Research map / Architectures and efficiency / Attention and transformers
Equivalence of Context and Parameter Updates in Modern Transformer Blocks
Adrian Goldwaser, Michael Munn, Javier Gonzalvo, Benoit Dherin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use numerically stable update via RMSNorm inversion instead of naive update in bfloat16 (ablation) | Implicit weight updates, Gemma 3 · Weather forecast prompt, bfloat16 | token prediction agreement: 87.5% -> 100% | Improved |
| use float32 instead of bfloat16 precision (ablation) | Implicit weight updates, Gemma 3 · Weather forecast prompt | logit difference and token match: float32 maintains perfect token matching; bfloat16 87.5% agreement | Improved |
| update only the final layer instead of several layers (ablation) | Implicit weight updates, Gemma 3 · Starting layer depth ablation | generation fidelity | Hurt |
| start update at earlier layers (spread across more layers) (ablation) | Implicit weight updates, Gemma 3 · Starting layer depth ablation | update norm | Improved |
| use scaling update variant instead of naive update (ablation) | Implicit weight updates, Gemma 3 · Scaling update variant analysis | TVD and token match in low precision | Improved |
| scaling update instead of stable update (ablation) | Implicit weight updates, Gemma 3 · Scaling update variant analysis | update norm | No clear effect |
| apply updates to Falcon (no post-norm) instead of Gemma | Implicit weight updates · Falcon-7B, bfloat16 | token matching | Improved |
| include an image in the context (Gemma 3 4B) | Implicit weight updates · Multimodal context | logit match | No clear effect |
Closest papers by meaning
- Transformer Block Coupling and its Correlation with Generalization in LLMs (ICLR 2025 · similarity 0.89)
- Relaxed Recursive Transformers: Effective Parameter Sharing with Layer-wise LoRA (ICLR 2025 · similarity 0.88)
- Trainable Transformer in Transformer (ICML 2024 · similarity 0.88)
- Bigram Subnetworks: Mapping to Next Tokens in Transformer Language Models (NeurIPS 2025 · similarity 0.88)
- Do Efficient Transformers Really Save Computation? (ICML 2024 · similarity 0.88)
- Talking Heads: Understanding Inter-Layer Communication in Transformer Language Models (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
