Search papers, labs, and topics across Lattice.
This paper investigates the feasibility of compressing task-relevant information from instruction prompts into a single activation vector for large language models (LLMs), aiming to enhance efficiency by reducing computational overhead. The authors demonstrate that a learned weighted sum of activations from an intermediate layer can effectively replace the original token sequence, achieving less than a 2% accuracy drop in performance. Their findings reveal significant insights into the structure of activation space, indicating that mid-layer representations can meaningfully transfer to early layers and that a single activation vector can encapsulate substantial semantic information.
Compressing prompts into a single activation vector can cut computational costs while maintaining nearly full accuracy in LLM responses.
Large language models process prompts by propagating activations through dozens of layers before generating a response. We ask whether the task-relevant information contained in an instruction prompt can be compressed into a single activation vector and re-injected into the model, replacing the original token sequence? We show this is achievable using a learned weighted sum of activations extracted at an intermediate layer and injected at an early layer of the target LLM. The compressed vector preserves task-relevant information, incurring an accuracy drop of under $2\%$ relative to full prompt processing. Beyond its practical implications, including reducing per-query computation for fixed instruction prompts without reprocessing the original token sequence, our analysis reveals structure in the activation space of LLMs: (i) mid-layer representations transfer meaningfully to early layers, suggesting a degree of cross-layer compatibility in how information is encoded; (ii) a single activation vector encodes a quantifiable and recoverable amount of semantic information; (iii) a weighted sum of activations is a robust representation compressor.
