AttnSR preprint reports improved agent success with fewer training interactions

Researchers report that AttnSR, a method for assigning training rewards to individual agent steps, raised final held-out success on ALFWorld from 91.2% to 97.1% in a controlled Qwen2.5-7B-Instruct study. The preprint also reports improved ScienceWorld performance at every checkpoint while reducing environment interactions by 29.9%. The method traces model attention backward during training replay to estimate which earlier steps contributed to the final response. It requires neither a separately trained reward model nor additional trial runs. An ablation experiment attributes the benefit to how intermediate rewards are assigned, rather than simply accumulating rewards over time.