WAMJET uses coding agents to auto-optimize World Action Model inference, reaching up to 9.95x lossless speedup across six models

WAMJET is an agentic harness that lets coding agents automatically accelerate World Action Model inference with reusable optimization guidance and measurement and validation tools. Prior acceleration techniques require substantial engineering to select and compose per model and hardware platform; WAMJET delegates that selection and composition to agents. The abstract states the harness equips coding agents with reusable optimization guidance and measurement and validation tools, and follows a bottleneck-driven workflow.

WAMJET follows a bottleneck-driven workflow in which the agent profiles inference, modifies targeted code, validates effects, and iteratively refines the acceleration stack as bottlenecks shift, while preserving action quality. The workflow frames acceleration-stack construction as iterative refinement as bottlenecks move, rather than applying a fixed optimization combination once. The abstract explicitly describes profiling, targeted modification, effect validation, and iterative refinement, and emphasizes preserving action quality.

Experiments span six World Action Models, three coding agents, and two GPU architectures, achieving up to 9.95x lossless speedup over upstream implementations. The result is reported across multiple models, agents, and hardware architectures, indicating acceleration stacks can be produced across these settings. The abstract gives the experimental scope of six WAMs, three coding agents, and two GPU architectures, and the figure of up to 9.95x lossless speedup.

Approximation and hardware-aware optimization yield additional latency reductions with comparable success rates. Beyond lossless speedup, approximation and hardware-aware optimization further compress latency while maintaining comparable success rates. The abstract states approximation and hardware-aware optimization yield additional latency reductions, with comparable success rates.

This work targets World Action Model inference acceleration, applicable where pretrained video foundation models are used for robot manipulation and inference cost is a deployment bottleneck. Its bottleneck-driven workflow and reusable optimization guidance are aimed at coding agents, and experiments cover six WAMs, three coding agents, and two GPU architectures, so results apply to acceleration-stack generation within these models and hardware. Approximation and hardware-aware optimization can further reduce latency while maintaining comparable success rates, suitable for deployments that accept approximation and have known hardware conditions.

As an abstract, this work does not provide per-model and per-hardware speedup breakdowns, full evaluation details of action quality and success rates, or how much latency reduction approximation and hardware-aware optimization each contribute. Readers should still watch the stability of the bottleneck-driven workflow across different WAMs and GPU architectures, the coverage of the reusable optimization guidance, and the success-rate behavior of approximation optimization on broader tasks and hardware.