🎰 MAB-DQA: Multi-Armed Bandit Document Question Answering

ACL 2026 — Thompson Sampling for aspect-aware document page retrieval.

Upload a PDF, ask a question, and watch the Multi-Armed Bandit explore & exploit to find the most relevant pages.

How It Works

1. Query Decomposition — Qwen2.5-VL breaks your question into aspect subqueries
2. ColPali Embedding — Pages and subqueries are encoded into a shared vision-language space
3. Late Interaction — MaxSim scoring between query tokens and page patch tokens
4. Page Similarity Graph — Cosine similarity between page embeddings builds a neighbor graph
5. Thompson Sampling MAB — Each subquery is an "arm"; the bandit explores high-value aspects and exploits promising pages over b=2 rounds
6. Final Ranking — Pages ranked by composite score: ColPali + VLM relevance + exploration bonus

📄 Paper · 💻 Code