Multimodal Instruction-Tuning Alters LLM Identity Encoding.

This study delves into how transformer language models geometrically encode identity-specifying system prompts within their hidden states. It compares four open-weight models across different post-training regimes: a base model, a multimodal RLHF model, an RL distillation model, and an SFT model. The researchers used various geometric metrics, primarily the 1-Wasserstein distance on Ollivier-Ricci curvature distributions, to analyze the hidden-state trajectories generated by different prompt conditions. A central finding is a significant qualitative shift in how identity is encoded after multimodal instruction-tuning. In the base model, the identity "fingerprint" is primarily direction-coded, meaning the distinctiveness comes from the orientation of the hidden state vectors. However, in the multimodal instruction-tuned model, this directional separation collapses, and the encoding migrates into the magnitude of the hidden states. This means the identity is distinguished more by the "strength" or "size" of the vector rather than its specific direction. Crucially, this "direction-to-magnitude" reorganization is unique to the multimodal instruction-tuning regime and was not observed in models trained with RL distillation or standard supervised fine-tuning (SFT). The research also introduces the use of W_1 on edge-wise Ollivier-Ricci distributions on k-NN trajectory graphs as a novel methodological contribution for analyzing transformer internal states.