Universal Attention Heads Found Across LLMs: Paper Reveals Neuron Scaling Laws
CSProfKGD · x · 2026-08-05
Nikhil Prakash utilized Goodfire's Silico platform to discover universal "Rosetta heads"—attention heads with similar attention patterns and OV circuits across 14 different language models of varying sizes and families.
This connects to a recent arXiv paper by Amil Dravid et al., Neuron Populations Exhibit Divergent Selectivity with Scale, which investigates how internal network structures evolve with scale. Key findings include:
- Sublinear Power Law: The absolute number of cross-model "Rosetta Neurons" grows with model size, but their fraction of total neurons shrinks.
- Neuron Polarization Effect: Rosetta Neurons become increasingly selective and monosemantic at larger scales, separating from a growing non-Rosetta population that remains less selective.
- Practical Implications: The study illustrates how these highly selective neurons can be used for targeted data filtering in continued pretraining, establishing a scaling law for interpretable neuron-level structures.
Related event: Study Reveals Universal Attention Heads Across LLMs(2 posts)→
More from Research
- Researcher Calls for a Refutations & Critiques Track at NeurIPS — RylanSchaeffer · 2026-08-05
- MuScriptor: Open-Weight Multi-Instrument Music Transcription Model Released — drscotthawley · 2026-08-05
- Zero-Latency Guitar Modulation Effects Modeling via Differentiable DSP — drscotthawley · 2026-08-05
- Omanta (YC S26): AI Personalized Research Lab for Cancer Patients — ycombinator · 2026-08-05
- Self-Reflection Fails: Paper Shows Model Introspection Drops Accuracy Up to 10% — omarsar0 · 2026-08-05
- Rumor: SSI Solved Continual Learning, But Alignment Becomes Harder — teortaxesTex · 2026-08-05