COCONUT - Training LLM to Reason in Latent Space
Meta FAIR's Coconut paper chains hidden states as "continuous thoughts". This enables superposition where one latent vector encodes multiple branching paths, yielding emergent breadth-first search for planning-heavy reasoning, all while bypassing discrete token bottlenecks.
• Superposition in latent vectors: One continuous thought encodes multiple branching paths → emergent BFS exploration. Precisely mirrors Insight Clusters holding relational tensions (reinforce/contradict/echo/recursive) in a single potent node.
• Bypass language bottlenecks: Verbal CoT wastes compute on coherence fluff; Coconut chains silently in unrestricted latent space for efficiency on planning-heavy tasks. Validates Thoughtbase distillation: strip raw data to semantic shards + structured payloads, freeing cognition from token overhead.