Super Weights: Why the Most Important Parameters Are the Hardest to Train in LLMs

> *"Nature's most exquisite designs are often hidden in the most inconspicuous corners."* — in the style of Richard Feynman

This is a Feynman-style deep dive into a recent paper on Super Weights in large language models and the failure of selective training (Subramanian, Akinfaderin & Sehwag, arXiv:2607.08733).

A Counterintuitive Discovery

Imagine a concert piano with a handful of strings so critical that breaking any of them ruins the entire instrument. The intuitive response would be to protect—or even specially train—those strings. Analogously, since researchers discovered Super Weights: a small set of parameters (tens to hundreds) whose removal causes a *cliff-edge* collapse in LLM performance, one might assume the most efficient fine-tuning strategy is to update only these critical parameters.

This paper shows the opposite:

> Training the most important parameters in isolation makes the entire model collapse.

What Are Super Weights?

An LLM like OLMo-7B has ~7 billion parameters. In 2024, researchers observed that zeroing out a tiny set of specific parameters causes performance to plummet not gradually but catastrophically—from coherent prose to gibberish. These parameters were named Super Weights. They act like a pacemaker: invisible in normal operation, but the system halts the moment they stop working.

The Experiments

Researchers tested "training only Super Weights" on OLMo-1B and OLMo-7B:

1. Extreme sparse training — updating only 100 to 8,192 Super Weights: accuracy drops to random-guess level. 2. Expanding the scope — training up to ~36,000 parameters in the Super Weights' neighborhood: still no improvement. 3. Control experiment — training a *randomly selected* equal number of parameters: works better than training Super Weights.

Why: Importance ≠ Trainability

Super Weights matter not because they independently "know" a lot, but because they sit at critical information hubs, receiving signals from across the network and distributing transformed outputs downstream. Their importance is relational, context-dependent, and emergent. Isolating them during training severs these connections—like a concertmaster practicing alone in a soundproof room while ignoring the rest of the orchestra.

Why LoRA Succeeds

LoRA updates only ~0.16% of parameters—sparser than training 8,192 Super Weights—yet it works. The key is structured decomposition: LoRA introduces low-rank transformations that influence *entire layers*, not isolated coordinates. The paper's decisive test:

• Applying the same low-rank update to the down_proj layer containing Super Weights: succeeds.
• Restricting the LoRA update to the exact coordinates of the Super Weights: collapses, just like direct Super Weight training.

The distinction is structure vs. isolation, not parameter count.

Deeper Implications

1. Importance is emergent — like a critical city intersection, a Super Weight's value comes from its position in the network, not its intrinsic properties. 2. Effective training needs a global view — parameter-efficient methods (LoRA, Adapters, Prefix Tuning) work by structurally influencing the whole network, not by training fewer parameters per se. 3. Rethinking compression — Super Weights are a "sensitive zone": they can neither be safely pruned nor naively trained. Future pruning/quantization methods must distinguish parameters that are safe to remove, safe to train, and best left untouched.

There is also an echo of neuroscience: hub neurons in the brain are both critically important and the hardest to modulate with local stimulation—much like Super Weights in artificial networks.

Practical Takeaways

Do:

• Use structured PEFT methods (LoRA, Adapters)
• Ask whether your method performs a local update or a global, structured influence
• If you must do selective training, random selection beats "cleverly" choosing Super Weights

Don't:

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

Conclusion

A Feynman-style twist: *what you cannot train in isolation is precisely what turns out to be the soul of the system.* Neural networks are not collections of parameters but an ecosystem; each parameter's importance derives from its position and relationships within it.

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.
• Sun, M., et al. (2024). Super Weights identification paper.
• Feynman, R. P. (1985). *Surely You're Joking, Mr. Feynman!*. W.W. Norton.