Local Coordination for Efficient Small-Model RL on TPUs
cs.LG, cs.DC
Submitted: 2026-11-12
Updated: 2026-11-18
Code: https://github.com/google/tunix
Terminology
Sources
- Pathways: Asynchronous Distributed Dataflow for ML
- TorchTPU - A Hybrid Eager/Compiled Architecture for Native PyTorch Execution on TPU
- Small Language Models are the Future of Agentic AI
- Deep reinforcement learning from human preferences
- Training Verifiers to Solve Math Word Problems
- Reinforcement Learning for Reasoning in Small LLMs: What Works and What Doesn't
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- SEED RL: Scalable and Efficient Deep-RL with Accelerated Central Inference
- AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning
- Gemma 4 Technical Report
- The Llama 3 Herd of Models
- HetRL: Efficient Reinforcement Learning for LLMs in Heterogeneous Environments
- Podracer architectures for scalable Reinforcement Learning
- RAPID: An Efficient Reinforcement Learning Algorithm for Small Language Models
- Ragged Paged Attention: A High-Performance and Flexible LLM Inference Kernel for TPU
- Efficient Memory Management for Large Language Model Serving with PagedAttention
- Tulu 3: Pushing Frontiers in Open Language Model Post-Training
- AXLearn: Modular, Hardware-Agnostic Large Model Training
- MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases
- ReaL: Efficient RLHF Training of Large Language Models with Parameter Reallocation
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis