让智能体学会推理自身计算开销：同等训练预算下 75% 游戏表现更优、算力省三倍

View PDF HTML (experimental)

Abstract:While reinforcement learning agents can achieve superhuman performance in many complex tasks, they typically do not become more computationally efficient as they improve. In contrast, humans gradually require less cognitive effort as they become more proficient at a task. If agents could reason about their compute as they learn, could they similarly reduce their computation footprint? If they could, we could have more energy efficient agents or free up compute cycles for other processes like planning. In this paper, we experiment with showing agents the cost of their computation and giving them the ability to control when they use compute. We conduct our experiments on the Arcade Learning Environment, and our results demonstrate that with the same training compute budget, agents that reason about their compute perform better on 75% of games. Furthermore, these agents use three times less compute on average. We analyze individual games and show where agents gain these efficiencies. An interactive version of this paper is available at this https URL.

Comments:	Accepted at the Reinforcement Learning Conference (RLC) 2026. Added interactive version: this https URL. Author name change
Subjects:	Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Cite as:	arXiv:2510.22833 [cs.AI]
	(or arXiv:2510.22833v3 [cs.AI] for this version)
	https://doi.org/10.48550/arXiv.2510.22833 arXiv-issued DOI via DataCite

Submission history

From: Adrian Orenstein [view email]
[v1] Sun, 26 Oct 2025 21:01:30 UTC (3,498 KB)
[v2] Mon, 12 Oct 2026 17:14:41 UTC (3,582 KB)
[v3] Wed, 14 Oct 2026 10:19:05 UTC (3,582 KB)