Step through how a small model's cache is reused inside a bigger one (same family, shared tokenizer).
The ρ dial Illustrative
ρ = fraction of prompt positions the big model recomputes. ρ=0 is plain transfer; ρ=1 is a full prefill. Which positions light up here is a schematic, not data.
ρ0.30
Motivating gap Reported
Qwen3-0.6B cache mapped into Qwen3-14B, GSM8K accuracy (%).
Where does the 14B look? Reported
Attention share on one GSM8K prompt (Qwen3-14B).
Pick the positions Illustrative
Schematic scores for a made-up GSM8K-style prompt. Choose a selector and a budget; orange outlines = recomputed. Hover for details.
ρ0.30
Hover a token.
Rank disagreement, geometrically Illustrative
Calibration predictions vary along a well-described subspace (the paper keeps the top 128 directions). A token's score is how much mapped KV lies outside it, summed over layers and K/V.
offset from subspacealong
Accuracy and retention Reported
Qwen3-0.6B → 14B unless noted. Pick a view.
Source-size spread collapses Reported
Spread across source sizes on GSM8K (percentage points).
Latency Reported
One A100, PyTorch eager, plain batching baseline. The small model's own prefill is excluded.
Why 30% recompute is not 30% cost
Two fixed costs the dial does not shrink.
What the evidence covers
Hover a row. Status reflects what the episode says was and wasn't run.
Hover a row.
Template vs. content Reported
Share of recomputed positions in the shared template prefix, full GSM8K test set, vs. a template-heavy selector.