Interpretability

Works

QKV Decomposition for Transformer XAI Paper · 2026

Diagnose transformer prediction failures from weights alone, then correct them by retraining one layer. GPT-2 capital-city accuracy 2/8 → 8/8 with zero side effects, achievable through any of attention, FFN, or V-only (590K params).

LLM
Interpretability
Model editing
Transformers
Insights on intelligence

Zenodo (DOI) ↗
HuggingFace dashboard ↗

Notes

2026-05-10
Deep learning is editable

We treat deep learning as something you train, not something you fix. That's a habit, not a fact. I edited 590,000 parameters in GPT-2 — half a percent of the model — and corrected six wrong answers without damaging anything else.
2026-05-10
Deep learning is editable

딥러닝을 고치는 도구는 늘 '재학습'이었다. 그건 사실이라기보다 습관이다. GPT-2의 0.5%만 한 번 손봐서 틀린 답 여섯 개를 모두 고쳤다. 다른 능력은 그대로다.

Interpretability

Works

Notes

Deep learning is editable

Deep learning is editable