Skip to main content
Back to timeline
arXivSource publication:

Streaming world models stop discarding rejected drafts: SpecWM turns verification into refinement for a 1.63x speedup on WBench

Synopsis

Streaming world models must hold visual quality, control following and responsiveness at once, yet large models are costly and small ones are fast but weaker on quality and control. Earlier speculative methods scored decoded frames with an external reward model and used the score only to accept a draft or regenerate it with the target, leaving rejected drafts no correction signal. SpecWM instead probes the target's final-stage latent velocity field once, reuses that same prediction as a correction, and routes chunks to Accept, Refine or Reject with two thresholds; on the 158-case WBench Navigation split it reaches a 1.63x generation-loop speedup over Target-only and raises NavScore from the draft's 75.99 to 85.63, with HPSv3 rising from 74.57 to 77.66.

Source-provided article image: SpecWM: Turning Verification into Refinement for Streaming World Models
arXiv

Interpretation

Speculative generation for streaming world models gains a middle path: a draft chunk judged off-target no longer faces only acceptance or full regeneration, because the target can apply a bounded correction before the chunk is committed. Earlier reward-based speculative methods scored decoded frames and used the score only to accept a chunk or send it through the target's full sampling process, so a rejected draft received no correction signal. On the full 158-case WBench Navigation split, SpecWM and SDVG-style both accept 68% of chunks, but SpecWM refines 13% and leaves 19% for full target generation including fallback, whereas SDVG-style needs full regeneration for 32%, lifting generation-loop speedup from 1.42x to 1.63x.

Correction depth is allocated by where control changes rather than fixed: refinement runs deeper at a turn and shallower during continued motion. Uniform-refinement experiments show complementary effects, with one-step refinement raising HPSv3 from the draft's 72.49 to 77.80 and three-step refinement raising NavScore from 75.92 to 85.17, while three steps lowers HPSv3 to 74.26 and speedup from 1.78 to 1.10. With three steps at a transition, two in the next chunk and one thereafter, SpecWM reaches NavScore 85.63 and HPSv3 77.66 at 1.63x speedup, against 76.39 NavScore for fixed one-step refinement and 81.49 when the intermediate two-step correction is removed.

Verification and refinement share a single target probe, and draft generation of the next chunk overlaps target verification of the current one. Decoded-frame scoring requires decoding and scoring each chunk, adding runtime cost, and the score itself supplies no correction to the draft. One final-stage probe takes 1.5 s under separate optimized timings versus 4.4 s for four-stage verification; removing overlapped execution drops speedup from 1.63x to 1.29x, and removing burst fallback drops it to 1.45x.

Perspective

The demonstrated scope is the pairing of a 1.3B draft with a 14B target from LingBot-World 2.0 on the full 158-case WBench Navigation split, plus a 280-case human evaluation of Hy-Nomas with shared weights and four draft steps against eight target steps; thresholds are set from 20 calibration cases at the 70th and 90th percentiles and then fixed. The authors propose extending draft-target pairing to cross-model latent alignment.

The authors report that one-step refinement improves perceptual quality with limited navigation-control gain, while three-step refinement strengthens camera following and trajectory consistency but lowers perceptual preference and aesthetics; thresholds depend on percentile settings from 20 calibration cases, and routing effectiveness depends on which chunks receive each action rather than on route shares alone.