Latent Thought Flow

Summary

Latent Thought Flow (LTF) is the method introduced by Latent Thought Flow: Efficient Latent Reasoning in Large Language Models. It trains an LLM to sample variable-length continuous latent reasoning trajectories with a continuous GFlowNet objective, then decode only the final answer.

Role In The Wiki

Use this page as the object card for the method. The source page carries the evidence details, limitations, and agenda mapping.

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in latent-thought-flow-2026 rather than duplicating verdict rows here. At the entity level, LTF is relevant as an adjacent dynamic-compute and multi-trajectory latent-state mechanism: it suggests how hidden trajectories could be sampled and scored by a quality/cost reward, but current evidence is LLM reasoning rather than multivariate time-series or action-conditioned world-model evidence.

Evidence

• Latent Thought Flow: Efficient Latent Reasoning in Large Language Models

Official Artifacts

• Preprint: arXiv 2606.16222
• Review: ArXivIQ review
• Code status: no official repository was verified at ingest time; Gonzo/ArXivIQ list Code: N/A.
• Model status: no official checkpoint was verified at ingest time; Gonzo/ArXivIQ list Model: N/A.

Related Pages

• Looped Transformers And Test-Time Memory
• LLM Post-Training
• Training Dynamics
• Representation Collapse
• Time-Series Scaling And Efficiency
• Foundation Time-Series Model Research Agenda
• GRAM
• Latent Thoughts
• Latent CoT Superposition Analysis