Growing and Transferring Pathways for LLM Reinforcement Learning
We propose GrowRL, an RL method that organizes learned updates into compact computational pathways through structural growth.
ProofPaper ↗
Key points
- Reinforcement learning (RL) can substantially improve the reasoning capabilities of large language models, but an effective training run requires considerable resources.
- GrowRL freezes the pretrained backbone and adds an independent causal attention branch and a feed-forward network branch at each Transformer layer.
- By keeping the learned attention and feed-forward computation separate from the pretrained weights, GrowRL provides an explicit structure for preserving and reusing the outcome of an RL run.
- These results support computational pathways as parameter-efficient and reusable units of reinforcement learning adaptation.
Sources (1)
- [1]Growing and Transferring Pathways for LLM Reinforcement LearningarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 09:43 AM
We propose GrowRL, an RL method that organizes learned updates into compact computational pathways through structural growth.
Reinforcement learning (RL) can substantially improve the reasoning capabilities of large language models, but an effective training run requires considerable resources.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 12, 2026AReA: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers
- Oct 11, 2026AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards
- Oct 11, 2026A Sovereign, Indic-First, Reasoning Model
- Oct 10, 2026Single-step Retrosynthesis via Reinforcement Learning from Forward Feedback
- Oct 10, 2026Target-ESS Tempered Policy Optimization (TETPO): Smooth Variance Control via Bridge Importance Sampling
- Oct 10, 2026BridgeWAM: Bridging Video and Action Experts via Latent Bridge Queries