Driving · 2026-10-06
WAMJET: A Harness for World Action Model Acceleration
Le Chen, Lixin Liu, Jan Schneider, Zeju Qiu, Simon Guist, Bernhard Sch\"olkopf, Dieter B\"uchler
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
WAMJET is an agentic harness for accelerating World Action Model (WAM) inference. It guides coding agents through a bottleneck-driven workflow: profile inference latency, modify targeted code for lossless or approximate accelerations, and validate numerical/state correctness plus paired task performance. Experiments across six WAMs and two GPU architectures show up to 9.95× lossless speedup over upstream implementations, with comparable success rates under approximate acceleration.
Why it matters
WAM inference can be expensive due to large video backbones and video-action co-prediction, and effective acceleration requires repeated profiling, implementation, and careful correctness checks to preserve action quality. WAMJET aims to reduce this deployment engineering burden by providing reusable optimization guidance and measurement/validation tools to automatically build model-and-hardware-specific acceleration stacks.
Method
- Bottleneck-driven workflow: establish a baseline, profile inference (without relying on slow profilers), implement lossless acceleration first, then approximate techniques when permitted, and re-profile after each change.
- Validation tiers and checks: tier A lossless with output/state checks; tier B validated approximation with paired task evaluation; tier C changed workload/unvalidated approximation rejected or bounded by user settings. Also checks output structure/numerical drift and statefulness via A→B→A sequences.
- Agent execution loop: measure GPU-synchronized warmed latency, record experiments and accept/revert decisions, and continue iterative search within a specified budget while storing validated notes in external memory.
Limitation
The paper notes that the numerical drift checker could possibly make the acceleration too conservative, so in practice it is usually turned off.
Abstract (from arXiv)
World Action Models (WAMs) leverage pretrained video foundation models for robot manipulation, but their large backbones and video-action co-prediction are expensive. Although existing acceleration techniques offer many ways to reduce this cost, selecting and composing them requires substantial engineering for each model and hardware platform. To tackle this bottleneck, we present WAMJET, an agentic harness that accelerates WAM inference by equipping coding agents with reusable optimization guidance and measurement and validation tools. WAMJET follows a bottleneck-driven workflow where the agent profiles inference, modifies targeted code, validates effects, and iteratively refines the acceleration stack as bottlenecks shift, while preserving action quality. Experiments span six WAMs, three coding agents, and two GPU architectures. WAMJET achieves up to 9.95x lossless speedup over upstream implementations. Approximation and hardware-aware optimization yield additional latency reductions, with comparable success rates. The results show that WAMJET can produce effective acceleration stacks for WAM deployment.
Related papers
- Staircase Policy: Streaming Inference for World-Action Models with Large Action Chunks
- MARS: Systematic Multi-Level Acceleration for World Action Models
- RealtimeWAM: How Fast Can I Run My World Action Model?
- 1Step-WAM: Your World Action Model but Faster using 1-Step Action Generation
- WAM-Cache: Staleness-Bounded KV Reuse for Efficient World Action Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-01 · Indexed by Pub-AI: 2026-10-06 UTC
WAMJET: A Harness for World Action Model Acceleration
Preprint posted: 2026-10-01 · Original record ↗
- arxiv
- 2610.03797
Explicit source versions (1)
- v3 · 2026-10-07
- Explicit source version detected · v3
Source date: 2026-10-07 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:id explicit version (official metadata)
- Preprint available
Source date: 2026-10-01 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
World Action Models (WAMs) leverage pretrained video foundation models for robot manipulation, but their large backbones and video-action co-prediction are expensive. Although existing acceleration techniques offer many ways to reduce this cost, selecting and composing them requires substantial engineering for each model and hardware platform. To tackle this bottleneck, we present WAMJET, an agentic harness that accelerates WAM inference by equipping coding agents with reusable optimization guidance and measurement and validation tools. WAMJET follows a bottleneck-driven workflow where the agent profiles inference, modifies targeted code, validates effects, and iteratively refines the acceleration stack as bottlenecks shift, while preserving action quality. Experiments span six WAMs, three coding agents, and two GPU architectures. WAMJET achieves up to 9.95x lossless speedup over upstream implementations. Approximation and hardware-aware optimization yield additional latency reductions, with comparable success rates. The results show that WAMJET can produce effective acceleration stacks for WAM deployment.