AWM: Answerable Working Memory for Long-Document VQA Agents
AWM-GRPO improves long-document VQA accuracy by 8.1 and 11.9 points.
Key Findings
Methodology
AWM-GRPO integrates the answerability of terminal working memory into the GRPO reward. It enhances the advantage of answer-correct trajectories whose terminal working memory remains answerable, tested on MMLONGBENCH-DOC and LONGDOCURL datasets.
Key Results
- AWM-GRPO improves final-answer accuracy over the RAG baseline by 8.1 and 11.9 points on MMLONGBENCH-DOC and LONGDOCURL datasets.
- Even with gold evidence pages, 42.5% of correct answers cannot be answered from terminal working memory alone.
- AWM-GRPO reduces the memory-missing-correct rate by 2.7 points.
Significance
This research fills a gap in evaluating memory quality in long-document VQA by introducing Answerable Working Memory (AWM). AWM-GRPO not only improves final-answer accuracy but also enhances the answerability of terminal working memory, advancing the VQA field.
Technical Contribution
AWM-GRPO introduces a new reward signal in the existing GRPO framework, emphasizing the quality of terminal working memory. This approach offers a new optimization direction by considering both final-answer correctness and memory answerability.
Novelty
AWM-GRPO is the first to incorporate terminal working memory answerability as a reward signal in GRPO, differing from traditional methods that focus solely on final-answer correctness.
Limitations
- AWM-GRPO faces challenges in handling complex tables and layouts, requiring further optimization.
- The method performs poorly in certain cross-source settings.
Future Work
Future research could explore optimizing AWM-GRPO's performance in structured evidence preservation and expanding its application to other VQA tasks.
AI Executive Summary
Long-document Visual Question Answering (VQA) systems face challenges in extracting and integrating information from numerous pages. Traditional methods focus on final-answer correctness, neglecting the quality of working memory, which may lead to correct answers unsupported by memory. AWM-GRPO addresses this by incorporating terminal working memory answerability into the GRPO reward.
AWM-GRPO's experiments on MMLONGBENCH-DOC and LONGDOCURL datasets demonstrate enhanced final-answer accuracy and terminal working memory answerability. This method prioritizes final answers while giving higher advantages to answerable terminal working memory.
Despite AWM-GRPO's excellent performance across datasets, it faces challenges with complex tables and layouts. Future research could explore optimizing its structured evidence preservation and expanding its application to other VQA tasks.
Deep Analysis
Background
Long-document Visual Question Answering (VQA) systems need to extract and integrate information from numerous pages. Existing methods focus on final-answer correctness, neglecting the quality of working memory, which may lead to correct answers unsupported by memory.
Core Problem
Current VQA systems primarily focus on final-answer correctness in evaluations, neglecting the quality of working memory. This may lead to correct answers unsupported by memory.
Innovation
AWM-GRPO is the first to incorporate terminal working memory answerability as a reward signal in GRPO, differing from traditional methods that focus solely on final-answer correctness.
Methodology
- �� AWM-GRPO integrates terminal working memory answerability
- �� Introduces new reward signal in GRPO
- �� Emphasizes terminal working memory quality
- �� Tested on MMLONGBENCH-DOC and LONGDOCURL datasets
Experiments
Experiments conducted on MMLONGBENCH-DOC and LONGDOCURL datasets, compared against RAG baseline. Key metrics include final-answer accuracy and terminal working memory answerability.
Results
AWM-GRPO improves final-answer accuracy over the RAG baseline by 8.1 and 11.9 points and reduces the memory-missing-correct rate by 2.7 points.
Applications
AWM-GRPO can be used to enhance the accuracy of long-document VQA systems, especially in scenarios requiring extraction and integration of information from numerous pages.
Limitations & Outlook
AWM-GRPO faces challenges in handling complex tables and layouts, requiring further optimization.
Plain Language Accessible to non-experts
Imagine you're in a large library searching for specific information in a book. Traditional methods are like flipping through the shelves, finding the answer, and responding immediately. But if you leave the shelf, you might forget the details. AWM-GRPO is like jotting down key points once you find the information, allowing you to answer even after leaving the shelf. This method ensures you can answer questions from memory even without the original book.
ELI14 Explained like you're 14
Imagine you're playing a game and need to find clues on a big map. Traditional methods are like finding the clue and answering right away, but if you leave the map, you might forget the details. AWM-GRPO is like noting down key points once you find the clue, so you can answer even after leaving the map. This method ensures you can answer questions from memory even without the original map.
Glossary
AWM (Answerable Working Memory)
A memory mechanism for long-document VQA systems that ensures terminal working memory contains enough information to support answering questions.
In AWM-GRPO, AWM is used as part of the reward signal.
GRPO (Group Relative Policy Optimization)
A reinforcement learning method that updates policies by comparing relative advantages within a group.
AWM-GRPO uses GRPO to optimize terminal working memory quality.
RAG (Retrieval-Augmented Generation)
A model architecture combining retrieval and generation to extract information from large-scale documents.
AWM-GRPO is compared against the RAG baseline in experiments.
MMLONGBENCH-DOC
A dataset for evaluating long-document VQA systems, containing multi-page documents and complex questions.
AWM-GRPO is tested on this dataset to validate its effectiveness.
Terminal Working Memory
A memory module in VQA systems that records key information extracted from pages.
AWM-GRPO emphasizes the answerability of terminal working memory.
Open Questions Unanswered questions from this research
- 1 How can AWM-GRPO be further optimized for complex tables and layout issues?
- 2 What is the applicability of AWM-GRPO to other VQA tasks?
Applications
Immediate Applications
Long-Document Information Extraction
AWM-GRPO can enhance accuracy in extracting and integrating information from long documents.
Long-term Vision
Intelligent Document Analysis
AWM-GRPO could drive the development of intelligent document analysis systems, helping users quickly access key information.
Abstract
Long-document visual question answering increasingly relies on VLM agents that retrieve candidate pages, inspect page images, write findings to working memory, and synthesize answers. Working memory should carry answer-supporting evidence across page inspections for later grounded answering, yet existing evaluation mainly checks final-answer correctness and evidence-page access. This creates a memory-quality blind spot: an agent may reach the right page and answer correctly while leaving behind memory too generic or incomplete to support answering once page context is removed. We introduce \emph{memory-only answerability}, a diagnostic that asks whether a reader can answer from the question and terminal working memory alone. Building on this diagnostic, \emph{Answerable Working Memory} (AWM) treats terminal working memory as an answerable evidence artifact, and AWM-GRPO incorporates this signal into the GRPO reward while preserving final-answer priority. Under GRPO, this reward assigns higher advantages to answer-correct trajectories whose terminal working memory remains answerable. On \textsc{MMLongBench-Doc}, even when gold evidence pages are provided, 42.5\% of correct answers still cannot be answered from terminal working memory alone. AWM-GRPO improves final-answer accuracy over the RAG baseline by 8.1 and 11.9 points on \textsc{MMLongBench-Doc} and \textsc{LongDocURL} and reduces the memory-missing-correct rate by 2.7 points over answer-only GRPO.