GTAGym: Real AAA Game Tasks Environment for Evaluation and Reinforcement Learning
GTAGym turns GTA V-inspired missions into a verifiable benchmark and RL environment, revealing where visual agents still fail at dynamic control.
ls ./papers | grep "#reinforcement learning"
Trending AI research papers from Hugging Face, summarized in plain language: what each paper does, why it matters and where to read more.
12 papers · #reinforcement learning
GTAGym turns GTA V-inspired missions into a verifiable benchmark and RL environment, revealing where visual agents still fail at dynamic control.
A new reinforcement-learning approach teaches language models to study textbooks while solving unfamiliar proofs, improving transfer beyond fixed retrieval.
EvoMemState treats personal-agent memory as evolving state, linking facts through updates and conflicts while learning how each memory should influence a request.
Microsoft’s Mage-VL-Cowork trains a 4B computer-use agent to combine GUI actions, code, and terminal tools through evolved harnesses and a five-stage curriculum.
PrimeSim turns a video generator into an action-conditioned robot simulator, combining cross-embodiment training, multiview attention, self-rollouts, and hierarchical memory.
NorthCUA combines executable tasks, persistent computer-use environments, verified trajectories, staged learning, and reinforcement learning to improve long-horizon desktop and web agents.
IntReal introduces a 100,000-sample dataset and seven-dimension benchmark for testing whether unified multimodal models can coordinate text, images, context, and sequence structure.
DTCA assigns reinforcement-learning credit to diffusion-model tokens using denoising confidence gains, POS grouping, and entropy calibration, improving reported math reasoning results.
EVIDENCECONNECT synthesizes multi-region visual questions to train models that connect evidence, improving fine-grained, spatial, and long-context multimodal perception.
TTME, a memory-based test-time adaptation method, outperforms costly test-time training in scientific discovery tasks, improving efficiency and results across models and frameworks.
ThinkWorld introduces Chain-of-Dynamics reasoning reinforced by world imagination, enabling robots to learn from human videos and generalize to unseen tasks and environments.
Multiscale Policy Optimization (MPO) improves long-horizon LLM agents by aligning credit assignment and policy updates across turns and tokens, boosting performance on diverse tasks.