Technical Summary: Super Weights in LLMs and the Failure of Selective Training

Problem Statement

Recent research (Yu et al., 2025) identified "Super Weights"—individual scalar parameters in Large Language Models (LLMs) whose removal causes performance to degrade by orders of magnitude. This finding suggested a hypothesis: if specific parameters are so critical, then fine-tuning strategies that deliberately target and update only these "important" parameters should be highly effective, potentially outperforming methods that update random or distributed subsets of parameters.

This paper challenges that hypothesis. It investigates whether parameter importance (as defined by the catastrophic failure upon removal) implies parameter trainability in isolation. The authors test whether training only Super Weights, or their immediate local neighborhoods, can successfully adapt a model, or if such sparse, coordinate-specific training leads to catastrophic failure.

Methodology

The authors conducted a comprehensive suite of experiments on OLMo-1B and OLMo-7B, with pruning replication across 10 different model families (including Llama, Phi, Mistral, and Gemma variants).

1. Validation of Super Weight Consistency

Prior work identified Super Weights via a single forward pass. The authors validated that these weights are structural properties of the model rather than input-specific artifacts by running 1,000 random WikiText-2 samples. They found that 9 specific weight positions consistently produced the largest activation spikes across 100% of inputs, confirming their structural stability.

2. Pruning Replication

The authors replicated the pruning experiments of Yu et al. (2025) across multiple architectures. They zeroed out the identified Super Weight coordinates and measured the impact on perplexity and accuracy (ARC-Easy, Winogrande). They also tested whether simple magnitude-based ranking could identify Super Weights in models not covered by the original paper.

3. Selective Training Experiments

The core of the study involved training isolated subsets of parameters while freezing the rest:

• Direct Super Weight Training: Training only the top-$k$ Super Weights ($k \in \{100, 1000, 4096, 8192\}$).
• Neighborhood Training: Training Super Weights plus all parameters within a radius of 1 (forming $3 \times 3$ patches), increasing the trainable count up to ~36,864 parameters.
• Random Control: Training an equal number of randomly chosen positions in the same layers (down proj), excluding Super Weight coordinates, to isolate the effect of sparsity vs. coordinate selection.
• LoRA Variants:
  • Vanilla LoRA: Standard low-rank adaptation on attention layers.
  • LoRA-dproj-SW-freeze: Restricting LoRA updates at attention positions whose indices correspond to Super Weight coordinates in the down-projection (MLP) layers.
  • LoRA-∆W-SW-freeze: Freezing the top-1,000 highest-magnitude positions within the LoRA update matrix ($\Delta W$) itself.

4. Theoretical Analysis

The authors provided theoretical arguments regarding:

• Intrinsic Dimensionality: Sparse coordinate-aligned updates capture only a vanishing fraction ($k/N$) of the fine-tuning subspace compared to low-rank decompositions.
• Optimizer Dynamics: Adaptive optimizers implicitly suppress updates at Super Weight positions due to the $M^2$-fold larger curvature (Hessian) at these outlier locations, where $M$ is the activation amplification ratio.

Key Results

1. Failure of Isolated Super Weight Training

Training Super Weights in isolation resulted in a complete collapse of performance, dropping accuracy to random-guessing levels (~25% for 4-choice ARC-Easy, ~50% for 2-choice Winogrande) on both OLMo-1B and OLMo-7B.

• Increasing the number of trained parameters from 100 to 8,192 provided no improvement.
• Expanding to local neighborhoods (up to 36,864 parameters) also failed to recover performance.
• Training loss decreased during optimization, but validation perplexity exploded, indicating overfitting to training data without generalization.

2. Specificity of the Failure

The failure is specific to the Super Weight coordinates, not sparsity or the module type:

• Random Control: Training an equal number of randomly chosen positions in the same down-projection layers improved accuracy over the baseline (64.18% vs. 60.65% for OLMo-1B).
• Low-Rank Control: Applying a rank-8 low-rank update to the down-projection layer also succeeded (68.77%).
• Conclusion: The collapse is caused by targeting the specific Super Weight coordinates, which breaks the relational structure the pretrained model relies on.

3. Robustness of LoRA

• Vanilla LoRA succeeded with 0.16% of parameters, achieving 66.88% (OLMo-1B) and 77.3% (OLMo-7B).
• Positional Restrictions: Freezing attention update positions that correspond to Super Weight coordinates (LoRA-dproj-SW-freeze) had no measurable effect on performance. A 10-seed ablation confirmed statistically indistinguishable results ($p > 0.05$) between vanilla LoRA and the restricted variant.
• LoRA-Internal Criticality: Freezing the highest-magnitude positions within the LoRA update matrix ($\Delta W$) also had no effect, contrasting with the catastrophic failure of freezing base model Super Weights.

4. Pruning Generalization

While pruning Super Weights caused catastrophic degradation in OLMo, Phi-3, Mistral, and Llama-3-8B (using the original coordinates), simple magnitude-based identification failed to identify critical weights in other models (e.g., Llama-3.1, Qwen2.5, Gemma-2). This suggests Super Weights are model-family specific activation outliers, not universally defined by magnitude alone.

Significance and Claims

The paper establishes that parameter importance does not imply parameter trainability in isolation. The authors argue that:

1. Structural vs. Trainable: Super Weights act as fixed, high-gain amplifiers. Modifying them in isolation breaks the coordinated structure of the network. Effective fine-tuning requires updating the representations flowing into these weights (via attention layers in LoRA) rather than the weights themselves.
2. Coordination over Sparsity: Success in PEFT relies on structured decompositions over entire layers (like LoRA's low-rank structure) rather than targeting individually important weights. LoRA succeeds because its low-rank factorization coordinates updates across the entire layer, preserving the relational structure, whereas sparse training isolates parameters and disrupts these dependencies.
3. Optimizer Dynamics: Theoretical analysis suggests that adaptive optimizers implicitly protect Super Weights by suppressing updates due to high curvature. Explicitly training these positions overrides this protection, leading to instability and collapse.

The authors conclude that future PEFT methods should derive efficiency from structured, layer-wide decompositions rather than attempting to identify and update "critical" individual parameters.