| Recursive Language Models (RLMs) enable long-context reasoning by iteratively generating and executing code over externalized inputs. While this approach improves reasoning capability beyond fixed context windows, it introduces practical challenges in deployment, including unpredictable latency, unbounded cost, sequential execution bottlenecks, and unsafe code execution. We present the Async Budgeted Sandbox RLM (ABS-RLM), a runtime infrastructure designed to improve the execution characteristics of recursive language model systems. Rather than proposing a new reasoning algorithm, our approach focuses on the execution layer, providing (i) asynchronous sub-call orchestration for parallel processing of independent tasks, (ii) strict budget enforcement to bound token usage and monetary cost, and (iii) a sandboxed execution environment to safely run model-generated code. We evaluate ABS-RLM on long-context reasoning tasks, combining controlled synthetic workloads with a subset of real-world examples. Results show that ABS-RLM reduces tail latency by up to $3.1x and cost variance by up to 69% compared to sequential recursive execution, while maintaining comparable task performance under relaxed budget settings. The goal is not to outperform alternative methods, but to improve the reliability, predictability, and safety of recursive reasoning systems. These results show that runtime-level design is essential in making recursive language model approaches practical for real-world deployment. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.