LLMs Can Learn from Other LLMs Without Seeing Tokens

The evolution of natural language processing has been driven by increasingly powerful transferable representations, from word embeddings to contextual embeddings. In contrast, modern large language models (LLMs) are primarily adapted through fine-tuning, while their hidden representations are rarely treated as reusable features. We revisit this paradigm by asking whether the hidden representations of one pretrained LLM can serve as transferable representations for another. We propose a sequential architecture in which a frozen backbone LLM generates contextual hidden representations that are projected directly into the hidden space of a second LLM. The backbone remains frozen throughout training, functioning as a reusable representation provider, while only the projection module and the second model's LoRA parameters are optimized. We evaluate this paradigm on seven SuperGLUE tasks using a frozen Llama-3.2-3B backbone paired with six downstream LLMs from different architecture families. Hidden representation transfer consistently improves weaker downstream models, with gains of up to 11.0 percentage points over standalone fine-tuning, while enabling a 1.5B model to outperform standalone models with substantially more parameters. The largest improvements occur on low-resource tasks, reaching 35.7 percentage points on Commitment Bank. Our results suggest that pretrained LLMs can serve not only as task solvers but also as providers of reusable hidden representations, opening a new representation learning paradigm for transferring knowledge between foundation models.