Calibrated Signals @ FinMMEval 2026 Task 2: Inference-Only Strategies for Cross-Lingual Financial Question Answering
This work evaluates eight prompting strategies (P0–P7) over a single Azure AI Foundry agent, four inference-time techniques (Structured IR, Self-Refine, LLM-as-Judge, Cross-lingual Consistency), and a reward-guided Best-of-N pipeline that generates DPO preference pairs as a by-product.
Gaurav Jain, H. Hitesh, SP Sundriyal et al.
· 1 citation