We investigate whether small LLM agents can operate under explicit wall‑clock time budgets while both respecting the allocated runtime and using any spare time productively. Experiments evaluate Qwen3.6‑27B on five competitions from MLE‑Bench Lite and Qwen3‑4B on Zork I (Jericho), two agentic benchmarks where extra computation can meaningfully boost performance.
In the simplest setting, the budget is only stated in the prompt. Agents fail to translate this into controlled time usage. The failures stem from gaps in time awareness: the harness provides no timing feedback, the models cannot reliably anticipate action durations, and they lack a learned mapping from available time to an appropriate strategy.
We explore two complementary interventions: (1) harness‑based mechanisms that expose timing information and enforce deadlines, and (2) reinforcement learning with budget‑aware rewards. Injecting timing information through the harness substantially improves budget adherence for Qwen3.6‑27B without measurable loss in performance, while enforcement hooks tighten adherence further. RL with GRPO (Generalized Reward‑Based Policy Optimization) achieves near‑perfect budget adherence on Zork I and generalizes to held‑out budgets not seen during training, but does not improve task performance on MLE‑Bench compared to the untrained harness.
Even when agents respect the budget, they still fail to use additional time to improve performance. RL‑trained policies learn when to stop but often fill extra time with repeated actions, and GRPO training on multiple budgets tends to collapse toward the strategy learned for the shortest budget.
Our results reveal a gap between time adherence and productive time allocation, which remains a central challenge for budget‑conditioned agents.
Review