Thought Engine: Learnable Cognitive Modules for LLM Agent Networks

We introduce the Thought Engine, where small sets of learnable embedding vectors (soft prompts) are injected into a frozen LLM's key-value cache to serve as silent cognitive modules. Three specialized modules (reasoning, imagination, critique) comprising only 73,728 parameters control a 1.5-billion-parameter frozen Qwen2.5 model—a control ratio of 1:20,000. The Thought Engine improves response quality by +13.3% over baseline while adding negligible inference cost.

Also available in French: Thought Engine : Modules Cognitifs Apprenables pour Réseaux d'Agents LLM