The paper introduces Representation Finetuning (ReFT) methods, which aim to adapt large neural models by editing representations rather than updating weights. The authors propose a family of ReFT methods, with a strong instance called Low-rank Linear Subspace ReFT (LoReFT), which learns task-specific interventions on hidden representations. LoReFT is a parametrization of ReFT that intervenes on hidden representations in the linear subspace spanned by a low-rank projection matrix. The authors also introduce an ablation of this method, DiReFT, which trades some performance for increased efficiency. LoReFT and DiReFT are evaluated on various benchmarks, including commonsense reasoning, arithmetic reasoning, instruction-following, and natural language understanding, showing that they achieve state-of-the-art performance with significantly fewer parameters compared to existing parameter-efficient finetuning (PEFT) methods like LoRA. The paper provides a generic ReFT training library and discusses the limitations and future directions of ReFT.The paper introduces Representation Finetuning (ReFT) methods, which aim to adapt large neural models by editing representations rather than updating weights. The authors propose a family of ReFT methods, with a strong instance called Low-rank Linear Subspace ReFT (LoReFT), which learns task-specific interventions on hidden representations. LoReFT is a parametrization of ReFT that intervenes on hidden representations in the linear subspace spanned by a low-rank projection matrix. The authors also introduce an ablation of this method, DiReFT, which trades some performance for increased efficiency. LoReFT and DiReFT are evaluated on various benchmarks, including commonsense reasoning, arithmetic reasoning, instruction-following, and natural language understanding, showing that they achieve state-of-the-art performance with significantly fewer parameters compared to existing parameter-efficient finetuning (PEFT) methods like LoRA. The paper provides a generic ReFT training library and discusses the limitations and future directions of ReFT.