Super Weights in LLMs: Why the Most Important Parameters Fail When Trained in Isolation

> A Feynman-style paper walkthrough, translated and adapted from zhichai.net.

A Counterintuitive Discovery

Large language models like OLMo-7B contain billions of parameters. In 2024, researchers discovered that removing certain specific parameters—often just a few dozen to a few hundred, dubbed Super Weights—causes model performance to fall off a cliff, degrading from coherent text to near-random output.

This raised a natural hypothesis: if these weights are the most important, wouldn't fine-tuning *only* them be the most efficient training strategy?

The paper's answer: No—the opposite. Training these critical parameters in isolation collapses the model.

The Experiments

Researchers tested multiple "train only the Super Weights" schemes on OLMo-1B and OLMo-7B:

• Extreme sparse training (100–8,192 Super Weights): accuracy crashed to random-guess levels.
• Expanding the scope (~36,000 parameters around Super Weights): still no improvement.
• Control experiment: training randomly selected parameters of the same count worked *better* than training Super Weights—the most striking result.

Why Importance ≠ Trainability

The core insight: parameter importance is not the same as parameter trainability.

Super Weights are not important because they independently encode knowledge, but because they occupy critical information-hub positions—receiving signals from across the network and passing transformed results downstream. Their importance is *relational, contextual, and emergent*. Isolating them for training severs these connections, like removing a concertmaster from the orchestra to practice alone in a soundproof room.

The paper echoes neuroscience: like "hub neurons" in the brain, these parameters are both critical and hardest to modulate locally.

Why LoRA Succeeds

LoRA (Low-Rank Adaptation) updates only ~0.16% of parameters—sparser than training 8,192 Super Weights—yet works. The difference is structured vs. isolated updates:

• Training Super Weights alone = isolated updates at specific coordinates.
• LoRA = structured low-rank transformations that influence all positions in a layer.

The researchers confirmed this: applying the same low-rank update to the down_proj layer containing Super Weights worked; restricting LoRA updates to the Super Weights' exact coordinates caused the same collapse as direct training.

Implications

1. Importance is emergent — a Super Weight matters because of its position in the network, like a busy intersection in a city. 2. Effective training needs a global view — parameter-efficient methods work through *structure*, not sparsity alone. 3. Model compression risks — Super Weights are a sensitive zone: they cannot be safely pruned, nor safely trained in isolation. Future compression methods should distinguish parameters safe to remove, safe to train, and best left untouched.

Practical Takeaways

Do:

• Use structured PEFT methods (LoRA, Adapters, Prefix Tuning).
• Ask whether your method performs local updates or global structured influence.
• If you must do selective training, random selection may beat "smart" Super Weight selection.

Don't:

• Train Super Weights in isolation, at any scale.
• Assume important parameters are the right ones to train.
• Equate parameter efficiency with parameter count—the key is structure.

Conclusion

Neural networks are ecosystems, not collections of parameters. Each parameter's importance derives from its position and relationships. Some things are important precisely because they cannot be isolated.

References

• Subramanian, S., Akinfaderin, A., & Sehwag, A. (2026). *Super Weights in LLMs and the Failure of Selective Training*. arXiv:2607.08733.
• Hu, E. J., et al. (2021). *LoRA: Low-Rank Adaptation of Large Language Models*. arXiv:2106.09685.
• Feynman, R. P. (1985). *Surely You're Joking, Mr. Feynman!*. W.W. Norton.