LLM reasoning, forgetting share root cause

This research reframes LLM "forgetting" as an architectural inability to reorganize dependent knowledge, essentially turning learning into an overwrite process.

• –An "External Metabolism Pipeline" organizing contradictions by time boosted logical consistency from 21.1% to 73.3% in long-turn dialogues.
• –"Structural Forgetting" describes the collapse of entire knowledge chains when a single underlying premise is modified during fine-tuning.
• –LoRA-based updates behave like overwriting rather than cumulative learning across model sizes up to 72B.
• –The framework mathematically models "Structural Persistence Potential" as an exponential decay driven by the log-ratio of state space reduction.