We are thrilled to announce that our paper has been accepted at NeurIPS 2026 (Main Track), one of the most prestigious conferences in machine learning and artificial intelligence!

Risk-Aware Action Repetition via Expected Skip Evaluation Hyunwoo Park, Joonhyeok Eom, Baek-Ryun Seong, and Sang-Ki Ko

Repeating an action for several steps, known as a skip, lets a reinforcement learning agent make decisions less often, but raises the question of how much a skip is actually worth. Existing Skip-MDP methods bootstrap from the greedy value at the end of the skip, implicitly assuming the agent can resume near-optimal control right away. This assumption breaks down during training and inflates skip values. The paper proposes Expected Skip Evaluation, which values the end of a skip under the agent’s expected future control rather than idealized greedy control, and works with both discrete and continuous action spaces. Across grid-world, continuous-control, and safety-critical benchmarks, it improves sample efficiency and makes better-calibrated skip decisions.

Huge congratulations to Hyunwoo Park for leading this work, and to Joonhyeok Eom and Baek-Ryun Seong for their contributions! 🎉