Surfing the Activation Wave: Analyzing the Information and Geometry of LLM Hidden States for Speculative Decoding
The paper argues that speculative decoding can pick better target-model layers by measuring hidden-state informativeness and redundancy, not by fixed layer positions.
The authors use matrix-entropy analysis and centered kernel alignment to map how decoder layers differ across depth. They add token-entropy filtering to remove low-uncertainty tokens from the layer-selection statistics. In EAGLE3-style speculative decoding, their selected layer sets beat fixed baselines, including the default `[2, N//2, N-3]` rule, while explaining why that default can still hold up. HF Daily Papers' note
The authors use matrix-entropy analysis and centered kernel alignment to map how decoder layers differ across depth. They add token-entropy filtering to remove low-uncertainty tokens from the layer-selection statistics. In EAGLE3-style speculative decoding, their selected layer sets beat fixed baselines, including the default `[2, N//2, N-3]` rule, while explaining why that default can still hold up. HF Daily Papers' note
score 4