Supernodes and Halos: Loss-Critical Hubs in LLM Feed-Forward Layers

Abstract: We study the organization of channel-level importance in transformer feed-forward networks (FFNs). Using a Fisher-style loss proxy (LP) based on activation-gradient second moments, we show that loss sensitivity is concentrated in a small set of channels within each layer. In Llama-3.1-8B, the top 1% of channels per layer accounts for a median of 58.7% of LP mass, with a range of 33.0% to 86.1%. We call these loss-critical channels supernodes. Although FFN layers also contain strong activation outliers, LP-defined supernodes overlap only weakly with activation-defined outliers and are not explained by activation power or weight norms alone. Around this core, we find a weaker but consistent halo structure: some non-supernode channels share the supernodes' write support and show stronger redundancy with the protected core. We use one-shot structured FFN pruning as a diagnostic test of this organization. At 50% FFN sparsity, baselines that prune many supernodes degrade sharply, whereas our SCAR variants explicitly protect the supernode core; the strongest variant, SCAR-Prot, reaches perplexity 54.8 compared with 989.2 for Wanda-channel. The LP-concentration pattern appears across Mistral-7B, Llama-2-7B, and Qwen2-7B, remains visible in targeted Llama-3.1-70B experiments, and increases during OLMo-2-7B pretraining. These results suggest that LLM FFNs develop a small learned core of loss-critical channels, and that preserving this core is important for reliable structured pruning.

• The paper identifies that a sparse subset of FFN channels, called supernodes, captures most loss sensitivity, with the top 1% channels accounting for a median of 58.7% of the loss proxy mass.
• It demonstrates that surrounding halo channels exhibit high redundancy with supernodes yet have a minimal individual impact on performance, highlighting a clear morphological separation.
• The study shows that structured pruning methods like SCAR, which protect supernodes, are crucial for maintaining low perplexity, as supernode ablation leads to drastic performance degradation.

Supernodes and Halos: Loss-Critical Hubs in LLM Feed-Forward Layers

Background and Motivation

Channel-level component analysis in LLMs has revealed extreme heterogeneity in importance, often observed as rare but critical activation events or "super weights." This work formalizes this observation in the context of feed-forward network (FFN) channels: most loss-sensitivity in LLMs arises from a highly restricted subset of these channels, termed supernodes. Loss sensitivity is measured via a Fisher-inspired loss proxy (LP) statistic that jointly captures activation and backpropagated gradient structure.

Loss Proxy and Supernode Discovery

The loss proxy ($LP_i = \frac{1}{2} \mathbb{E}[(u_i s_i)^2]$ for channel $i$) captures channels whose masking would cause largest increases in the empirical Fisher information of model loss. Empirically, in Llama-3.1-8B, the top 1% of channels per layer capture a median of 58.7% of total LP mass, with layer-level ranges between 33.0% and 86.1%. Supernodes show a max-to-mean LP gap regularly exceeding $10^3\times$ per layer, indicating extreme outlier status not attributable purely to activation magnitude or weight norms. This concentration is reproduced across Mistral-7B, Llama-2-7B, Qwen2-7B, and Llama-3.1-70B, and during progressive pretraining of OLMo-2-7B.

Figure 1: Supernodes (red) and halos (orange) localize LP concentration in FFN channels. The top 1% of channels dominate the LP mass, especially in early and late layers.

A key finding is that supernodes, identified by LP, only partially overlap with activation outliers or activation-defined top channels (Jaccard overlap ≈ 5–10% at default fraction), and LP-activation rank correlations are moderate. This is further quantified via domain stability measurements: LP concentration as a phenomenon is robust, but the actual channel identities are domain-conditioned.

Figure 2: Write-halo channels exhibit more redundancy with the supernode core than do non-halo counterparts. LP and activation outlier sets have weak overlap.

Halo Structure: Redundancy and Topology Around the Core

Surrounding the sharply concentrated supernode cores, the authors identify "write halos": non-supernode channels whose output-support (as defined by the $W_\text{down}$ column structure) considerably overlaps that of supernodes. These halos are defined as the top-10% of non-supernodes by write pattern connectivity to the aggregated supernode output support. Using a directed redundancy metric based on correlation of loss-relevant contributions, halo channels are shown to be more redundant with supernodes than matched controls, yet their ablation has much weaker effects than supernode ablation. This reveals a morphological separation: while halos encode shared read/write structure, true functional centrality is reserved to the sparse supernode set.

Structured Pruning and Failure Modes

The channel-level LP organization has significant implications for practical LLM pruning. The authors implement the SCAR (Supernode-Constrained Allocation for Resource-aware pruning) method, which strictly protects the LP-defined supernode set and uses further connectivity and redundancy-driven ranking for the remaining channels.

At 50% FFN channel sparsity, SCAR achieves perplexity of 54.8–55.8 on Llama-3.1-8B, while channel-adapted Wanda and SparseGPT (which prune significant fractions of supernodes) degrade to perplexities of 989.2 and 5256.3, respectively—a gap of over an order of magnitude:

Figure 3: Across Llama, Mistral, Qwen, and Llama-2, SCAR-LP (solid lines) maintains markedly lower perplexity at aggressive sparsity compared to baselines, with divergence widening as sparsity increases.

Furthermore, direct supernode removal, either by explicit highest-LP ablation or by random masks with controlled supernode hit-rate, causes catastrophic loss increases. This dose–response validates the core hypothesis: extreme functional reliance is localized in the supernode set.

Figure 4: Downstream accuracy and perplexity changes track directly with supernode hit-rate—removal of more supernodes translates to marked performance collapse.

Emergence During Pretraining

Longitudinal checkpoint analysis during OLMo-2-7B pretraining reveals that LP concentration and supernode structure are not static, but arise over the course of optimization. The top-1% LP mass fraction begins near 18.6% at early checkpoints and stabilizes over 70% late in training, with the set identities stabilizing later than the bulk concentration itself.

Figure 5: Top-1% LP mass and max/mean LP ratio both increase steadily during OLMo-2-7B training, while Jaccard overlap with final supernodes rises in later stages.

Robustness Across Models and Domains

LP concentration and the SCAR robustness effect generalize across architectures, scales (including 70B parameters), and input domains. Although the exact identity of LP supernodes may drift with domain (mean Jaccard ≈ 6% at 1% threshold), the loss-concentration pattern and necessity of supernode protection for successful pruning remain invariant.

Theoretical and Practical Implications

This work firmly establishes that functional capacity in LLM FFN layers is highly non-uniform: an extremely sparse set of channels mediate most loss sensitivity, with "halo" neighborhoods of redundant but sub-critical channels. The implications are twofold:

• Pruning: Channel-level pruning must explicitly isolate and preserve these supernodes. Failing to do so results in dramatic model degradation at moderate sparsities.
• Mechanistic Interpretability: Supernodes serve as clear targets for mechanistic, diagnostic, and even causal interventions. Their LP-based definition is stronger than mere activation magnitude and more informative about true loss gradients.
• Compression and Domain Adaptation: Calibration-domain dependence implies that adaptive recalibration (i.e., re-estimating LPs) may be needed for domain-specific deployment, bridging mechanistic insights and practical resource constraints.

Limitations and Future Directions

Supernode identification is calibration data-dependent; thus, distributional or task shifts may change the relevant supernode set. The authors restrict their study to FFN channels—analogous loss-critical cores in attention mechanisms, MoE routers, or other architectures remain open for study. Further, the paper uses a Fisher/Gauss–Newton LP approximation; non-Gaussian or higher-order perturbative effects are not explicitly modeled.

There is substantial future scope in using LP-based analyses for fine-tuning, quantization robustness, dynamic routing policies, domain adaptation recalibration, and discoverability of core components during model editing.

Conclusion

A salient, extremely sparse set of FFN channels—supernodes—emerges during LLM training as the locus of loss-critical capacity. This feature holds across models and scales and dominates structured pruning behavior. The identification and preservation of these supernodes are necessary and sufficient for robust channel-level sparsification. This delineates a foundational facet of LLM geometry: highly structured, hub-like channel importance that dictates both compression efficacy and offers new axes for mechanistic scrutiny.