The news. On July 13, 2026, a work-in-progress paper introduced TIGER, which asks why speculative decoding helps text-only models more than it helps vision-language models. Its answer: the draft model diverges from the target on vision-critical content, so those guesses get rejected. TIGER proposes text-conditioned sparse visual routing — feeding the drafter only the image patches relevant to its current text — and trains the drafter to maximize how much of its guess survives verification rather than to imitate the target. Read the paper →

Picture a junior docent walking a tour group through one enormous painting, narrating as they go, with a senior guide a step behind who stops them the instant they say something wrong. The junior is fast, which is the point — they can rattle off a few lines while the senior would still be clearing their throat. The catch is the painting. If the junior tries to keep the whole canvas in their head at once, their attention smears across it, and the moment the sentence turns to a specific detail — the colour of a collar, the number of figures in a doorway — they guess, the senior catches it, and the run stops there.

That is a vision-language model under speculative decoding, and swapping the docent for a small draft model changes almost nothing. The trick is the standard one: the draft model guesses several tokens ahead, and the big target model checks all of those guesses in one forward pass. On text-only prompts it pays off handsomely, because a small model imitating a big one is right often enough that long runs of its guesses get accepted. The paper's starting observation is that this pays off less on images: the draft model diverges from the target precisely on the content that depends on the picture, and those are the guesses the target rejects.

Why does the drafter diverge there? A picture is not a few tokens, it is often hundreds of visual tokens — the image cut into patches — and the draft model has to produce its next words while attending across all of them. The intuition the paper works from is that attending to everything leaves the drafter unfocused exactly where it needs to be sharp, while a fixed, pre-compressed summary of the image throws detail away before the drafter knows which detail this sentence will need. Either way the drafter is looking in the wrong place: at all of the image, or at a frozen sketch of it, never at the part the current sentence is actually about.

TIGER's move is to make the drafter's view of the image depend on what it is currently saying. As the draft model writes, it uses its own current text state to select the handful of visual tokens relevant to that moment and route only those into the guess — the same way the docent looks only at the section they are narrating right now. When the sentence is about the collar, the collar patches are what the drafter sees; when it moves to the doorway, the view moves with it. This is a routing decision over the image, made fresh at each step, and it is meant to keep the drafter's guesses from smearing across a canvas it cannot hold all at once.

The second half is how the drafter is trained, and it is the part that stops this from being ordinary imitation. A drafter trained the usual way learns to sound like the target — distillation that copies the target's next-token distribution, with a term anchoring it close. But sounding like the target is not the same as surviving it. So TIGER starts from that anchored distillation and then switches the objective to a verifier-derived reward: the drafter is rewarded for the accepted prefix length its guesses actually earn under the target's check. It optimizes the thing the speed-up is made of — how many guessed tokens in a row survive — instead of a proxy for it.

Hold the arithmetic still and the payoff is easy to see. A verification pass costs about the same whether it confirms one token or eight, so the tokens-per-pass you get is just the accepted prefix length. Say the drafter proposes 8 tokens, and on a vision-critical sentence the all-patches drafter has its 4th guess rejected — an accepted run of 3, so you netted roughly 3 tokens for one pass (illustrative: the paper is a work in progress and does not publish this breakdown). Now route the drafter to the two patches that sentence is about, and suppose the first mismatch slips to the 7th guess — an accepted run of 6, roughly 6 tokens for the same pass. Nothing about the verification got cheaper. The drafter simply stopped guessing blind on the parts that depend on the picture, so more of each run survives — and on a draft-verify loop, a longer accepted run is the speed-up.

The reason this reads as an attention result rather than a decoding one is that it moves the decision to the drafter's eyes. A standard drafter treats the image as fixed context and hopes its imitation carries the vision-critical words; TIGER treats which patches to look at as a choice the drafter remakes every step, and treats surviving verification as the thing to optimize for. Framing the drafter's problem as "look where the sentence is going, and earn a long accepted run" is TIGER's bet for making speculative decoding pay off on images the way it already does on text — a bet the work-in-progress paper proposes rather than yet reports numbers on.

Goes deeper in: LLM Serving → Speculative Decoding → The Draft-Verify Idea

Related explainers

• VIA-SD — Tiered speculative verification — a different lever on the same loop: make verification itself cheaper rather than making the guesses survive longer
• DominoTree — Conditional draft-tree scoring — spends a single request's draft budget better by scoring a tree of continuations
• ViQ — Text-aligned quantized visual tokens — the other half of the multimodal cost story: how the image becomes tokens in the first place