PaperScope
LIVE · 2026-09-09 05:40 UTC

Elastic Horizon: Discovering the Effective Interaction Frontier in Agentic Reinforcement Learning

Gangyi Zhang, Junjie Meng, Letian Zhang, Wei Wu, Yang Zheng, Dong Wang, Yang Liu, Guanjun Jiang, Chongming Gao

Latestcs.CLcs.LGcs.AIcs.CV
arXiv ID
2609.07247 v1
Category
Submitted
2026-09-07

Abstract

Scaling the interaction horizon-the maximum number of environment interactions per episode-improves LLM agents on long-horizon tasks, and curriculum-based methods that progressively expand the horizon outperform fixed-horizon alternatives. However, existing schedules are open-loop: they monotonically increase the horizon until a manually specified maximum, with no mechanism to detect when further expansion stops helping. We propose the effective interaction frontier hypothesis: a dynamic boundary beyond which additional interactions yield diminishing returns while cost grows linearly. We then introduce Elastic Horizon, a closed-loop controller that tracks this boundary via the 90th percentile of successful trajectory lengths. On AppWorld and BFCL, fixed-horizon sweeps reveal clear saturation plateaus; Elastic Horizon stabilizes the horizon inside the saturation band from both under- and over-capacity initializations, attains the best success rates across 7B and 14B backbones, and saves up to 25% of per-step trajectory tokens. Our work shifts the paradigm from how to scale interaction horizons to when to stop scaling.

Comment: 17 pages, 6 figures, 12 tables. Accepted to EMNLP 2026

arXiv abs page · PDF · same-day batch