BudgetBench
Adaptive reasoning-budget control for cost-efficient LLM inference.
🎯The Problem
Models expose reasoning budget knobs, but we don't know which tasks benefit or how to allocate budget optimally.
💡The Solution
BudgetBench protocol for standardized evaluation. Adaptive Budget Controller that escalates budget only when needed. Reproducible pipeline using open weights and datasets.
✨Key Highlights
- Accuracy–cost Pareto curves
- Adaptive budget escalation
- $0 reproducible pipeline