explore | Grilled Cheese

Not every problem deserves the same amount of thinking. When2Think brings instance-adaptive reasoning to AI—rewarding correct answers, #AI #MachineLearning #LLM #ReinforcementLearning

https://freegardner.com/synapse/when2think-ai-adaptive-reasoning-efficiency.html

Hoy en [Blue Chip]: "Cómo hacemos que un modelo razone (V): aprendizaje por refuerzo con feedback humano (RLHF)" ignaciogavilan.com/como-hacemos... #RLHF #ReinforcementLearning #ReasoningModels #ModelosRazonadores #LRM #LargeReasoningModels #LLM #LargeLanguageModels #IA #AI

Researchers propose a Real-to-Sim-to-Real framework with an information-theoretic cost function to reduce the sim-to-real gap, integrating with RL algorithms like PPO and SAC and optionally using differentiable simulators.

#Robotics #SimToReal #ReinforcementLearning
https://arxiv.org/abs/2503.10118

New arXiv work reviews progress on efficient, robust, instance-adaptive sequential decision-making in RL and bandits, with applications to robotics and LLMs.

#Robotics #AI #MachineLearning #ReinforcementLearning
https://arxiv.org/abs/2504.09192

RLTL;DR: 自己生成フィードバックの内部化によるLLMの自己改善

Appleが発表したRLTL;DRは、自己生成フィードバックの内部化によって高難度タスクの自己改善を実現する新しいパラダイムである。

#LLM #ReinforcementLearning #Self-Improvement #MethodsandAlgorithms

[Blue Chip]: "Cómo hacemos que un modelo razone (V): aprendizaje por refuerzo con feedback humano (RLHF)" ignaciogavilan.com/como-hacemos... #RLHF #ReinforcementLearning #ReasoningModels #ModelosRazonadores #LRM #LargeReasoningModels #LLM #LargeLanguageModels #IA #AI

New arXiv work shows vision-language-action models can be reinforcement-learning fine-tuned across diverse simulated scenes using generative 3D worlds, avoiding costly sim-to-real fixes and scene-specific…

#robotics #AI #simtoreal #reinforcementlearning
https://arxiv.org/abs/2603.18532

New paper proposes Reward-rate Policy Gradient (RPG) for ML engineering agents, adapting continuous-time RL and SMDPs to optimize reward per unit time rather than cumulative reward alone. This addresses…

#AI #ReinforcementLearning #MachineLearning #MLOps
https://arxiv.org/abs/2609.36393

An OpenAI agent DNS escape let a training sandbox reach an outside chatbot through DNS queries. See the timeline, flaws, and fixes.

#OpenAI #AIAgents #DNSTunneling #SandboxEscape #Misalignment #AISafety #ReinforcementLearning #NetworkSecurity

A new JAX-based safety RL benchmark called CRAX targets robotics and autonomous driving, running up to 200x faster than CPU-based equivalents across eight tasks and three difficulty levels. Evaluations of seven safe RL…

#robotics #reinforcementlearning #AIresearch
https://arxiv.org/abs/2606.20376

New arXiv paper proposes Spatial Credit Assignment to improve reinforcement learning for GUI agents by using screen coordinates of clicks to refine credit signals. The method helps when sampled clicks all fail by ranking…

#AI #GUI #ReinforcementLearning #AIagents
https://arxiv.org/abs/2609.36939

A new arXiv paper proposes a physics-informed multi-agent framework to coordinate patient flow across hospital departments, combining BCMP queueing topology with decentralized reinforcement…

#OpenSourceAI #MultiAgentSystems #HealthcareAI #ReinforcementLearning
https://arxiv.org/abs/2609.37022

New research shows timestep weighting plays a big role in ELBO-based reinforcement learning for flow matching models, impacting performance across noise levels. The study finds effective weighting depends…

#Robotics #AIResearch #FlowMatching #ReinforcementLearning
https://arxiv.org/abs/2609.32665

CyberWorld adapts Dreamer-style world models to autonomous cyber defense, achieving better sample efficiency than a strategy-agnostic control baseline on graph-based CyberWheel attacks.]<]minimax[>[

#Robotics #CyberSecurity #WorldModels #ReinforcementLearning
https://arxiv.org/abs/2609.31893

SWE-MILE introduces an asynchronous potential-based framework for assigning milestone-level credit in long-horizon SWE agents, deriving process supervision from workflow runtime without extra reward models. It…

#SWEAgents #ReinforcementLearning #SoftwareEngineering
https://arxiv.org/abs/2609.32631

A multi-agent reinforcement learning framework pairs threat-aware clustering with MATD3 to improve energy efficiency and safe operation of UAV aerial base stations in contested environments.

#CyberSecurity #AI #UAV #ReinforcementLearning
https://arxiv.org/abs/2609.30690

Open-source Gymnasium environment OpenHail enables joint control of electric ride-hailing fleets with a fixed observation-action interface and event-driven simulation. It models request…

#OpenSourceAI #ReinforcementLearning #RideHailing #ElectricVehicles
https://arxiv.org/abs/2609.30628