiny 0.6B model just crushed formal math proofs, topping the MiniF2F leaderboard

Kimina-Prover hits 92.2% on math benchmark by learning to reuse its own lemmas

PipelineRL ditches value functions, matches Open-Reasoner-Zero with inflight weight updates

How a 3B open model taught itself to rethink math—no human data needed

DeepSeek-R1 cracked open: New Open-R1 project reverse-engineers the secret RL recipe

RLOO Slashes RLHF Memory by 70% While Matching PPO Quality

JAT: One Transformer That Plays Atari, Controls Robots, and Reads Wikipedia