An Adoption-Aware Crop Recommender from Farmer World Feedback

We build an adoption-aware crop recom- mender from offline farmer survey panel data, treating recorded adoption decisions and re- alized yields as world feedback. The rec- ommender targets Expected Realized Benefit (predicted yield gain weighted by predicted adoption probability) and is fit on 8,744 cul- tivated plots from the LSMS-ISA Ethiopia ESS Wave 4 panel. On 800 enumeration- area-disjoint test plots, it beats an accuracy- only baseline by 50 kg/ha (bootstrap 95% CI [24, 74]), with Rosenbaum sensitivity Γ⋆ ≥ 5 and sign-exchangeable permutation p < 10−3 .The Ethiopia-trained adoption head repro- duces the treatment-effect ordering of the Du- flo et al. (2011) Kenya SAFI fertilizer RCT across all four arms. Farmer feedback dif- fers from robot or LLM feedback in that the logging policy is the agent, so its propen- sities cannot be recovered from the data; we describe the architectural and evaluation choices that let direct-method offline policy learning work in this regime.