Super Weights in LLMs: Why High-Salience Parameters Fail as Fine-Tuning Targets
Amazon
This COLM 2026 paper investigates 'Super Weights' — the tiny subset of parameters (100–36K) whose removal collapses LLM performance. Counterintuitively, targeting only these high-importance parameters for fine-tuning causes accuracy to collapse to random-guessing levels on OLMo models. Training an equal number of randomly selected parameters in the same layers succeeds, proving the failure is specific to Super Weight targeting. Vanilla LoRA — spreading updates across full weight matrices via low-rank decomposition — succeeds with only 0.16% of parameters. Findings confirmed across 10 random seeds.
Why it matters
It directly refutes a popular hypothesis that high inference-time parameter salience predicts fine-tuning leverage, a premise underlying several published efficient-training methods. Practitioners cannot safely use Super Weight coordinates to guide parameter-efficient fine-tuning without risking catastrophic performance degradation. Strong positive evidence for structured low-rank approaches over targeted sparse ones.
Importance: 3/5
COLM 2026 accepted; refutes popular assumption underlying several efficient fine-tuning methods