Large language models are increasingly used to help people arrange appointments, plan journeys, and allocate limited resources. Such tasks are not open-ended generation; they require satisfying explicit constraints on time, cost, capacity, and order, or correctly concluding that no valid plan exists.
To test this capability in Bangla, the new benchmark BANGLAPLAN provides a dedicated evaluation suite. It focuses on constraint-based planning and decision-making in the Bangla language, a domain that has received less attention than English-centric planning benchmarks.
The benchmark is designed to probe whether models can reason about multiple interacting constraints and produce feasible plans, rather than merely generating plausible-sounding responses. By framing the task in Bangla, it also checks whether these planning abilities transfer across languages.