GainCtrl: Predicting Remaining Gains for Efficient LLM Agent Execution

under review arXiv

Kangwen Zhao, Jianfeng Cai, Ruopei Sun, Mingxiao Feng, Peng Zhang, Wengang Zhou*, and Houqiang Li*.

* Corresponding authors.

Repeated attempts can improve the quality of tool-using LLM agents, but continuing an attempt after meaningful progress has become unlikely wastes model calls and environment interactions. GainCtrl addresses this problem with a lightweight CPU predictor that estimates the expected positive gain remaining in an interaction history. Its calibrated stopping thresholds control the loss in task-level Best-of-N quality while reducing unnecessary execution. GainCtrl-T measures potential gains against the score retained by the current attempt, whereas GainCtrl-P also accounts for the best result retained across previous attempts.

Trajectory-replay experiments span WebShop, ScienceWorld, and ALFWorld with three LLM actors. Across these settings, GainCtrl generally consumes fewer actor tokens than the baselines at matched target quality and achieves higher quality under comparable token budgets. Fixed stopping policies also remain close to full-execution quality while substantially reducing token usage. The predictor operates at millisecond latency on a single CPU thread and relies only on visible interaction histories, allowing it to monitor both black-box and white-box agents without access to model internals.

Download paper from here or view the arXiv page.

BibTeX formatted citation:

@misc{zhao2026gainctrlpredictingremaininggains,
      title={GainCtrl: Predicting Remaining Gains for Efficient LLM Agent Execution},
      author={Kangwen Zhao and Jianfeng Cai and Ruopei Sun and Mingxiao Feng and Peng Zhang and Wengang Zhou and Houqiang Li},
      year={2026},
      eprint={2611.20736},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2611.20736}
}