ReFT (Representation Finetuning) is a parameter-efficient method for adapting large language models (LLMs) by modifying hidden representations rather than model weights. This approach leverages the rich semantic information encoded in representations, offering a more powerful alternative to traditional parameter-efficient fine-tuning (PEFT) methods like LoRA. ReFT methods operate on a frozen base model and learn task-specific interventions on hidden representations. A strong instance of ReFT, Low-rank Linear Subspace ReFT (LoReFT), intervenes on hidden representations in a low-rank linear subspace, achieving state-of-the-art performance with significantly fewer parameters than LoRA. LoReFT and its ablation DiReFT outperform existing PEFTs across multiple tasks, including commonsense reasoning, arithmetic reasoning, instruction-following, and natural language understanding. ReFT methods are drop-in replacements for weight-based PEFTs and are more efficient and effective. The ReFT framework is generalizable and can be applied to various tasks and models. The paper evaluates ReFT on multiple benchmarks and shows its effectiveness in improving model performance while maintaining efficiency. The results demonstrate that ReFT methods achieve the best balance of efficiency and performance, outperforming state-of-the-art PEFTs in most cases. The ReFT training library is publicly available for further research and application.ReFT (Representation Finetuning) is a parameter-efficient method for adapting large language models (LLMs) by modifying hidden representations rather than model weights. This approach leverages the rich semantic information encoded in representations, offering a more powerful alternative to traditional parameter-efficient fine-tuning (PEFT) methods like LoRA. ReFT methods operate on a frozen base model and learn task-specific interventions on hidden representations. A strong instance of ReFT, Low-rank Linear Subspace ReFT (LoReFT), intervenes on hidden representations in a low-rank linear subspace, achieving state-of-the-art performance with significantly fewer parameters than LoRA. LoReFT and its ablation DiReFT outperform existing PEFTs across multiple tasks, including commonsense reasoning, arithmetic reasoning, instruction-following, and natural language understanding. ReFT methods are drop-in replacements for weight-based PEFTs and are more efficient and effective. The ReFT framework is generalizable and can be applied to various tasks and models. The paper evaluates ReFT on multiple benchmarks and shows its effectiveness in improving model performance while maintaining efficiency. The results demonstrate that ReFT methods achieve the best balance of efficiency and performance, outperforming state-of-the-art PEFTs in most cases. The ReFT training library is publicly available for further research and application.