HARTS: A Prefix-Sharing Training System for Agentic RL with Hybrid-Attention Models

arXiv:2608.28158 · cs.LG, cs.DC · Submitted 2026-08-28 · Read on arXiv

cs.LG, cs.DC

Submitted: 2026-08-28

Updated: 2026-10-26

Code: https://github.com/NVIDIA/TransformerEngine

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Terminology

• Reinforcement Learning — 1× in this paper · explained in 4 episodes

Sources

• $\tau^2$-Bench: Evaluating Conversational Agents in a Dual-Control Environment
• DeepSpeed Ulysses: System Optimizations for Enabling Training of Extreme Long Sequence Transformer Models
• Kimi Linear: An Expressive, Efficient Attention Architecture
• Efficient Sequence Packing without Cross-contamination: Accelerating Large Language Models without Impacting Performance
• Proximal Policy Optimization Algorithms
• DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
• Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism
• Retentive Network: A Successor to Transformer for Large Language Models
• Tree Training: Accelerating Agentic LLMs Training via Shared Prefix Reuse
• Qwen3 Technical Report
• AREAL-DTA: Dynamic Tree Attention for Efficient Reinforcement Learning of Large Language Models

Related papers

• Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
• AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
• Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
• Convergence issues in Relational Concept Analysis based on AOC-posets
• Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
• Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks