Tag
A paper introducing T/E/S, a structural representation separating text, local events, and activity spans to model LLM reasoning dynamics, showing improved next-outcome prediction across reasoning tasks and deriving efficient bounds for budgeted continuation analysis that cut inference CPU time by 77.7%.
This paper systematically investigates reward function design for reinforcement learning to improve the quality of LLM-generated BPMN process models, finding that equal reward weighting outperforms targeted weighting and that design choices interact with model architecture in non-trivial ways.