Is One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL Training
2026-07-01 • Machine Learning
Machine LearningComputation and Language
AI summaryⓘ
The authors studied how reinforcement learning (RL) changes different layers inside large language models after training. They found that only a few layers, often just one in the middle of the model, are responsible for most of the improvements from RL training. This pattern was consistent across different models, RL methods, and tasks like math, coding, and decision-making. Their work challenges the usual approach of updating all layers equally and shows that focusing on key layers might be more efficient.
reinforcement learninglarge language modelstransformer layerspost-training adaptationlayer contributionRL algorithmsmodel fine-tuningmathematical reasoningcode generation
Authors
Zijian Zhang, Rizhen Hu, Athanasios Glentis, Dawei Li, Chung-Yiu Yau, Hongzhou Lin, Mingyi Hong
Abstract
Reinforcement learning (RL) has become a central component of post-training large language models (LLMs), yet little is understood about how RL adaptation is distributed across transformer layers. Existing approaches typically update all model parameters uniformly, implicitly assuming that every layer contributes similarly to the gains obtained during RL post-training. In this work, we challenge this assumption through a systematic layer-wise study of RL training. Surprisingly, we find that training a single transformer layer can recover most of the gains achieved by full-parameter RL training, and in some cases even surpass it. To quantify this phenomenon, we introduce the quantity layer contribution, which measures the fraction of full RL improvement recovered by training a layer in isolation. Across seven models spanning two model families (Qwen3, Qwen2.5), three RL algorithms (GRPO, GiGPO, Dr. GRPO), and multiple task domains including mathematical reasoning, code generation, and agentic decision-making, we observe a remarkably stable pattern: RL gains are highly concentrated in a small subset of, and in many cases even a single, transformer layers. More strikingly, the same structural pattern consistently emerges: high-contribution layers concentrate in the middle of the transformer stack, while layers near the input and output ends contribute substantially less. The resulting layer rankings remain strongly correlated across datasets, tasks, model families, and RL algorithms.