IRREPPROMPT: Symmetry-Preserving Prompt Tuning for Parameter-Efficient Molecular Force Fields

Pretrained equivariant interatomic potentials provide transferable geometric representations, but specializing them to a new molecule or thermodynamic regime can still require scarce quantum-mechanical labels. Existing equivariant parameter-efficient fine-tuning (PEFT) methods adapt symmetry-compatible operators or directly modulate equivariant features. We study a complementary question: can adaptation instead be parameterized by routing intermediate irreducible representations over a finite bank of learned prompts? We introduce IRREPPROMPT, which softly routes invariant node features over a small bank of multiplicity-space prompts. The resulting prompt translates even scalar channels and rescales complete higher-order irreducible blocks, yielding a lightweight representation-space adaptation mechanism that preserves geometric symmetry. We prove that the prompt map is O(3)-equivariant and, consequently, that its insertion into an equivariant energy model preserves energy invariance and force equivariance. On rMD17, standalone IRREPPROMPT adaptation improves over full fine-tuning in 18 of 20 reported energy/force comparisons. More importantly, IRREPPROMPT composes effectively with ELoRA: across rMD17, 3BPA, and AcAc, IRREPPROMPT+ELoRA improves ELoRA in 29 of 32 comparisons. These results identify representation prompting as a lightweight and complementary adaptation mechanism for molecular force fields under limited and shifted data.