Moir: New AI editing method preserves 79.9% math accuracy vs 10.9% baseline
Model's own outputs can prevent reasoning collapse during knowledge editing.
Knowledge editing lets LLMs update facts without full retraining, but often destroys core reasoning abilities — math accuracy can plummet while encyclopedic recall stays intact. Researchers trace this to a distributional mismatch: covariance-based editors preserve subspaces from static external corpora (like Wikipedia) that don't match the model's post-training distribution (shaped by SFT/DPO). They propose Moir, which estimates the preservation covariance directly from the model itself by sampling its own decoding distribution. A single random token seed bypasses instruction-following templates, exposing the broader subspaces the model has internalized. Moir requires no external data and works as a drop-in component for any covariance-based editor (MEMIT, AlphaEdit).
Tested on OLMo-2, Llama-3.1, and Qwen-3 (7–8B) in batch and sequential regimes, Moir consistently extends preservation in vulnerable domains. Most striking: on Qwen3-8B after 20,000 AlphaEdit batch edits, Moir retains 79.9% GSM8K accuracy compared to 10.9% with the Wikipedia baseline. These results show aligning the preservation distribution with the model's operative distribution is key to non-destructive editing — and the model itself may be the most accessible source of that distribution for deployed systems.
- Moir estimates preservation covariance from the model's own generation (starting with a random token) to capture post-training distribution.
- On Qwen3-8B after 20,000 AlphaEdit batch edits, Moir retained 79.9% GSM8K accuracy vs 10.9% with Wikipedia baseline.
- Drop-in replacement for covariance-based editors (MEMIT, AlphaEdit) — no external data required.
Why It Matters
Enables safe knowledge updates in deployed LLMs without sacrificing mathematical and reasoning capabilities.