Driving · 2026-10-15
Complementary Future Token Pruning for World Action Models
Baekseung Kim, Junhoo Lee, Sohyeon Hong, Suhyun Jeon, Minkyu Kim, Seungyeon Kim, Donghoon Han, Nojun Kwak
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
CF-Prune is a training-free method to prune future-frame tokens in pretrained world action models by selecting task-relevant future tokens from the initial denoising forward pass using instruction and action attention. It further disables pruning during “precision phases” detected via attention entropy contrast. Experiments on RoboLab, RoboCasa, and real-world manipulation report inference speedups with maintained or improved success rates.
Why it matters
WAMs jointly generate actions and multiple future frames, and dense future-frame tokens dominate compute and latency. CF-Prune reduces future-token inference cost without additional training by identifying redundant tokens early (when representations are noisy) and by selectively relaxing pruning to avoid performance degradation during precision interaction phases.
Method
- Training-free future-frame token pruning in pretrained WAMs: compute instruction relevance (observation-to-instruction attention) and action relevance (action-to-visual/future attention) in the initial denoising forward pass, then select complementary token sets under pruning ratio α.
- Instruction-action complementary selection: retain (1-α)/2 fraction by action relevance and (1-α)/2 fraction by instruction relevance, with wrist tokens selected per future frame and third-person tokens selected shared across future frames.
- Entropy-based phase identification: define an entropy contrast score from attention entropy differences across view (third vs wrist) and temporal position (early vs late action tokens), and disable pruning during precision phases.
Abstract (from arXiv)
World action models (WAMs) jointly predict future visual states and robot actions, but processing dense future-frame tokens incurs substantial inference latency. We propose CF-Prune, a training-free framework for pruning future-frame tokens in pretrained WAMs. Our key insight is that future visual predictions support action generation and therefore need not preserve every visual detail required for high-fidelity video synthesis. CF-Prune combines instruction and action attention to identify task-relevant tokens during the initial denoising forward pass, even when future-frame and action representations remain highly noisy. An entropy-based phase identification mechanism further retains dense computation during precision phases to mitigate pruning-induced performance degradation. Experiments on RoboLab, RoboCasa, and real-world manipulation tasks demonstrate consistent inference speedups. At a 35% pruning ratio, CF-Prune achieves average speedups of 1.19-1.27x while improving task success rates. At a 50% pruning ratio, it achieves 1.29-1.33x average speedups on both simulation benchmarks, with overall success rates within 1.0 percentage point of the baseline.
Related papers
- Sparse-WAM: Accelerating World Action Models via Action-Guided Sparse Imagination
- Self-Distilling Future Video Representations for World Action Models
- CtrlWAM: Controllable World Action Models with Aligned Intent and Foresight
- De-Diffusing WAMs: Action-Sufficient Direct Prediction for Robot Control
- {\Delta}WAM: Distilling Action Tangent Fields into World Action Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-14 · Indexed by Pub-AI: 2026-10-15 UTC
Complementary Future Token Pruning for World Action Models
Preprint posted: 2026-10-14 · Original record ↗
- arxiv
- 2610.19957
Explicit source versions (1)
- v1 · 2026-10-14
- Preprint available
Source date: 2026-10-14 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
World action models (WAMs) jointly predict future visual states and robot actions, but processing dense future-frame tokens incurs substantial inference latency. We propose CF-Prune, a training-free framework for pruning future-frame tokens in pretrained WAMs. Our key insight is that future visual predictions support action generation and therefore need not preserve every visual detail required for high-fidelity video synthesis. CF-Prune combines instruction and action attention to identify task-relevant tokens during the initial denoising forward pass, even when future-frame and action representations remain highly noisy. An entropy-based phase identification mechanism further retains dense computation during precision phases to mitigate pruning-induced performance degradation. Experiments on RoboLab, RoboCasa, and real-world manipulation tasks demonstrate consistent inference speedups. At a 35% pruning ratio, CF-Prune achieves average speedups of 1.19-1.27x while improving task success rates. At a 50% pruning ratio, it achieves 1.29-1.33x average speedups on both simulation benchmarks, with overall success rates within 1.0 percentage point of the baseline.