Moir: Let the Model Direct Its Own Story for Robust Cross-Domain Knowledge Editing
Moir addresses the asymmetric degradation of core capabilities (math/programming) during knowledge editing by aligning preservation distributions with the model's operative distribution. The method estimates preservation covariance directly from the model’s own decoding distribution using random vocabulary token seeding, bypassing instruction-following templates. Moir requires no external data and serves as a drop-in component for covariance-based editors like MEMIT and AlphaEdit. Experiments on
Analysis
TL;DR
- Moir addresses the asymmetric degradation of core capabilities (math/programming) during knowledge editing by aligning preservation distributions with the model's operative distribution.
- The method estimates preservation covariance directly from the model’s own decoding distribution using random vocabulary token seeding, bypassing instruction-following templates.
- Moir requires no external data and serves as a drop-in component for covariance-based editors like MEMIT and AlphaEdit.
- Experiments on OLMo-2, Llama-3.1, and Qwen-3 show significant improvements in retaining reasoning abilities, with Qwen-3-8B maintaining 79.9% GSM8K accuracy after 20,000 edits compared to 10.9% with Wikipedia baselines.
Why It Matters
This research solves a critical bottleneck in deploying knowledge editing systems: the collapse of complex reasoning skills while factual recall remains intact. By demonstrating that models can self-supervise their own preservation distributions, it offers a scalable, data-efficient solution for maintaining model utility during continuous updates. This is particularly relevant for practitioners managing large-scale, iterative knowledge updates where access to original training corpora is restricted.
Technical Details
- Problem Identification: Asymmetric capability erosion is traced to a distributional mismatch where static external corpora fail to capture the shifted manifold created by post-training steps like SFT and DPO.
- Methodology: Moir estimates the preservation covariance $C$ by sampling from the model's internal decoding distribution. Seeding generation with a single random vocabulary token exposes broader internalized subspaces without being dominated by instruction-following templates.
- Implementation: It acts as a modular drop-in component for existing covariance-based editors (MEMIT, AlphaEdit), requiring no additional external datasets or access to proprietary pre/post-training corpora.
- Evaluation: Tested across OLMo-2, Llama-3.1, and Qwen-3 (7-8B) models in both batch and sequential editing regimes, focusing on encyclopedic recall and mathematical/programmatic reasoning benchmarks like GSM8K.
Industry Insight
- Self-Supervised Preservation: Organizations should prioritize methods that leverage the model's internal state for preservation constraints, reducing dependency on hard-to-obtain or non-existent training data.
- Robustness in High-Volume Editing: For systems requiring frequent knowledge updates, integrating Moir-like mechanisms is essential to prevent the rapid degradation of reasoning capabilities, which are often more valuable than simple factual recall.
- Standardization of Editing Metrics: The stark contrast between baseline and Moir performance highlights the need for rigorous evaluation of reasoning retention, not just factual accuracy, when assessing knowledge editing tools.
Disclaimer: The above content is generated by AI and is for reference only.