Implicit Weight Patch Theory
🔗 Source: arXiv
Equivalence of Context and Parameter Updates in Modern Transformer Blocks
🚀 Technical Novelty
- Mechanism: Derives a constructive proof and algorithm showing that the entire computational effect of a prompt can be exactly absorbed into rank-1 patches on MLP weights and RMSNorm scales, unified under “input controllability” and “output controllability” properties.
- Nuance: Extends prior vanilla-transformer proofs to bias-free modern architectures (Gemma/Llama), multi-layer networks, gating, normalization, and MoE blocks, eliminating the historical reliance on bias terms for context absorption.
💡 Yield
- Establishes a unified theorem proving implicit weight updates exist across diverse modern LLM architectures.
- Empirically validates near-perfect logit matching and identical token generation between patched models (without context) and original models (with context) on Gemma 3.
⚠️ Limitations
- Derived parameter updates are strictly token-dependent and require recomputation at every generation step.
- Serves as a descriptive theoretical framework for understanding ICL rather than a prescriptive method for efficient inference or cross-step patch reusability.