Perception-RFT is introduced, a training framework that applies Group Relative Policy Optimization (GRPO) to multimodal document QA, bypassing intermediate reasoning tokens to directly align visual features with structured grounding outputs, and a reasoning variant under identical reward settings is constructed.
Abstract
Efficient multimodal document question answering with explicit visual grounding, locating the precise document region that supports each answer remains an open challenge. Current approaches bifurcate into Supervised Fine-Tuning (SFT), which requires large annotated datasets and reaches optimization plateaus, and reasoning-centric Reinforcement Learning (RL), which depends on verbose intermediate traces that inflate inference token cost without clear benefit. We introduce Perception-RFT, a training framework that applies Group Relative Policy Optimization (GRPO) to multimodal document QA, bypassing intermediate reasoning tokens to directly align visual features with structured grounding outputs. To rigorously evaluate the necessity of reasoning, we construct a reasoning variant under identical reward settings. We find that reasoning-enabled models suppress their reasoning traces during training, converging to direct perception-based policies at the 4B parameter scale, reducing per-query inference token length by more than 60%, while reasoning-enabled RL underperforms perception-only training. Through a fine-grained analysis of Qwen3-VL-4B optimization dynamics, we confirm that SFT saturation and cold-start RL instability established in text-domain post-training extend to multimodal, and identify a previously uncharacterized Grounding Divergence: a selective trade-off between semantic robustness and geometric precision on two out of distribution (OOD) benchmarks (4,828 samples) under joint RL optimization. We further show that an early SFT$\rightarrow$RL transition achieves comparable precision with 65% less training data.
Video question answering requires models to ground language queries in visual evidence and, when necessary, reason over that evidence across time. Existing methods typically rely on long textual chain-of-thought rationales, even though many questions can be answered as soon as the relevant object, action, or frame is localized. We propose Dynamic Latent Reasoning (DyLaR), which first grounds a question in a short block of perception latents (continuous hidden states that encode query-relevant visual evidence), and then adaptively decides whether to append reasoning latents (continuous thoughts that reason over this evidence in latent space) before answering. DyLaR learns this behavior by grounding perception latents in verified visual evidence and distilling verified rationales into reasoning latents, followed by reinforcement learning that further refines when to reason. Across nine video benchmarks and four multimodal language model backbones, DyLaR improves average accuracy over same-backbone baselines while generating fewer than 20 tokens per query. On Qwen3-VL-4B, for example, DyLaR improves average accuracy over Qwen3-VL-4B-Thinking from 54.0 to 58.2 while reducing response length from 1,220.7 to 18.5 tokens per query. Ablations further show that grounded perception latents, rationale-supervised reasoning latents, and adaptive routing each improve accuracy.
Haotian Xia, Zilin Xiao, Junbo Zou et al.· 0 citations
Document question answering over scanned pages requires two coupled abilities: (i) canonicalizing complex layouts into a faithful textual structure, and (ii) selecting and reasoning over query-relevant evidence from that structure. Most existing pipelines decouple OCR from retrieval-augmented reasoning and optimize OCR for global reconstruction, which often misaligns with evidence needs and causes brittle grounding in multi-page settings. We propose UniDocVLM, an end-to-end framework that unifies OCR and visual RAG within a single vision-language model: the model first generates a structured parse of retrieved pages, then activates question-relevant evidence from the parse to support grounded reasoning and answering. To train UniDocVLM under heterogeneous supervision, we introduce a unified JR-GRPO reinforcement learning recipe with lightweight, verifiable rewards, including format, layout-aware OCR, evidence-consistency, and answer-correctness signals, and route them to the corresponding parts of the output to improve credit assignment and reduce interference. Experiments on multi-page document QA benchmarks show that UniDocVLM yields more reliable evidence grounding and improves downstream accuracy under complex layouts.
Zong-Sheng Cao, Anran Liu, Jun Xie et al.· Proceedings of the 32nd ACM...· 0 citations
Knowledge-based Visual Question Answering aims to answer questions about an image by integrating external knowledge with visual and textual information. Recent approaches often rely on in-context learning to prompt Large Language Models (LLMs) with multimodal context in a zero-shot or few-shot manner. However, we observe that directly concatenating heterogeneous visual descriptions and retrieved knowledge into long, unstructured prompts often degrades reasoning performance, due to both excessive irrelevant context and the lack of explicit relational structure. In this paper, we propose an LLM-based Structured Context Reasoning (SCoRe) framework that infers both explicit and implicit relationships for prediction. SCoRe consists of three stages: Context Acquisition, which generates diverse visual notes and retrieves explicit knowledge via an efficient two-stage multimodal retrieval strategy; Context Selection, which filters relevant visual, explicit, and implicit knowledge using LLM-guided selection; and Context Compression, which performs Relational Logic Distillation (RLD) to transform raw text into explicit entity-relation triplets. These relational triplets serve as a concise and structured prompt for final answer prediction. Extensive experiments on the OK-VQA and A-OKVQA benchmarks demonstrate that SCoRe consistently outperforms state-of-the-art methods.
Qiyou Liu, Yong Zhang, Jianjie Luo et al.· 0 citations
Knowledge-based Visual Question Answering (KB-VQA) aims to answer queries that necessitate reasoning over external knowledge sources beyond the visual content. Typically, current methods fuse multimodal features to retrieve external information, subsequently leveraging Multimodal Large Language Models (MLLMs) to derive answers from the retrieved evidence. However, these methods often struggle to capture structural associations within complex contexts to effectively filter noise. Furthermore, they frequently fail to ensure that the reasoning process remains strictly faithful to the retrieved evidence. To address these challenges, we propose SAFE-G, a Structure-Aware Faithful Evidence-guided Generation framework, which enables precise evidence localization and trustworthy reasoning. Specifically, we first employ a coarse-grained hybrid search fusing visual and textual modalities to recall candidate documents, and subsequently implement a structure-aware fine-grained graph retrieval that captures structural dependencies to filter noise and pinpoint precise evidence. Moreover, we introduce a reinforcement learning (RL) strategy with an evidence-grounded reward that assigns credit to correct answers only when the selected evidence is correct. This strict alignment constraint compels the model to anchor its response in the retrieved context, effectively enhancing its capability to locate evidence via multimodal features and perform faithful reasoning. Extensive experiments on the Encyclopedic-VQA and InfoSeek benchmarks demonstrate that SAFE-G outperforms prior methods by a margin of 8.9% and 3.5%, substantially enhancing the overall reasoning accuracy. Our source code is publicly available at: https://github.com/MINE-USTC/SAFE-G.
Long Shu, Shuochen Liu, Wei Chen et al.· 0 citations
Q-Guide is built, a small agent that reads a question, works out what evidence it is still missing, and calls targeted tool(s) to recover it---reading text where text is needed, zooming in where detail is needed, or grounding a region where position matters.
Multimodal question answering remains sensitive to noisy, incomplete, and weakly grounded evidence. Long unstructured contexts can introduce redundancy and encourage unsupported generation, while flat retrieval may overlook relations needed for multi-step reasoning. We present HANIA, a planner-guided multimodal graph framework for evidence-grounded question answering. HANIA processes the supplied image and text using a frozen vision-language model to extract concise question-relevant visual evidence with explicit abstention. It then constructs an input-grounded multimodal graph and applies a two-group finite-state planner to coordinate descriptive and relational evidence. Coverage-aware pruning retains a compact evidence set based on relevance, graph confidence, concept coverage, and modality diversity. The selected passages, visual statements, and graph triples are provided to a frozen instruction-tuned decoder. We evaluate HANIA on ScienceQA using answer accuracy, evidence-filtering quality, evidence-budget sensitivity, and efficiency. The results show that structured evidence planning and compact graph-guided retrieval can support competitive multimodal question answering without target-dataset fine-tuning or iterative retrieval. The code is available at https://github.com/Zafar-southeast/HANIA.
Zafar Ali, Asad Khan, Nimbeshaho Thierry et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
