Signal
8 articles covering "reinforcement-learning"
Hugging Face's TRL v1.17.0 adds a single-node LoRA example for Mistral Large 4, changes the AsyncGRPO default objective to score centering and makes GRPOConfig raise on top_p or min_p without top_k when vLLM importance sampling is on.
Researchers from CMU, NYU, Stanford and MIT report in Nature that Ataraxos beat Pim Niemeijer with 15 wins, 1 loss and 4 draws, using far less compute than DeepMind's DeepNash.
EvoHarness-RL trains Qwen3-8B to hit 96.9% on ALFWorld, edging out Claude Opus 4.5's 96.4% baseline by rethinking agent memory and state.
A KAIST team uses optimal transport over video-foundation-model embeddings to learn robot rewards from a handful of preference labels, winning a top-0.7% ICML 2026 oral slot.
Boston Dynamics engineers detail the five-step reinforcement learning pipeline that taught Atlas to lift over 100 pounds using whole-body control — from animation to factory floor in weeks.
Sony AI's Project Ace, published in Nature on April 22, 2026, is the first autonomous robot to win competitive matches against elite and professional table tennis players under official rules.
Researchers from MIT and warehouse automation firm Symbotic have developed a hybrid deep reinforcement learning system that dynamically prioritizes hundreds of autonomous robots, preventing traffic jams and increasing package throughput by 25 percent in simulated warehouse environments.
Nvidia open-sources a 30B mixture-of-experts model that activates only 3B parameters yet matches DeepSeek's 671B model on olympiad-level math and coding benchmarks.