Prof. Zhang's team proposes a two-stage safety framework for offline RL-based #sepsistreatment employing Constraint-Penalized Q-learning combined with Implicit Q-Learning (CPQ-IQL) and a runtime safety filter. Please access the full text at

www.sciltp.com/journals/tai...

#reinforcementlearning

Aalto University is offering two PhD positions in continual reinforcement learning for real-world systems as part of an ERC Starting Grant project.

mlscientist.com/phd-continua...

#reinforcementlearning

Chronocooked is a new reinforcement learning benchmark for studying interval timing in artificial agents, built on cooking scenarios inspired by Overcooked and drawn from psychology literature. The tasks hide…

#AI #ReinforcementLearning #MachineLearning #Research
https://arxiv.org/abs/2608.16666

The University of Toulouse is offering a fully funded PhD opportunity in social reinforcement learning for the agroecological transition, jointly supervised by INRAE Toulouse...

mlscientist.com/phd-social-r...

#reinforcementlearning

I enjoyed the #EWRL workshop and discovered an awesome community !
Thank you @ewrl-org.bsky.social for organising this event and inviting me to give a talk on
Interpretable representations of complex tasks in activity recognition and robot decision-making

#ReinforcementLearning #NeuroSymbolicAI

AGAR reframes LLM program evolution as a Markov decision process over the conditioning prefix, letting RL estimators handle parent selection, mutation, diversity, memory, and credit assignment separately rather…

#GenerativeAI #Prompting #LLMs #ReinforcementLearning
https://arxiv.org/abs/2610.09215

The method formalizes algorithm search as a Markov decision process, enabling credit assignment to specific code components. This removes the need for manually tuning five constants that govern program evolution.

#ReinforcementLearning #LLM #ProgramEvolution

A meta-agent analyzes task trajectories to diagnose failures and revise tools, achieving a 62% success rate across 42 RoboDojo tasks without updating model weights.

#Robotics #SelfImprovement #ReinforcementLearning

Researchers propose using a foundation model to improve multi-agent reinforcement learning efficiency across diverse random access network optimization tasks, with convergence analysis for their…

#AI #MachineLearning #WirelessNetworks #ReinforcementLearning
https://arxiv.org/abs/2610.07550

Small AI model (~4.5B params) learns to negotiate on a single 48GB GPU, matching frontier sellers. Learning rate tuning mattered more #AI #MachineLearning #ReinforcementLearning #Negotiation

https://freegardner.com/synapse/small-ai-model-learns-to-negotiate-on-one-gpu.html

Agent Lightning v1.0：実証済みの環境ハーネスをそのまま活用する軽量エージェントRLフレームワーク

Microsoft Research Asiaが公開した約3,500行の軽量エージェントRLフレームワークAgent Lightning v1.0の技術解説。

#LLM #ReinforcementLearning #Kubernetes #OpenSource #SoftwareEngineering

#WorkingMemory #ReinforcementLearning #Cognition #Vision

#ReinforcementLearning #Books with #Algorithms to Read! #BigData #Analytics #DataScience #IoT #IIoT #PyTorch #Python #RStats #TensorFlow #Java #JavaScript #ReactJS #CloudComputing #Serverless #DataScientist #Linux #Programming #Coding #100DaysofCode
geni.us/93-RL-AI

A new arXiv method, FC-SWE, uses failure-conditioned trajectories in reinforcement learning to help long-horizon software engineering agents recover from failed patches by reusing diagnostic feedback from…

#SWEAgents #ReinforcementLearning #DevTools #AIResearch
https://arxiv.org/abs/2610.07898

A new arXiv paper shows that for vision-language-action models like π0.5 and GR00T N1.5, applying image augmentation only to the critic during reinforcement learning post-training boosts out-of-distribution success…

#Robotics #VLA #ReinforcementLearning #Robustness
https://arxiv.org/abs/2610.05994

A new arXiv paper, CORE-RL, proposes a pipeline for evaluating black-box reinforcement learning policies against multi-objective safety and robustness specifications without needing internal access. The framework…

#AI #reinforcementlearning #cybersecurity
https://arxiv.org/abs/2610.04418

GPlaceRL is an open-source framework applying graph reinforcement learning to detailed placement refinement, using graph representations of legalized placements with modular support for encoders, policies,…

#GPlaceRL #ReinforcementLearning #OpenSourceAI #ChipDesign
https://arxiv.org/abs/2610.06489

Deep RL just shattered its depth limit: 1000-layer networks now deliver 2-50x gains in self-supervised locomotion & manipulation, with #DeepRL #ReinforcementLearning #AIResearch #MachineLearning

https://freegardner.com/synapse/deep-rl-depth-scaling-breakthrough.html

MATE is a new multi-agent reinforcement learning method that uses visual attention on salient objects to help AI agents coordinate with unseen partners zero-shot. Unlike symmetry-breaking, it…

#AI #ReinforcementLearning #MultiAgentSystems #ZeroShotCoordination
https://arxiv.org/abs/2610.06025

New arXiv paper VideoResearchAgent introduces a training framework for agents that autonomously discover and reason over open-web video, using controllable task synthesis and sim-to-real RL to…

#OpenSourceAI #DeepResearch #VideoReasoning #ReinforcementLearning
https://arxiv.org/abs/2610.04911