Research map / Vision-language and multimodal / Multimodal LLMs

Resolving Evidence Sparsity: Agentic Context Engineering for Long-Document Understanding

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SLEUTH, a training-free multi-agent framework that builds evidence-dense multimodal contexts from retrieved pages for long-document QA, using clue discovery, page screening, difficulty assessment, and core decision agents on top of ColPali retrieval.

Problem

VLMs degrade on long documents where key evidence is sparse and scattered across pages and modalities, and retrieved pages still contain substantial redundancy.

Approach

A coarse-to-fine, page-wise multi-agent pipeline (ColPali top-K retrieval, then clue extraction and visual page screening, then difficulty-aware choice between instruct and thinking VLMs) that keeps effective context length fixed and distills evidence before final answering, without any training.

Key results

• SLEUTH with Qwen3-VL-8B reaches 52.77% average on MMLongBench-Doc vs 46.76% for Base and 48.75% for MoLoRAG.
• On LongDocURL, SLEUTH achieves 59.96% average vs 55.18% for Base; PaperTab 43.09% and FetaTab 70.46%.
• With Gemini-2.5-Flash, SLEUTH Top-5 reaches 51.86% on MMLongBench-Doc vs 46.37% Base.

Surprising findings

• Purely visual page inputs outperform a multimodal (visual + OCR text) retrieval input on both MMLongBench-Doc and LongDocURL.
• Increasing retrieval Top-K improves accuracy monotonically while the effective context length stays fixed.

Limitations

• Relies on the initial retriever; missed critical pages cannot be recovered downstream.
• Purely prompt-driven design without training limits self-evolution.
• Binary difficulty estimation is coarse; experiments only on English administrative and academic documents.

Experimental setup

Datasets: MMLongBench-Doc LongDocURL PaperTab FetaTab

Benchmarks: MMLongBench-Doc LongDocURL PaperTab FetaTab

Metrics: accuracy

Baselines: Direct Base M3DocRAG MoLoRAG MDocAgent GPT-5 Gemini 2.5 Pro VRAG-RL MACT

Builds on: ColPali MDocAgent M3DocRAG MoLoRAG

Closest papers by meaning

• Visual Document Understanding and Reasoning: A Multi-Agent Collaboration Framework with Agent-Wise Adaptive Test-Time Scaling (CVPR 2026 · similarity 0.90)
• DocSeeker: Structured Visual Reasoning with Evidence Grounding for Long Document Understanding (CVPR 2026 · similarity 0.89)
• MMLONGBENCH-DOC: Benchmarking Long-context Document Understanding with Visualizations (NeurIPS 2024 · similarity 0.89)
• URaG: Unified Retrieval and Generation in Multimodal LLMs for Efficient Long Document Understanding (AAAI 2026 · similarity 0.89)
• DocHop: Benchmarking Out-of-domain Multi-hop Reasoning in Information-Dense Documents (ICML 2026 · similarity 0.87)
• Not Search, But Scan: Benchmarking MLLMs on Scan-Oriented Academic Paper Reasoning (ICLR 2026 · similarity 0.87)