Recycling a target model's own predictions provides a training‑free source of proposals for speculative decoding. Existing recycling methods for vision‑language models index predictions solely by the last token, which conflates predictions generated under different images and prefixes; we observe that this reuse fails most often at image‑sensitive decoding states. Longer keys reduce aliasing but match fewer states. To address this, we introduce RECAP (\method), which stores target predictions under one‑, two‑, and three‑token keys: request‑local memories exist only for the current request and are thus implicitly conditioned on its image, while token‑level memory persists across requests to improve coverage. At each lookup, entropy‑confidence weighting fuses the retrieved predictions according to key specificity and concentration. A best‑first search then composes the fused transitions into a candidate tree, supplemented by suffix‑matched continuations; a final target pass verifies the tree under greedy decoding. Without any additional training or a draft model, RECAP achieves mean speedups of $1.66\times$ and $1.56\times$ over autoregressive decoding on seven single‑turn benchmarks for Qwen2.5‑VL‑7B and LLaVA‑1.6‑7B, surpassing the strongest training‑free baseline, Token Recycling ($1.51\times$ and $1.37\times$). The implementation is open‑source: https://github.com/Senmo996/RECAP-official-code.
Review: RECAP's fine‑grained key design and dual‑layer memory effectively mitigate cross‑image prediction conflicts, delivering notable inference acceleration while preserving a zero‑training footprint, highlighting the practical potential of prediction recycling in multimodal speculative decoding.