SpecWM: Turning Verification into Refinement for Streaming World Models
Jiaming Li, Lingzhen Li, Xiaze Yang, Chenyu Zhu, Kaiqi Liu, Xiaotong Zhao, Songlian Li, Toyota Li, Alan Zhao, Haobin Zhong, Yue Ma, Zhiyuan Ma
cs.CV
2026-11-19
Training-free SpecWM scores a 1.3B draft with one 14B latent-velocity probe and writes that estimate back. On WBench, NavScore moves from 75.99 to 85.63 at 1.63× the target.
A streaming video world model rolls the scene forward under the user's camera actions, and it still has to hold image quality, control, and latency. A larger model keeps the picture and the trajectory steadier, and every chunk pays for a full denoising run. A smaller model is fast, and both the image and the controls drift.
Speculative generation lets the small model draft and the large model regenerate the bad chunks. SDVG decodes the draft and asks an external reward what to keep. SDVG-style in this paper is that router on the same draft-target pair. The reward only covers the external scorer's own axes. Decoding and scoring add latency on every chunk, and the score cannot edit the draft. A rejected chunk is a full target resample, so time saved by the draft often returns as regeneration cost.
SpecWM is training-free. Draft and target stay frozen, share the latent, the chunk length, and the control interface, and both condition on the committed history, the current camera action, and the initial image and text. The main pair is LingBot-World 2.0 causal-fast: a 1.3B draft and a 14B target, four denoising stages each.
Sampling follows flow matching: a straight path from noise to a clean latent. At noise level σ the state is a mix of the two, the network predicts a velocity, and subtracting it yields a clean estimate. SpecWM queries the target once, at the last stage. The draft is aligned to that noise, the 14B model returns a clean candidate, and the relative L1 between candidate and draft is the score. The candidate can be committed directly, so the probe is already a one-step edit.
Two thresholds route the chunk. On 20 WBench calibration videos, the 70th and 90th percentiles of the accept-all gaps set τ = 0.2381 and τr = 0.2704. Below τ, keep the draft. In between, refine. At or above τr, rerun all four target stages.
How many steps depends on whether the control is changing. One step lifts perceptual scores and barely moves navigation. Three steps restore camera following and trajectory consistency, and perceptual scores fall relative to one step. Among chunks routed to Refine, a control change gets three target steps, the next chunk two, and later chunks one. The one-step path does no extra denoising. Longer paths re-enter at a higher noise level. After two rejections in a row, the next four chunks are written by the target, with no draft and no probe.
While the target checks the current chunk, the draft writes the next one in parallel. If the current chunk is changed, the next one is proposed again on the new history.
Hy-Nomas uses one 8B network, four steps for the draft and eight for the target. The final probe and the transition schedule stay. The gap is the denoising budget.
The main split is all 158 WBench Navigation cases: 253 frames, 16 fps, about 16 seconds, 16 chunks. NavScore averages navigation accuracy and trajectory consistency. HPSv3 is a human-preference score. Table entries are multiplied by 100. Speedup is generation-loop wall-clock against the 14B target, including rollback and communication, excluding load and final decode. One model uses 4 GPUs; overlap uses 8. A/F/R/T is the share of chunks accepted, refined, rejected, or written in target-only fallback.
| Method | Speedup | HPSv3 | Imaging | NavScore | Route A/F/R/T |
| Target-only 14B | 1.00× | 74.57 | 66.82 | 85.49 | — |
| Draft-only 1.3B | 4.35× | 72.49 | 64.67 | 75.99 | — |
| SDVG-style | 1.42× | 73.98 | 66.04 | 75.92 | 68/0/32/0 |
| Refine-all, 1 step | 1.78× | 77.80 | 66.35 | 75.92 | 0/100/0/0 |
| Refine-all, 3 steps | 1.10× | 74.26 | 66.38 | 85.17 | 0/100/0/0 |
| SpecWM 3→2→1 | 1.63× | 77.66 | 67.19 | 85.63 | 68/13/10/9 |
SpecWM's NavScore is 85.63, against 85.49 for the target and 9.64 above the draft's 75.99. HPSv3 rises from 74.57 to 77.66, Imaging from 66.82 to 67.19. Acceptance matches SDVG-style at 68%, but SDVG-style regenerates 32% of chunks in full. SpecWM refines 13% and sends 19% through a full target run, fallback included. Speedup moves from 1.42× to 1.63×.
Refine-all at one step scores HPSv3 77.80 at 1.78×, with NavScore stuck at 75.92. Three steps reach NavScore 85.17, drop HPSv3 to 74.26, and slow to 1.10×. Removing the two-step stage drops NavScore from 85.63 to 81.49 and Spatial from 80.83 to 76.90, while speedup only rises to 1.67×.
Without Refine, speedup is 1.21× and NavScore is 82.18. Without Reject, speedup is 2.07× and NavScore is 80.46. Random routing at the same shares runs at 0.95×, slower than the target, with NavScore 75.98. Dropping fallback leaves 1.45×. Turning overlap off leaves 1.29×. In a five-chunk pilot, one stage took 1.5 s and all four took 4.4 s. Final-stage gaps varied by 1.55× from smallest to largest, against 1.22× and 1.30× on the first two stages.
LongSpecBench has 80 evaluation rollouts of 509 frames, about 32 seconds, with ten to twelve turns and revisits. SpecWM runs at 1.67×. Revisit consistency is 76.56, against 76.77 for the target and 66.40 for the draft. Direction, response, and camera following are 91.89, 83.61, and 22.03, against the target's 90.92, 83.41, and 18.87. WMQ is 72.41, against 71.62 and 67.13. Once speed is folded into QES, SpecWM is at 70.32 and the draft, at 4.26×, still reaches 69.69.
On Hy-Nomas, human comparisons against the target are mostly ties, with wins and losses close. Against the draft and against SDVG-style, wins exceed losses on all four axes. Prompt following (n=120, win/tie/loss) is 55.8/32.5/11.7 against the draft, 28.3/46.7/25.0 against the target, and 43.3/44.2/12.5 against SDVG-style. Relative preference, ties counted as half-wins and scaled to a maximum of 100, lands at 100, 94.23, 52.44, and 42.33. The two high scores are the target and SpecWM. Generation-loop speedup is 1.33×, about 2.09× for the draft and 1.29× for SDVG-style. Both roles share weights and run in sequence.
Interactive use cares about the latency of the next chunk. SpecWM spends 14B steps where the draft and the target already disagree: trust the 1.3B output on a small gap, spend one to three target updates in the middle, and pay for a full resample only on a large gap. Shared latents and controls are enough, with no extra student to train. A single network can play both roles by changing the step budget.
Perceptual preference and speed still favor a one-step correction on every chunk. Navigation only gets back to the 14B model. The extra steps sit on camera changes.
Draft and target must share the latent, the chunk length, and the control interface. Cross-model latent alignment is left open, so arbitrary video generators do not plug in.
Thresholds are frozen percentiles from 20 videos, and the paper does not sweep them. When the shares are right but the chunks are wrong, random routing drops navigation back toward the draft.
The 1.63× and 1.67× figures are wall-clock. LingBot uses 4 GPUs for one model and 8 when the two overlap. That matches an interaction budget, where extra GPUs buy lower latency. It is not a same-GPU accounting. Hy-Nomas at 1.33×, shared weights and sequential execution, is the cleaner compute comparison.
LongSpecBench is a new 80-case set. Camera-following scores sit around 16 to 22 on a 0 to 100 scale from MegaSaM poses, so a lead of about 3 points is a ranking on a narrow scale. NavScore 85.63 versus 85.49 is a tie.
The paper discloses that GPT was used on the writing, the experimental analysis, and the research code.