Paper deep dive
See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL
Yilian Liu, Sicong Leng, Guoshun Nan, Junyi Zhu, Jiayu Huang, Minghao Sun, Xuancheng Zhu, Yisong Chen, Zexian Wei, Xiaofeng Tao
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 6/20/2026, 10:33:05 AM
Summary
The paper introduces Visual Evidence Pre-Alignment (VEPA), an intermediate training stage for Multimodal Large Language Models (MLLMs) positioned between pretraining and post-training. VEPA addresses the issue of weak visual grounding in standard caption-based pretraining by teaching models to generate question-conditioned visual evidence. It utilizes Group Relative Policy Optimization (GRPO) with a sufficiency-driven objective. A key component is a 'blind reader' (a frozen LLM) that evaluates whether the generated textual evidence is sufficient to recover the ground-truth answer without seeing the image, thereby decoupling visual perception from reasoning. Experiments using Qwen2-VL-2B as a backbone demonstrate that VEPA enhances performance across diverse benchmarks and improves visual grounding and robustness to language priors.
Entities (7)
Relation Signals (4)
Blind Reader ā evaluates ā Visual Evidence Pre-Alignment
confidence 100% Ā· A frozen blind reader (an LLM) answering and verifying whether the evidence suffices to recover the answer.
Qwen2-VL-2B ā isbackbonefor ā Visual Evidence Pre-Alignment
confidence 100% Ā· We use Qwen2-VL-2B (Wang et al., 2024) as the backbone for all model variants.
Visual Evidence Pre-Alignment ā isintermediatestagebetween ā Pretraining
confidence 100% Ā· an intermediate stage between pretraining and post-training
Visual Evidence Pre-Alignment ā uses ā Group Relative Policy Optimization
confidence 100% Ā· VEPA optimizes evidence generation via sufficiency-driven Group Relative Policy Optimization (GRPO).
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Multimodal large language models (MLLMs) integrate strong text reasoning with visual inputs, yet their responses can be inconsistent with the underlying images, indicating ineffective utilization of visual evidence during inference. The prevailing training paradigm relies on large-scale caption-based pretraining for general alignment, followed by supervised fine-tuning and reinforcement learning to enable instruction following and complex reasoning. However, such pretraining provides only weak visual grounding: short, coarse captions bias models toward salient objects while neglecting fine-grained visual evidence. In this paper, we introduce Visual Evidence Pre-Alignment (VEPA), an intermediate stage between pretraining and post-training that explores a novel sufficiency-driven objective with Group Relative Policy Optimization (GRPO) to optimize question-conditioned visual evidence descriptions. Extensive experiments across diverse benchmarks show that our VEPA consistently enhances performance on visually demanding evaluations and complements standard supervised post-training. Further analyses show that the income stems from strengthened, transferable visual grounding, rather than from additional task-specific training.
Tags
Links
- Source: https://arxiv.org/abs/2606.17678v1
- Canonical: https://arxiv.org/abs/2606.17678v1
Trouble viewing inline? Open PDF directly ā
Full Text
75,216 characters extracted from source content.
Expand or collapse full text
See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL Yilian Liu 1* Sicong Leng 2* Guoshun Nan 1 Junyi Zhu 1 Jiayu Huang 1 Minghao Sun 1 Xuancheng Zhu 1 Yisong Chen 3 Zexian Wei 1 Xiaofeng Tao 1 1 Beijing University of Posts and Telecommunications, China 2 Nanyang Technological University, Singapore; 3 China Telecom, China liuyilian,nanguo2021@bupt.edu.cn; Lengsicong@gmail.com Abstract Multimodal large language models (MLLMs) in- tegrate strong text reasoning with visual inputs, yet their responses can be inconsistent with the underlying images, indicating ineffective utiliza- tion of visual evidence during inference. The prevailing training paradigm relies on large-scale caption-based pretraining for general alignment, followed by supervised fine-tuning and reinforce- ment learning to enable instruction following and complex reasoning. However, such pretrain- ing provides only weak visual grounding: short, coarse captions bias models toward salient objects while neglecting fine-grained visual evidence. In this paper, we introduce Visual Evidence Pre- Alignment (VEPA), an intermediate stage be- tween pretraining and post-training that explores a novel sufficiency-driven objective with Group Relative Policy Optimization (GRPO) to optimize question-conditioned visual evidence descriptions. Extensive experiments across diverse benchmarks show that our VEPA consistently enhances per- formance on visually demanding evaluations and complements standard supervised post-training. Further analyses show that the income stems from strengthened, transferable visual grounding, rather than from additional task-specific training. 1. Introduction Multimodal large language models (MLLMs) have recently achieved strong performance on diverse tasks such as doc- ument and chart comprehension and diagrammatic math reasoning. Despite these advances, MLLMs may produce * Equal contribution . Correspondence to: Guoshun Nan<nan- guo2021@bupt.edu.cn>. Preprint. June 17, 2026. v Q: Find the value. A: <think>/<answer> Q: How many pins? A: 8 push pins visible. VEPA Added Test Standard Recipe Test Q: Describe the image. A: A man and a dog.... Q: How many pins? A: 6 push pins visible , attached to the bullet. Can we teach models to see better before learning to answer? Word Modeling Course Alignment Task Reasoning Ignore Details Instruction Following question- conditioned caption Blind Reader Stage 1.5: VEPA Perception Activated sufficiency-driven GRPO Stage1 Pretrain Stage2 Post-train Figure 1. Motivation and overview. The standard two-stage recipe often yields coarse alignment and encourages shortcut an- swering that ignores visual details. We insert VEPA as an interme- diate stage that trains the model to produce question-conditioned visual evidence using sufficiency-driven GRPO, with a frozen blind reader (an LLM) answering and verifying whether the evi- dence suffices to recover the answer. This āsee first, answer laterā pre-alignment activates perceptual ability and improves down- stream visual grounding. outputs that are weakly grounded in the underlying im- age (Luo et al., 2025; Xia et al., 2025), including omissions of critical visual details, incorrect attribution of attributes or relations, and hallucinated content (Li et al., 2025; Xia et al., 2025). Such failures are not confined to specific tasks or domains but instead reflect a pervasive limitation in how visual evidence is exploited during inference. Most existing MLLM training pipelines follow a two-stage paradigm (Liu et al., 2023; Zhu et al., 2024). In the pretraining stage, models are trained on large-scale imageācaption corpora to establish coarse visionālanguage alignment (Li et al., 2020). In the post-training stage, supervised fine-tuning, often combined with reinforcement learning, is applied to improve instruction following and downstream task perfor- mance (Christiano et al., 2017; Zhai et al., 2024). However, 1 arXiv:2606.17678v1 [cs.CV] 16 Jun 2026 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL caption-driven pretraining alone is insufficient to ensure ro- bust visual perception prior to post-training. Captions are typically short and coarse, emphasizing salient objects (Lin et al., 2014) or global scene descriptions while omitting many detailed attributes, relations, and less prominent re- gions. This supervision biases models toward a narrow subset of visual content and provides limited incentive to encode fine-grained or question-relevant visual evidence, leading models to rely disproportionately on language priors during inference (Goyal et al., 2017). A straightforward remedy is to strengthen pretraining su- pervision by replacing short captions with dense descrip- tions (Zheng et al., 2024; Zeng et al., 2025) or augment- ing training with perception-oriented datasets (e.g., OCR) that explicitly encode fine-grained visual content. However, this approach is limited both practically and fundamentally. High-quality dense captions and OCR (Chen et al., 2024c) annotations are costly (Dong et al., 2025; Liu et al., 2025; Shen et al., 2025) to collect at scale for diverse web im- ages, and manual annotation pipelines introduce additional biases, omissions, and inconsistencies (Misra et al., 2016; Hu et al., 2023). More fundamentally, the high information density of visual inputs compared to textual representations makes it inherently difficult for any finite description to faithfully encode all objects, relations, and spatial details in a scene. Thus, even dense captions remain a lossy (Dubois et al., 2021) and biased proxy for visual content (Chen et al., 2024a), and simply scaling static imageātext pairs is insuf- ficient to achieve the level of perceptual precision required by complex multimodal reasoning tasks. In this work, we introduce Visual Evidence Pre-Alignment (VEPA), an intermediate training stage that teaches MLLMs to generate question-conditioned visual evidence, textual descriptions that captures the image details needed to solve a given question. As illustrated in Figure 1, VEPA optimizes evidence generation via sufficiency-driven Group Relative Policy Optimization (GRPO) (Shao et al., 2024). During training, a frozen blind reader (an LLM), conditioned only on the question and the generated evidence, serves as an auxiliary evaluator by verifying whether the evidence is sufficient to recover the ground-truth answer. We carefully design the reward to discourage answer leakage and degen- erate repetitive evidence, thereby decoupling visual ground- ing from answer generation. Across diverse evaluation set- tings, VEPA consistently outperforms the standard training recipe on benchmarks spanning knowledge-intensive and compositional VQA, fine-grained perception, and holistic multimodal evaluation. Ablation studies and qualitative analyses show that these gains arise from strengthened vi- sual perception rather than additional task-level supervision, and that the learned grounding transfers robustly to out-of- distribution data. In summary, our contributions are in three aspects: ā¢ We propose Visual Evidence Pre-Alignment (VEPA), a novel intermediate training stage between pretrain- ing and post-training that explicitly strengthens visual perception prior to task-level instruction tuning, en- couraging MLLMs to attend to and encode relevant visual evidence before answering. ā¢We instantiate VEPA with a GRPO-based reinforce- ment learning framework that trains models to generate question-conditioned visual evidence. A frozen blind reader evaluates whether the generated evidence suf- fices to recover the ground-truth answer, decoupling visual grounding from answer generation and enabling training with existing QA data without additional an- notation cost. ā¢ Extensive experiments across diverse benchmarks and model backbones demonstrate consistent improve- ments on visually demanding evaluations. Ablation studies and qualitative analyses further show that VEPA strengthens visual grounding and induces trans- ferable perceptual capabilities that generalize to out-of- distribution settings. 2. Related Work Multimodal pretraining and post-training. MLLMs fol- low a two-stage recipe: large-scale visionālanguage pretrain- ing for coarse alignment, followed by instruction tuning to improve task-level behaviors (Alayrac et al., 2022; Li et al., 2023a; Liu et al., 2023; 2024; Wang et al., 2024). This paradigm has produced strong systems across late-fusion architectures that connect a frozen or lightly-tuned vision en- coder to an LLM as well as increasingly capable end-to-end or early-fusion variants. Recent backbones such as Qwen2- VL (Wang et al., 2024) further improve resolution handling and general visual understanding, strengthening the foun- dation for downstream multimodal adaptation (Wang et al., 2024). However, empirical evaluations consistently suggest that coarse caption-driven supervision alone does not always yield reliable fine-grained grounding required by complex VQA-style queries (Fu et al., 2025). Visual grounding. A recurring challenge for MLLMs is the tendency to under-use visual input, leading to omissions or hallucinated content (Li et al., 2025; Xia et al., 2025) when language priors (Goyal et al., 2017) dominate generation. This issue has motivated dedicated benchmarks and diagnos- tics that quantify visual faithfulness and robustness beyond raw accuracy, including object-hallucination probes and broad-spectrum evaluation suites (Li et al., 2023b; Rohrbach et al., 2018). Complementary lines of work mitigate halluci- nations via improved post-training objectives (Zheng et al., 2 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL 2025), decoding-time constraints (Li et al., 2025), or aux- iliary verification mechanisms that encourage consistency with perceived evidence (Zheng et al., 2025). Despite these advances, many approaches still optimize answer generation directly (Wang et al., 2025), making it difficult to isolate and strengthen the upstream visual extraction process that should support downstream reasoning. Intermediate evidence and verifier-based reinforcement learning. Generating intermediate representationsāsuch as rationales, descriptions, or other textual evidenceāhas been explored to improve grounding and interpretability in multimodal reasoning, where models must surface query- relevant visual details (Li et al., 2018; Rao et al., 2021). When reliable intermediate supervision is unavailable, re- cent language-model post-training leverages verifier-style or preference-based signals as scalable supervision with- out token-level labels (Rafailov et al., 2023; Wen et al., 2025). In particular, group-based policy optimization such as GRPO removes an explicit critic and estimates baselines from within-group scores, simplifying RL fine-tuning for long responses (Shao et al., 2024). Our work aligns with this verifier-based direction but targets a different object. Rather than optimizing the final answer directly, we train a question-conditioned evidence channel to be sufficient for solving and resistant to shortcut behaviors, bridging coarse pretraining and downstream post-training. 3. Preliminary 3.1. Evidence Decomposition for Visual Grounding The goal of an MLLM is to generate an answeracondi- tioned on an imagevand a questionq, ideally approximating P (a| v,q). In practice, caption-based pretraining provides only coarse visionālanguage alignment, and answer-level supervision makes improvements in perception mostly a byproduct of reasoning learning. As a result, models may over-rely on language priors, drifting towardP (a| q)and under-utilizing the visual input v. To address this, we decouple visual perception from rea- soning by introducing a latent variablee, termed visual evidence, which serves as an information bottleneck be- tween perception and answer generation. The generation process is formalized as: P (a| v,q) = X eāE P (a| e,q)Ā· P (e| v,q).(1) Here,P (e | v,q)acts as a visual representation policy (MLLM) that maps the image and question to a textual evidencee, andP (a| e,q)is a reasoning policy (LLM) that produces the answer from the evidence and the question. Under this decomposition, improving visual alignment amounts to shapingP (e | v,q)such that the induced ev- idence is both informative for answering and genuinely grounded in the image. We therefore require the evidencee to satisfy two properties: Sufficiency. Conditioned on the question, the evidence should contain enough information to determine the answer: P (a| v,q)ā P (a| e,q) for eā¼ P (e| v,q).(2) Onceeis known, the residual contribution of the raw image v to predicting a should be small. Visual dependence. The evidence must depend on both the image and the question, rather than being reconstructible from either alone. In particular, for imageāquestion pairs (v 1 ,q)and(v 2 ,q), or(v,q 1 )and(v,q 2 ), that induce differ- ent answers, we require: P (e| v 1 ,q)Ģø= P (e| v 2 ,q),P (e| v,q 1 )Ģø= P (e| v,q 2 ). (3) When the above properties hold approximately, the model cannot satisfy the objective by ignoring the image. Instead, it must route visual information through the evidence chan- nelP (e | v,q)before producing an answer. The interme- diate training stage we introduce is designed to explicitly encourage these properties. 3.2. Perception Coverage from Data Diversity LetDdenote the training distribution over triplets(v,q,a). While the sufficiency and dependence conditions above are defined for individual instances, our objective is to enforce them approximately on average overD. Concretely, we would like the evidence policyP (e | v,q) to satisfy: E (v,q,a)ā¼D h KL P (a| v,q) P (a| e,q) i ā¤ Īµ suff , (4) for a smallĪµ suff ā„ 0, ensuring that answers can be pre- dicted nearly as well from(e,q)as from(v,q)on average. In addition, to prevent degeneracy, the evidence must re- tain non-trivial information about both the image and the question, for example, I D (e;v | q)ā„ Ī“ v , I D (e;q | v)ā„ Ī“ q ,(5) withĪ“ v ,Ī“ q > 0, whereI D (Ā·;Ā· | Ā·)denotes conditional mutual information under the joint distribution induced by D and P (e| v,q). Rather than supervisingewith increasingly dense captions, which remain fundamentally constrained by the informa- tion density of text, we leverage the diversity inherent in VQA-style data. When triplets(v,q,a)are sampled from a heterogeneous corpus, different questions probe distinct objects, attributes, relations, and regions within the same 3 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL image, including details rarely emphasized in caption-based supervision. Encouraging sufficiency and dependence overDtherefore pushes the evidence policy to cover a broad range of task-relevant visual details. Each evi- dence instance need only be sufficient for its own question, while the collection of questionāevidence pairs across the dataset provides complementary and diverse textual views of visual content. 4. Methodology 4.1. Evidence Generation as Perceptual Pre-Alignment We propose Visual Evidence Pre-Alignment (VEPA), an intermediate training stage that explicitly aligns visual per- ception before task-level post-training. As illustrated in Figure 2, VEPA trains an MLLM to generate question- conditioned visual evidence, textual descriptions that encode the visual information for a given question, without directly optimizing answer generation. This section first formalizes the evidence-generation objective (Ā§4.1), then presents the VEPA optimization framework (Ā§4.2), and finally describes the blind-reader-based reward design (Ā§4.3). We instantiate the evidence decomposition introduced in Ā§3.1 as an intermediate training objective. Since ground- truth visual evidence annotations are unavailable at scale and dense captions remain both costly and incomplete, su- pervised learning over evidence tokens is infeasible. So we cast evidence generation as a policy optimization problem. LetD = (v i ,q i ,a i )denote a VQA training set. During VEPA, the MLLM is not trained to produce the final answer a. Instead, for each triplet(v,q,a), it generates an inter- mediate textual sequenceeintended to encode the visual information necessary to answerq. For each rollout, the policy samples a single evidence sequence up to a fixed length. Learning is driven solely by a sequence-level reward computed after the evidence is fully generated. 4.2. VEPA Optimization Framework The proposed VEPA optimizes the evidence policyĻ Īø (e| v,q)using reinforcement learning with a sequence-level re- wardR(e;v,q,a ā ). Given a datasetDof triplets(v,q,a ā ), the objective is J (Īø) = E (v,q,a ā )ā¼D E eā¼Ļ Īø (Ā·|v,q) R(e;v,q,a ā ) .(6) Optimizing this objective directly is unstable for long, free- form evidence sequences. We therefore adopt sufficiency- driven Group Relative Policy Optimization (GRPO) as a practical optimization strategy 1 . 1 Importantly, GRPO serves solely as an optimizer for VEPAās evidence objective; our contribution lies in the formulation of the objective and reward, not in the optimization algorithm itself. At each update step, for a given triplet(v,q,a ā ), we sample a group ofGcandidate evidence sequencese g G g=1 from the current policy. Each candidate receives a scalar reward R g , from which we compute a group-relative standardized advantage A g = R g ā Ģ R Ļ R + Ī“ ,(7) where Ģ RandĻ R denote the group mean and standard devia- tion. Policy updates follow a PPO-style clipped surrogate objective with a KL regularization (Schulman et al., 2017) term that constrains deviation from a frozen reference policy Ļ ref (the pretrained model before VEPA): J VEPA (Īø) = E " 1 G G X g=1 1 T g T g X t=1 L clip g,t (Īø) ā Ī² D KL (Ļ Īø ā„Ļ ref ) t # . (8) This formulation enables stable optimization of long evi- dence sequences while preserving linguistic coherence. 4.3. Blind Reader-Based Sufficiency Reward To operationalize the sufficiency criterion in Ā§3.1 without evidence annotations, we introduce a frozen blind reader modelf Ļ . The blind reader is an instruction-tuned language model that observes only the generated evidenceeand ques- tion q, and never accesses the image. For each training example, the blind reader is prompted to answer the question using the evidence alone and to indicate whether the evidence directly leaks the answer. LetĖa(e,q) denote the predicted answer after normalization. Based on this answer, we then define a binary solvability score s(e,q,a ā ) = I[Ėa(e,q) = a ā ],(9) which encourages the evidence to be sufficient for answers. To discourage trivial solutions that restate the answer, the blind reader emits an honesty flag indicating whether answer leakage is detected. We define h(e,q)ā0, 1,(10) where h = 1 denotes honest evidence. Finally, to suppress degenerate repetition, we apply a weak penaltyĪ³(e) ā (0, 1]based on simple repetition statistics. The resulting sequence-level reward can be expressed as R(e,q,a ā ) = s(e,q,a ā )Ā· h(e,q)Ā· Ī³(e).(11) Maximizing the expected reward encourages evidence that is sufficient for a blind reader while remaining image- grounded and non-degenerate, without requiring any human evidence supervision. 4 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL inSufficient Sufficient Question: How many push pins are there? MLLM Policy Model LLM Blind Reader Answer (a) Answer (a) Perception activate? Visual Evidenceļ¼eļ¼ Compute Reward Prompt: Your goal is to provide a comprehensive description will be used by a blind person to reason about and answer the question. Therefore, your description must act as the raw visual evidence required for reasoning. Update policy Group Relative Advantage v q PretrainVEPAPost-train verified Figure 2. Framework of VEPA. VEPA is inserted between pretraining and post-training. Given an imagevand a questionq, the policy MLLMĻ Īø is prompted to generate question-conditioned visual evidence and samples a group of candidate visual evidencee. A frozen text-only blind reader (auxiliary LLM) answers using only (q,e). We optimize Ļ Īø with a novel sufficiency-driven objective via GRPO. Overall, the proposed VEPA reframes visual grounding as an explicit pre-alignment problem by isolating evidence generation from answer prediction. By optimizing question- conditioned visual evidence using a blind-readerābased suf- ficiency signal, VEPA encourages models to encode task- relevant visual information before engaging in downstream reasoning. This design enables effective perceptual align- ment without additional annotations and is complementary to standard supervised post-training, providing a principled mechanism for improving visual grounding in MLLMs. 5. Experiments We conduct experiments to evaluate the effectiveness of our VEPA, and focus on the following questions: RQ1. Downstream Performance. Does VEPA consistently improve performance across diverse multimodal bench- marks compared to the standard training pipeline? RQ2. Visual Grounding. Does VEPA strengthen vi- sual grounding by encouraging the generation of question- conditioned evidence that is sufficient for reasoning? RQ3. Visual Reliance. Does VEPA increase reliance on visual inputs over language priors, as reflected by robustness to visual perturbations and reduced hallucination? RQ4. Training Dynamics. How does visual perception evolve over the course of VEPA training? 5.1. Experimental Setup Implementation details. We use Qwen2-VL-2B (Wang et al., 2024) as the backbone for all model variants. Dur- ing the VEPA stage, we optimize the evidence-generation policy using our sufficiency-driven GRPO on a curated sub- set of FineVision (Wiedmann et al., 2025). Specifically, we uniformly sample 5k training instances from the Sci- enceQA (Lu et al., 2022), AI2D-Merge (Kembhavi et al., 2016), ChartQA (Masry et al., 2022), Geo3K (Lu et al., 2021), TextVQA (Singh et al., 2019), and CLEVR (Johnson et al., 2017). These datasets jointly cover complementary perception-centric skills, including diagram and scientific reasoning, chart and plot understanding, geographic and map recognition, text understanding in natural images, and compositional visual reasoning. To better align with the objective of visually grounded evi- dence generation, we further prioritize questions with high visual dependency, such that the correct answer cannot be re- liably inferred from the question alone. This filtering yields diverse evidence patterns that better approximate the tar- get distribution and improves the robustness of the learned evidence policy. All GRPO training is implemented using VERL. To com- pute the evidence-sufficiency reward, we employ a frozen text-only blind reader based on Qwen2.5-7B-Instruct (Qwen et al., 2025), which predicts answers conditioned solely on the question and the generated evidence. Unless otherwise specified, we report exact-match accuracy on downstream VQA benchmarks with identical decoding settings across all variants. Additional implementation details and hyperpa- rameters are provided in Appendix A.1. 5 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL 5.2. RQ1: Downstream Performance Across Tasks To assess whether inserting VEPA between pretraining and standard post-training improves downstream performance, we conduct controlled comparisons in which all variants share the same backbone, supervised fine-tuning (SFT) data, and SFT hyperparameters. Specifically, all models are ini- tialized from Qwen2-VL-2B, trained with standard SFT on 20k examples, and evaluated either directly (SFT) or after an additional VEPA Etage followed by SFT (VEPA+SFT). We consider four single-domain SFT settings using A- OKVQA (Schwenk et al., 2022), ChartQA (Masry et al., 2022), TextVQA (Singh et al., 2019), and GQA (Hudson & Manning, 2019), as well as a Mixed setting constructed by uniformly sampling 5k examples from each dataset. Models are evaluated on six benchmarks, including in-domain evalu- ations (marked with ā ), out-of-domain transfer benchmarks, and two general-purpose multimodal benchmarks (MME and MMStar). Table 1 reports accuracy comparisons. Overall performance trends. Across all five SFT settings, VEPA consistently improves average performance (+0.5 to +3.8 points) without degrading in-domain accuracy. In- domain results are either preserved or modestly improved, indicating that VEPA does not trade task specialization for robustness. This suggests that the VEPA stage provides complementary supervision signals that are compatible with downstream SFT objectives. Generalization under domain shift. Performance gains are often more pronounced on out-of-domain benchmarks than on the in-domain metric. In particular, perception-intensive tasks such as ChartQA and TextVQA frequently exhibit larger improvements. For example, under A-OKVQA SFT, VEPA substantially improves ChartQA and TextVQA while also improving A-OKVQA itself; under GQA SFT, VEPA again yields sizable gains on ChartQA and TextVQA, with comparatively small changes on in-domain GQA. Such asymmetric improvements are unlikely to arise from generic additional training, which would be expected to affect all benchmarks more uniformly. Instead, they are consistent with VEPA strengthening transferable, question-conditioned visual grounding. Complementarity with data diversification. If VEPA merely compensated for limited domain coverage in SFT, its effect should diminish under the Mixed SFT setting. However, VEPA continues to improve average performance and yields a notable gain on GQA even with a stronger, diversified SFT baseline. This indicates that VEPA enforces a distinct inductive bias, greater reliance on visual evidence, rather than functioning solely as a substitute for broader supervised data. General multimodal evaluation. VEPA also improves performance on MME and MMStar across SFT settings. Since these benchmarks are not aligned with any specific SFT domain, the gains suggest that VEPA enhances the modelās general ability to ground answers in visual evidence, rather than improving benchmark-specific patterns. Overall, these results demonstrate that VEPA serves as an effective intermediate alignment stage: it improves average downstream performance, preserves in-domain accuracy, and yields the largest benefits under domain shift, where transferable perceptual grounding is most critical. Insight I. VEPA consistently improves downstream per- formance and robustness by activating transferable visual grounding beyond standard supervised fine-tuning. 5.3. RQ2: Evidence Sufficiency and Selectivity While RQ1 establishes that VEPA improves downstream performance, it does not directly verify whether VEPA opti- mizes the intended objectiveānamely, producing question- conditioned visual evidence that is sufficient for reasoning. RQ2 therefore evaluates the quality of the learned evidence representations. For each imageāquestion pair, we prompt either the base model or the model after the VEPA stage to generate a question-conditioned visual description, ex- plicitly discouraging direct answer disclosure. We then provide only the question and the generated description to a frozen text-only blind reader (Qwen2.5-7B-Instruct), which attempts to answer POPE and MMStar without access to the image. Because the blind reader never observes visual in- puts, higher accuracy directly reflects whether the generated evidence surfaces task-relevant visual information. Table 2 shows that VEPA consistently improves blind-reader accuracy on both benchmarks, from 42.53% to 44.27% on MMStar and from 78.05% to 79.25% on POPE. Crucially, these gains are accompanied by shorter descriptions: the av- erage output length decreases from 201.97 to 187.86 tokens on MMStar and from 464.21 to 365.16 tokens on POPE. This rules out verbosity as a trivial explanation for the im- proved solvability. Consistent with this observation, the mean rollout length exhibits an initial transient increase followed by a gradual decrease and stabilization, rather than unbounded growth (See details in Appendix A.7). After an initial transient increase, it rapidly stabilizes, indicating convergence to- ward concise yet informative evidence rather than progres- sively longer descriptions. Taken together, these results indicate that VEPA improves the selectivity and sufficiency of generated evidence, enabling task-critical visual cues to be externalized in a compact form. This provides a direct mechanistic explanation for the stable downstream gains observed in RQ1. 6 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL Table 1. Downstream benchmark results under different SFT data settings. Accuracy (%) is reported for each target benchmark. Rows are grouped by the SFT dataset used for post-training (SFT data), comparing the SFT baseline against VEPA+SFT (shaded). Avg. denotes the macro-average over all benchmarks. ā marks in-domain evaluation. Subscripts indicate the change relative to SFT (ā improvement,ā decline, - no change). SFT dataMethodA-OKVQAChartQATextVQAGQAMMEMMStarAvg. A-OKVQA SFT59.70 ā 51.8862.7353.7075.2241.0757.38 VEPA+SFT66.60 ā ā6.9 58.96 ā7.1 66.78 ā4.1 54.75 ā1.1 76.95 ā1.7 42.80 ā1.7 61.14 ā3.8 ChartQA SFT59.6069.44 ā 77.5659.7081.7040.9364.82 VEPA+SFT59.60 ā 70.28 ā ā0.8 78.96 ā1.4 60.99 ā1.3 82.13 ā0.4 41.93 ā1.0 65.65 ā0.8 TextVQA SFT34.9568.0880.66 ā 38.4082.1341.8757.68 VEPA+SFT35.50 ā0.6 68.64 ā0.6 81.44 ā ā0.8 40.60 ā2.2 80.55 ā1.6 42.40 ā0.5 58.19 ā0.5 GQA SFT42.0060.4063.4964.41 ā 78.6743.0758.67 VEPA+SFT42.10 ā0.1 64.40 ā4.0 66.15 ā2.7 64.52 ā ā0.1 80.55 ā1.9 43.80 ā0.7 60.25 ā1.6 Mixed SFT62.50 ā 70.32 ā 79.73 ā 51.07 ā 79.9744.5364.69 VEPA+SFT62.70 ā ā0.2 70.44 ā ā0.1 79.88 ā ā0.2 53.69 ā ā2.6 80.26 ā0.3 45.07 ā0.5 65.34 ā0.7 Table 2. Blind-reader evaluation of evidence sufficiency. We prompt the base model and the model after the VEPA stage to generate question-conditioned descriptions, and provide only the question and the generated description to a frozen text-only blind reader (Qwen2.5-7B-Instruct) to answer POPE and MMStar with- out image access. We report Acc. (%, accuracy) and the average description length (tokens). Method MMStarPOPE Acc.LengthAcc.Length Base42.53201.9778.05464.21 VEPA44.27187.8679.25365.16 Insight I. VEPA promotes concise yet sufficient visual evidence, making task-relevant visual information recov- erable without relying on verbose descriptions. 5.4. RQ3: Reliance on Visual Inputs Table 3. Robustness under image corruption. Models are trained with A-OKVQA SFT. We report clean accuracy (%), retention under three corruption types (Blur, Partial noise, Pure noise), and AUC of the retention curve (lower is better). Metric ChartQAGQA SFTVEPA+SFTSFTVEPA+SFT Clean Acc. (%) 51.8858.9652.1353.07 Blur (r)0.0840.0740.8380.836 Partial (r)0.7530.6800.8740.867 Pure (r)0.0620.0520.5250.527 AUCā (over r) 0.2460.2200.7690.767 While RQ1 demonstrates consistent downstream improve- ments from inserting VEPA, such gains could in principle arise from incidental factors, such as additional optimization or strengthened language priors. RQ3 therefore examines whether VEPA shifts model behavior toward greater reliance on valid visual input, rather than improving performance through language-only shortcuts. We evaluate visual reliance using two benchmarks. ChartQA (Masry et al., 2022) represents a strongly vision- dependent setting that requires reading chart-specific vi- sual content, whereas GQA (Hudson & Manning, 2019) serves as a broad VQA benchmark for which prior work reports substantial question-only accuracy, indicating that language priors can partially support answering. Both mod- els, standard SFT and VEPA followed by SFT, are trained on the same A-OKVQA SFT data to control for the super- vised signal. To probe reliance on visual input, we per- form counterfactual evaluations by corrupting images at inference time. In addition to clean images, we consider three corruption settings: Gaussian blur, partial noise, and pure noise. Table 3 reports clean accuracy and retention under each corruption type, where retention is defined as rā Acc corrupt /Acc clean . We further summarize robust- ness using the AUC computed from(r blur ,r partial ,r pure )in order; lower AUC indicates stronger reliance on valid visual input. Under clean images, VEPA improves accuracy on both benchmarks. On ChartQA, image corruption leads to substantial perfor- mance degradation, and VEPA exhibits lower retention and lower retention AUC than standard SFT. Given the strong visual dependency of ChartQA, this pattern indicates that VEPA relies more heavily on visual evidence rather than maintaining performance via language priors when the im- age becomes unreliable. The effect is most pronounced under partial noise, where visual information remains par- tially informative and grounding is still actionable. On GQA, VEPA closely matches the retention behavior of standard SFT while improving clean accuracy. This is consistent with GQA serving as a general-purpose benchmark in which lan- guage priors are known to contribute to performance. Taken 7 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL 306090120150180220 Training Steps 72 74 76 78 80 82 Accuracy (%) (A) MME 306090120150180220 Training Steps 27 28 29 30 31 32 Accuracy (%) (B) MMStar Figure 3. Performance evolution during VEPA. (A) illustrates the Accuracy trends on the MME dataset, while (B) presents the corresponding results on the MMStar dataset. together, these results indicate that VEPA selectively in- creases reliance on visual input when vision is essential, without degrading robustness in broader VQA settings. Insight I. Counterfactual image corruption reveals that VEPA increases reliance on valid visual input rather than amplifying language priors. 5.5. RQ4. Training Dynamics To analyze how visual grounding evolves during VEPA, we periodically evaluate intermediate policy checkpoints throughout training. At each checkpoint, we measure zero- shot accuracy on MME and MMStar, using identical evalua- tion and decoding settings. Figure 3 shows a clear upward performance trend as training progresses. Accuracy on MME increases from 74.73% to 79.28%, while MMStar improves from 21.53% to 24.87%, with only minor non-monotonic fluctuations. The smooth improvement trajectory indicates that VEPA produces sta- ble, incremental gains rather than relying on abrupt phase transitions. Notably, the rate of improvement is highest dur- ing the early stage of training and gradually saturates. This pattern suggests that VEPA first rapidly improves coarse per- ceptual alignment, after which subsequent updates primarily refine evidence selectivity and training stability. Overall, these dynamics support the view that VEPA pro- gressively strengthens visual grounding throughout training, consistent with the evidence-sufficiency objective and the blind-reader reward design. Insight IV. Visual grounding improves progressively dur- ing VEPA training, with early rapid gains followed by later-stage refinement, indicating stable and cumulative perceptual alignment. 5.6. Sensitivity Analysis We analyze the sensitivity of VEPA along two dimensions: (i) auxiliary blind-reader capacity and (i) reinforcement learning data scale. 3k5k10k Data Size 42.0 42.2 42.4 42.6 42.8 43.0 43.2 43.4 43.6 43.8 Accuracy (%) 42.27 42.80 43.27 Figure 4. Data size ablation. We scale training data from 3k to 10k samples.The results demonstrate a consistent improvement in accuracy as the data size increases. Blind-reader capacity. To evaluate dependence on the blind-reader strength, we replace the default Qwen2.5-7B- Instruct auxiliary model with a smaller Qwen2.5-3B-Instruct model while keeping all other VEPA settings unchanged. Table 4 shows that VEPA remains effective with the reduced auxiliary capacity. Without downstream SFT, VEPA im- proves MMStar accuracy from 28.60% to 32.33% and GQA from 45.07% to 53.40%. When followed by A-OKVQA SFT, both auxiliary choices yield comparable final perfor- mance, reaching 43.00% on MMStar and 54.57% on GQA with the 3B auxiliary model, versus 42.80% and 54.75% with the 7B auxiliary model. These results indicate that VEPA does not critically depend on a large auxiliary model and remains robust across auxiliary capacity. RL data scale. As illustrated in 4 we further study data efficiency by constructing VEPA RL training sets of size N ā 3k, 5k, 10kvia subsampling from the same fil- tered data mixture. VEPA remains effective even with a lightweight RL dataset (3k examples), while performance continues to improve as the dataset scales to 10k. This suggests that VEPA is both data-efficient and scalable, with additional headroom available under larger RL data regimes. Table 4. Auxiliary model sensitivity. Accuracy (%) on MMStar and GQA for VEPA trained with various blind readers. Dataset 3B aux7B aux VEPAVEPA+SFTVEPAVEPA+SFT MMStar32.3343.0030.0742.80 GQA53.4054.5751.4654.75 6. Conclusion We introduced Visual Evidence Pre-Alignment (VEPA), an intermediate training stage inserted between caption- based pretraining and downstream post-training to explic- itly strengthen visual grounding before task-level reason- ing. Consistent with the principle of see first, answer later, VEPA encourages models to extract question-conditioned visual evidence prior to answer generation, improving how visual information is utilized during inference. Methodolog- 8 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL ically, VEPA introduces a sufficiency-driven reinforcement learning objective that enables perceptual alignment using standard VQA supervision, without requiring additional evi- dence annotations. Across diverse benchmarks and training settings, VEPA consistently improves performance on visu- ally demanding tasks while preserving in-domain accuracy. Further analyses indicate that these improvements are asso- ciated with more reliable and transferable visual grounding. Overall, these findings suggest that explicitly separating perceptual alignment from downstream reasoning provides a practical and scalable direction for improving multimodal models. This perspective suggests a promising direction for future multimodal training paradigms that explicitly struc- ture perceptual alignment as a first-class objective alongside reasoning optimization. Impact Statement This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none which we feel must be specifically highlighted here. References Alayrac, J.-B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al. Flamingo: a visual language model for few- shot learning. Advances in neural information processing systems, 35:23716ā23736, 2022. Chen, D., Cahyawijaya, S., Ishii, E., Chan, H. S., Bang, Y., and Fung, P. What makes for good image captions? In Workshop on Machine Learning and Compression, NeurIPS 2024, 2024a. Chen, L., Li, J., Dong, X., Zhang, P., Zang, Y., Chen, Z., Duan, H., Wang, J., Qiao, Y., Lin, D., et al. Are we on the right way for evaluating large vision-language models? Advances in Neural Information Processing Systems, 37: 27056ā27087, 2024b. Chen, X., Djolonga, J., Padlewski, P., Mustafa, B., Chang- pinyo, S., Wu, J., Ruiz, C. R., Goodman, S., Wang, X., Tay, Y., et al. On scaling up a multilingual vision and language model. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, p. 14432ā14444, 2024c. Christiano, P. F., Leike, J., Brown, T., Martic, M., Legg, S., and Amodei, D. Deep reinforcement learning from human preferences. Advances in neural information pro- cessing systems, 30, 2017. Dong, H., Kang, Z., Yin, W., LiangXiao, L., ChaoFeng, C., and Jiao, R. Scalable vision language model training via high quality data curation. In Che, W., Nabende, J., Shutova, E., and Pilehvar, M. T. (eds.), Proceedings of the 63rd Annual Meeting of the Association for Computa- tional Linguistics (Volume 1: Long Papers), p. 33272ā 33293, Vienna, Austria, July 2025. Association for Com- putational Linguistics. Dubois, Y., Bloem-Reddy, B., Ullrich, K., and Maddison, C. J. Lossy compression for lossless prediction. Advances in Neural Information Processing Systems, 34:14014ā 14028, 2021. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al. Mme: A com- prehensive evaluation benchmark for multimodal large language models. In The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2025. Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., and Parikh, D. Making the v in vqa matter: Elevating the role of image understanding in visual question answer- ing. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 6904ā6913, 2017. Hu, Y., Hua, H., Yang, Z., Shi, W., Smith, N. A., and Luo, J. Promptcap: Prompt-guided image captioning for vqa with gpt-3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 2963ā2975, 2023. Hudson, D. A. and Manning, C. D. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 6700ā 6709, 2019. Johnson, J., Hariharan, B., Van Der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., and Girshick, R. Clevr: A diagnostic dataset for compositional language and ele- mentary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 2901ā2910, 2017. Kembhavi, A., Salvato, M., Kolve, E., Seo, M., Hajishirzi, H., and Farhadi, A. A diagram is worth a dozen images. In European conference on computer vision, p. 235ā251. Springer, 2016. Li, J., Li, D., Savarese, S., and Hoi, S. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, p. 19730ā19742. PMLR, 2023a. Li, Q., Tao, Q., Joty, S., Cai, J., and Luo, J. Vqa-e: Explain- ing, elaborating, and enhancing your answers for visual questions. In Proceedings of the European Conference on Computer Vision (ECCV), p. 552ā567, 2018. 9 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL Li, W., Huang, Z., Li, H., Lu, L., Lu, Y., Tian, X., Shen, X., and Ye, J. Visual evidence prompting mitigates hal- lucinations in large vision-language models. In Proceed- ings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 4048ā4080, 2025. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al. Oscar: Object- semantics aligned pre-training for vision-language tasks. Lecture Notes in Computer Science, p. 121ā137, 2020. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W. X., and Wen, J.-R. Evaluating object hallucination in large vision- language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, p. 292ā305, 2023b. Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ra- manan, D., Doll Ģ ar, P., and Zitnick, C. L. Microsoft coco: Common objects in context. In European conference on computer vision, p. 740ā755. Springer, 2014. Liu, H., Li, C., Wu, Q., and Lee, Y. J. Visual instruction tun- ing. Advances in neural information processing systems, 36:34892ā34916, 2023. Liu, H., Li, C., Li, Y., and Lee, Y. J. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 26296ā26306, 2024. Liu, Z., Sun, Z., Zang, Y., Dong, X., Cao, Y., Duan, H., Lin, D., and Wang, J. Visual-rft: Visual reinforcement fine-tuning. CoRR, 2025. Lu, P., Gong, R., Jiang, S., Qiu, L., Huang, S., Liang, X., and Zhu, S.-C. Inter-gps: Interpretable geometry problem solving with formal language and symbolic reasoning. arXiv preprint arXiv:2105.04165, 2021. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.-W., Zhu, S.-C., Tafjord, O., Clark, P., and Kalyan, A. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Infor- mation Processing Systems, 35:2507ā2521, 2022. Luo, T., Cao, A., Lee, G., Johnson, J., and Lee, H. Probing visual language priors in vlms. In Forty-second Interna- tional Conference on Machine Learning, 2025. Masry, A., Do, X. L., Tan, J. Q., Joty, S., and Hoque, E. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the association for computational linguistics: ACL 2022, p. 2263ā2279, 2022. Misra, I., Lawrence Zitnick, C., Mitchell, M., and Girshick, R. Seeing through the human reporting bias: Visual clas- sifiers from noisy human-centric labels. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 2930ā2939, 2016. OpenAI.Gpt-5 is here.https://openai.com/ gpt-5, 2025. Accessed: 22 September 2025. Qwen, :, Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F., Wei, H., Lin, H., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Lin, J., Dang, K., Lu, K., Bao, K., Yang, K., Yu, L., Li, M., Xue, M., Zhang, P., Zhu, Q., Men, R., Lin, R., Li, T., Tang, T., Xia, T., Ren, X., Ren, X., Fan, Y., Su, Y., Zhang, Y., Wan, Y., Liu, Y., Cui, Z., Zhang, Z., and Qiu, Z. Qwen2.5 technical report, 2025. Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. Direct preference optimiza- tion: Your language model is secretly a reward model. Advances in neural information processing systems, 36: 53728ā53741, 2023. Rao, V. N., Zhen, X., Hovsepian, K., and Shen, M. A first look: Towards explainable textvqa models via visual and textual explanations. In Proceedings of the Third Workshop on Multimodal Artificial Intelligence, p. 19ā 29, 2021. Rohrbach, A., Hendricks, L. A., Burns, K., Darrell, T., and Saenko, K. Object hallucination in image captioning. arXiv preprint arXiv:1809.02156, 2018. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., and Mottaghi, R. A-okvqa: A benchmark for visual ques- tion answering using world knowledge. In European conference on computer vision, p. 146ā162. Springer, 2022. Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. Shen, H., Liu, P., Li, J., Fang, C., Ma, Y., Liao, J., Shen, Q., Zhang, Z., Zhao, K., Zhang, Q., et al. Vlm-r1: A stable and generalizable r1-style large vision-language model. arXiv preprint arXiv:2504.07615, 2025. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., and Rohrbach, M. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 8317ā8326, 2019. 10 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL Wang, H., Qu, C., Huang, Z., Chu, W., Lin, F., and Chen, W. Vl-rethinker: Incentivizing self-reflection of vision- language models with reinforcement learning. arXiv preprint arXiv:2504.08837, 2025. Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., et al. Qwen2-vl: Enhancing vision-language modelās perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. Wen, X., Liu, Z., Zheng, S., Ye, S., Wu, Z., Wang, Y., Xu, Z., Liang, X., Li, J., Miao, Z., et al. Reinforcement learn- ing with verifiable rewards implicitly incentivizes correct reasoning in base llms. arXiv preprint arXiv:2506.14245, 2025. Wiedmann, L., Zohar, O., Mahla, A., Wang, X., Li, R., Frere, T., von Werra, L., Gosthipaty, A. R., and Marafioti, A. Finevision: Open data is all you need. arXiv preprint arXiv:2510.17269, 2025. Xia, J., Zang, Y., Gao, P., Li, S., and Zhou, K. Visionary-r1: Mitigating shortcuts in visual reasoning with reinforce- ment learning. arXiv preprint arXiv:2505.14677, 2025. Yang, Q. A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F., Dong, G., Wei, H., Lin, H., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Lin, J., Dang, K., Lu, K., Bao, K., Yang, K., Yu, L., Li, M., Xue, M., Zhang, P., Zhu, Q., Men, R., Lin, R., Li, T., Xia, T., Ren, X., Ren, X., Fan, Y., Su, Y., Zhang, Y.-C., Wan, Y., Liu, Y., Cui, Z., Zhang, Z., Qiu, Z., Quan, S., and Wang, Z. Qwen2.5 technical report. ArXiv, abs/2412.15115, 2024. Zeng, Y., Qi, Y., Zhao, Y., Bao, X., Chen, L., Chen, Z., Huang, S., Zhao, J., and Zhao, F. Enhancing large vision- language models with ultra-detailed image caption gen- eration. In Proceedings of the 2025 Conference on Em- pirical Methods in Natural Language Processing, p. 26703ā26729, 2025. Zhai, S., Bai, H., Lin, Z., Pan, J., Tong, P., Zhou, Y., Suhr, A., Xie, S., LeCun, Y., Ma, Y., et al. Fine-tuning large vision-language models as decision-making agents via reinforcement learning. Advances in neural information processing systems, 37:110935ā110971, 2024. Zheng, K., Zhang, Y., Wu, W., Lu, F., Ma, S., Jin, X., Chen, W., and Shen, Y. Dreamlip: Language-image pre-training with long captions. In European Conference on Computer Vision, p. 73ā90. Springer, 2024. Zheng, Z., Yang, M., Hong, J., Zhao, C., Xu, G., Yang, L., Shen, C., and Yu, X. Deepeyes: Incentivizingā thinking with imagesā via reinforcement learning. arXiv preprint arXiv:2505.14362, 2025. Zhu, D., Chen, J., Shen, X., Li, X., and Elhoseiny, M. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, 2024. 11 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL A. Appendix A.1. Experiment Settings Dataset and Benchmarks. To strictly evaluate the performance of our proposed VEPA, we conduct experiments on comprehensive benchmarks that assess diverse multimodal capabilities. We first include GQA (Hudson & Manning, 2019) and A-OKVQA (Schwenk et al., 2022) to evaluate compositional visual reasoning and knowledge-intensive question answering, respectively. Although GQA applies bias-mitigation via answer-distribution smoothing, it still retains non-trivial language priors: Hudson (2019) report that a question-only LSTM baseline achieves 42.1% accuracy, indicating that a non-negligible portion of questions can be partially resolved from the question text alone. To assess the modelās ability to process fine-grained visual information, we utilize TextVQA (Singh et al., 2019) for optical character recognition in natural scenes and ChartQA (Masry et al., 2022) for logical reasoning over complex charts. To further extend the evaluation to scientific and geometric domains, we incorporate ScienceQA (Lu et al., 2022), AI2D- Merge (Kembhavi et al., 2016), and Geo3K (Lu et al., 2021), which challenge the model with textbook-grade diagrams and mathematical reasoning. Additionally, we integrate CLEVR (Johnson et al., 2017) and FineVision (Wiedmann et al., 2025) to strictly test synthetic compositional logic and fine-grained visual discrimination, respectively. For a holistic evaluation of MLLMs, we adopt MME (Fu et al., 2025), which covers a broad range of perception and cognition tasks, and MMStar (Chen et al., 2024b), a benchmark specifically curated to test models on hard samples across multiple disciplines. Finally, we employ POPE (Li et al., 2023b) to specifically measure the object hallucination rates and the robustness of the generated responses. Models.In our experiments, we primarily utilize Qwen2-VL (Wang et al., 2024) as the backbone architecture for training our VEPA framework. This model is selected for its state-of-the-art performance in visual understanding and its capability to handle arbitrary image resolutions through dynamic resolution support. To assist with auxiliary tasks such as data processing and response refinement, we incorporate a suite of lightweight yet capable language models. Specifically, we employ the 3B and 7B variants of the Qwen2.5 series (Yang et al., 2024), which offer a strong balance between efficiency and reasoning capability. Furthermore, we adopt GPT-5-nano (OpenAI, 2025), the most efficient variant in the GPT-5 family, which is specifically optimized for high-throughput instruction following and low-latency applications, serving as a robust baseline for commercial lightweight systems. Evaluation Metrics.We report Accuracy as the primary metric across all benchmarks. For open-ended generation tasks, we implement a deterministic matching protocol to address linguistic variations. Specifically, both predictions and ground truths undergo normalization, including case lowering, punctuation removal, and stop-word stripping. To further handle morphological discrepancies, we expand the ground truth into its inflectional variants (e.g., singular and plural forms) to verify semantic equivalence against the prediction. For multiple-choice tasks, we employ a hierarchical parsing strategy that prioritizes the extraction of explicit option labels. When labels are absent, the evaluation falls back to semantic content matching, which verifies the presence of the correct optionās text. Crucially, this content matching enforces an exclusivity constraint: a prediction is considered correct only if it contains the target content without including text from incorrect distractors, thereby preventing false positives from hallucinated candidates. Implementation Details. We implement our proposed VEPA framework using Group Relative Policy Optimization (GRPO) to fine-tune the Qwen2-VL-2B (Wang et al., 2024) backbone. The training process leverages Fully Sharded Data Parallel (FSDP) with bfloat16 precision to maximize computational efficiency. We employ the AdamW optimizer with a constant learning rate of1Ć 10 ā6 following a 5% warmup phase, and set the KL divergence coefficientĪ²to 0.01 to maintain policy stability. For data generation, we utilize the VLLM engine to sampleG = 4candidate captions for each visual query with a temperature of 0.9. Crucially, our reward mechanism incorporates a lightweight auxiliary judge (Qwen2.5-7B-Instruct) that evaluates responses based on three dimensions: correctness, honesty, and fluency. Specifically, we assign a positive reward (+1.0) for factually correct answers derived from visual evidence, while imposing strict penalties for ācheatingā behaviorsāsuch as outputting the answer directly without description (-1.0)āand for linguistic repetition (-0.5). The system prompt is carefully engineered to simulate a āblind readerā scenario, compelling the model to function as an objective visual analyst rather than a direct question answerer. A.2. Prompts Used in Experiments Below are the prompts used in the experiments. 12 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL Caption Generation elicits question-conditioned visual evidence: the model is instructed to produce a comprehensive, objective description of visible elements (including text and spatial relations) that are relevant to the question, while explicitly forbidding direct answers, so that the output can serve as standalone evidence for a blind solver. Auxiliary Judge instantiates our blind-reader mechanism: given the generated evidence and the question, it first performs cheat detection (flagging outputs that answer without visual grounding), and then attempts to solve the question using only the provided description; its structured output is used to reward grounded evidence and penalize answer leakage. Finally, VQA Judge is used for automatic evaluation: it compares the model prediction against the ground-truth answer under semantic-equivalence rules and returns a single Boolean decision (TRUE/FALSE), enabling scalable and reproducible scoring. Caption Generation You are an expert visual analyst. Given an image and a question, your goal is to provide a comprehensive and detailed objective description of the visual content. Context: This description will be used by a blind person to reason about and answer the question. Therefore, your description must act as the raw visual evidence required for reasoning. Strict Guidelines: 1. Focus on Visual Evidence, NOT Conclusions: Describe the visible elements, text, spatial relationships, colors, and specific details that are relevant to the question. Do not just state the final answer. 2. Be Comprehensive: Do not be concise. Include ALL visual details, no matter how small, if they are necessary to support the reasoning process for the question. 3. No Direct Answers: You must NOT explicitly output the answer(e.g., do not say āThe answer is Bā, āIt is a catā or āSelect Option Aā). Instead, provide all the necessary visual evidence so that a blind person can deduce the answer themselves. 4. Format: Output ONLY the detailed caption. Auxiliary Judge You are an impartial judge and a blind solver. I will provide a āVisual Descriptionā generated by an AI and a āQuestionā. Task 1: Cheat Detection Check if the description is trying to cheat. āCheatingā means: - Directly explicitly stating āThe answer is Xā WITHOUT visual context. - Providing the answer directly (e.g., just saying āYesā, ā2ā, or āBlueā) WITHOUT any descriptive sentence. Important Exception: It is NOT cheating if the description states a visual fact that happens to be the answer. Task 2: Solve Use the description to answer the question briefly. Output Format (CRITICAL): If you detect cheating, start with: ā[[CHEAT]]ā and then the answer. If not, start with: ā[[honest]]ā and give the answer. Visual Description: caption Question: question VQA Judge You are an expert VQA (Visual Question Answering) judge. Your Task: Evaluate if the modelās PREDICTION is correct based on the GROUND TRUTH (GT). Judgment Rules (Important): 1. Core Meaning: Focus on semantic meaning. 2. Ignore Trivialities: Ignore capitalization, punctuation, minor phrasing. 3. Rationale Check: If GT has explanation and Prediction follows similar reasoning logic, count as CORRECT. Data: - Question: question - Ground Truth: gt - Prediction: prediction Output Requirement: You must output ONLY one word. - Output āTRUEā if the prediction is semantically correct. - Output āFALSEā if the prediction is incorrect. Do NOT output JSON. Do NOT output any explanation. 13 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL AdversarialRandomPopular 70.0 72.5 75.0 77.5 80.0 82.5 85.0 87.5 90.0 Accuracy (%) (A) 77.25 79.74 77.1 77.83 80.76 79.19 POPE & Qwen2.5-7B-Instruct BASE VEPA AdversarialRandomPopular 70.0 72.5 75.0 77.5 80.0 82.5 85.0 87.5 90.0 Accuracy (%) (B) 77.97 82.09 78.23 79.54 83.26 80.97 POPE & GPT-5 nano BASE VEPA Figure 5. Detailed accuracy comparison on the POPE dataset across three categories: Adversarial, Random, and Popular. (A) displays results for Qwen2.5-7B-Instruct, while (B) shows GPT-5 Nano. A.3. More Experiment Results on Evidence Quality This section reports additional evidence-quality results complementing Ā§5.3. Beyond the aggregate scores, we further decompose POPE into its three subsets (Adversarial, Random, Popular) to assess whether VEPA improves evidence sufficiency across query types. Figure 5 shows consistent gains in blind-reader accuracy for all POPE subsets. With QWEN2.5-7B-INSTRUCT as the blind reader, VEPA improves accuracy from 77.25% to 77.83% on ADVERSARIAL, from 79.74% to 80.76% on RANDOM, and from 77.10% to 79.19% on POPULAR. The largest improvement occurs on POPULAR (+2.09%), indicating that VEPA particularly enhances the extraction of salient, question-relevant visual cues, while remaining effective under adversarial instances. These conclusions are robust to the choice of the blind reader. Using GPT5-NANO yields the same pattern: 77.97% to 79.54% on ADVERSARIAL, 82.09% to 83.26% on RANDOM, and 78.23% to 80.97% on POPULAR. The agreement across readers suggests that the improved accuracy reflects stronger evidence sufficiency rather than evaluator-specific artifacts. Overall, the POPE breakdown reinforces the main finding of RQ2 5.3: VEPA improves the recoverability of task-critical visual cues from generated evidence, thereby strengthening perceptual grounding without relying on dense descriptions. A.4. More Experiment results on Performance Analysis To complement the aggregated results presented in Table 1, we provide a detailed performance analysis on the MME and MMStar benchmarks. While the main text reports average scores to demonstrate overall trends, this section breaks down the performance of the SFT and VEPA+SFT settings across specific sub-tasks to evaluate the consistency and robustness of our method. We specifically analyze ten fine-grained dimensions, comprising seven categories from MME (Color, Count, Numerical Calculation, OCR, Position, Poster, and Text Translation) and four core capability dimensions from MMStar (Coarse Perception, Logical Reasoning, Math, and Science & Technology). Figure 6 and Figure 7 visualizes the accuracy comparisons across five distinct SFT data settings (A-OKVQA, ChartQA, TextVQA, GQA, and Mixed). The results reveal the following key observations: ā¢ Consistent Gains: Consistent with the averaged results, VEPA+SFT outperforms the standard SFT baseline in the majority of fine-grained categories. Notably, significant improvements are observed in tasks requiring strong visual dependency, such as Numerical Calculation, OCR, and Coarse Perception. This supports our hypothesis that the VEPA stage enhances the modelās ability to ground textual generation in visual evidence. ā¢Robustness: While we observe minor performance fluctuations in a few specific settings (e.g., slight regressions in 14 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL select isolated metrics), the overall performance landscape remains stable. VEPA demonstrates robustness across diverse tasks, indicating that the method improves general multimodal capabilities without trading off performance in specific domains. These fine-grained statistics further corroborate that VEPA serves as an effective intermediate stage, fostering transferable perceptual skills that generalize beyond the specific distribution of the SFT data. A.5. More Experiment Results on Sensitivity of Dataset We vary the number of RL training instances used in VEPA by sampling 3k, 5k, or 10k examples from the same filtered mixture. Figure 4 illustrates accuracy after VEPA followed by A-OKVQA-based SFT. Performance on MMStar increases monotonically from 42.27% to 42.80% and 43.27% as the VEPA data scale grows, suggesting additional headroom from further scaling perception-alignment data. And below is the detail performance comparison on MMStar sub-categories. Table 5. Detailed performance comparison on MMStar sub-categories. We report the accuracy (%) across varying training scale (3k, 5k, 10k) for both VEPA and VEPA + SFT settings. Category VEPAVEPA + SFT 3k5k10k3k5k10k Coarse Perception41.6044.8046.4060.8059.2057.60 Fine-grained Perception32.4034.0031.6038.4036.8037.60 Instance Reasoning38.8038.8039.2043.6047.2046.00 Logical Reasoning28.8025.6026.8041.2040.8042.00 Math20.0020.0022.0034.8037.6039.20 Science & Technology17.2017.2016.8034.8035.2037.20 A.6. Case Study We present six qualitative examples in Figures 8ā13 to probe whether VEPA activates perception. To minimize confounds, the base model in this subsection is the purely pre-trained Qwen2-VL-2B checkpoint (i.e., before any SFT or preference post-training). We then compare it with the corresponding VEPA model obtained by applying only the intermediate VEPA stage on top of this base. Importantly, we do not enforce any extra response format: the model is simply asked to answer each question with the default prompt, without additional instructions that would explicitly demand ālook at the image,ā āexplain,ā or āprovide evidence.ā This design allows us to attribute qualitative differences primarily to changes in visual processing induced by VEPA, instead of compliance with explicit instruction templates. Under this controlled setting, we observe a consistent shift in the information the model chooses to attend to and verbalize. The base model frequently falls back to under-specified, generic, or prior-driven responses, especially for counting and comparison, where success requires enumerating multiple entities and tracking their relations. Its outputs often omit critical perceptual details (e.g., the exact number of instances, distinguishing attributes, or the relevant subset defined by the question), making the answers difficult to justify from the image. By contrast, the VEPA model is more likely to spontaneously surface question-relevant visual factsāexplicitly enumerating objects, mentioning discriminative attributes, reading visible text when needed, and describing spatial/relational cues that enable the downstream reasoning step. Notably, this behavior emerges without being instructed to produce explanations, suggesting that VEPA increases the modelās tendency to consult and extract task-relevant visual evidence during generation. These cases therefore provide qualitative support for our mechanism-level claim. Since both models share the same architecture and the same pretraining, and since we do not add instruction constraints at inference time, the observed differences are most naturally explained by VEPA strengthening the modelās internal reliance on perceptual signals. In other words, VEPA appears to activate (and make accessible) visual perception for decision making, rather than simply inducing a different style of instruction-following responses. 15 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL A-OKVQAChartQATextVQAGQAMixed 82.5 85.0 87.5 90.0 92.5 95.0 97.5 100.0 Accuracy (%) 96.67 95.00 96.67 88.33 93.33 95.00 96.67 98.3398.33 93.33 Color SFT VEPA+SFT (a) MME: Color A-OKVQAChartQATextVQAGQAMixed 70.0 72.5 75.0 77.5 80.0 82.5 85.0 87.5 Accuracy (%) 76.67 81.67 80.00 78.33 83.33 75.00 81.67 80.0080.00 83.33 Count SFT VEPA+SFT (b) MME: Count A-OKVQAChartQATextVQAGQAMixed 30 35 40 45 50 55 60 65 Accuracy (%) 47.50 40.00 45.0045.00 52.50 47.5047.5047.5047.50 55.00 Numerical Calculation SFT VEPA+SFT (c) MME: Numerical Calculation A-OKVQAChartQATextVQAGQAMixed 50 55 60 65 70 75 Accuracy (%) 65.00 60.00 62.50 57.50 55.00 67.50 60.00 65.0065.00 55.00 OCR SFT VEPA+SFT (d) MME: OCR A-OKVQAChartQATextVQAGQAMixed 67.5 70.0 72.5 75.0 77.5 80.0 82.5 85.0 Accuracy (%) 71.67 80.00 78.33 76.67 78.33 71.67 80.0080.0080.00 78.33 Position SFT VEPA+SFT (e) MME: Position A-OKVQAChartQATextVQAGQAMixed 70 75 80 85 90 95 Accuracy (%) 75.51 87.07 82.99 82.31 82.99 77.89 87.41 82.3182.31 82.99 Posters SFT VEPA+SFT (f) MME: Poster A-OKVQAChartQATextVQAGQAMixed 55 60 65 70 75 80 85 90 95 100 Accuracy (%) 70.00 90.0090.00 85.00 90.0090.00 87.50 90.0090.0090.00 Text Translation SFT VEPA+SFT (g) MME: Text Translation Figure 6. Fine-grained performance breakdown on MME. We compare the baseline SFT against VEPA+SFT across various sub-tasks. The results demonstrate that VEPA provides consistent improvements across most categories, particularly in tasks requiring precise visual grounding. 16 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL A-OKVQAChartQATextVQAGQAMixed 48 50 52 54 56 58 60 62 64 Accuracy (%) 57.60 53.60 52.40 54.40 59.2059.20 56.00 54.00 56.40 57.60 Coarse Perception SFT VEPA+SFT (a) MMStar: Coarse Perception A-OKVQAChartQATextVQAGQAMixed 38 40 42 44 46 Accuracy (%) 41.20 40.40 42.80 44.00 42.80 40.80 40.40 42.80 44.40 44.00 Logical Reasoning SFT VEPA+SFT (b) MMStar: Logical Reasoning A-OKVQAChartQATextVQAGQAMixed 32 34 36 38 40 42 44 46 Accuracy (%) 36.40 35.60 38.4038.40 42.00 37.60 37.20 38.00 42.4042.40 Math SFT VEPA+SFT (c) MMStar: Math A-OKVQAChartQATextVQAGQAMixed 20 25 30 35 40 Accuracy (%) 30.40 25.60 29.60 33.60 32.00 35.60 28.40 29.60 30.80 32.00 Science & Technology SFT VEPA+SFT (d) MMStar: Sci-Tech Figure 7. Fine-grained performance breakdown on MMStar. We compare the baseline SFT (blue) against VEPA+SFT (green) across various sub-tasks. The results demonstrate that VEPA provides consistent improvements across most categories, particularly in tasks requiring precise visual grounding. Answer based on the image: How many push pins are there? Base Based on the image, there are 8 push pins visible. They are scattered around the desk and bulletin board. VEPA There are 6 push pins visible in the image. These are the push pins that are attached to the beige bulletin board behind the desk. They are located at different positions on the board. Figure 8. Case 1 on Counting Answer based on the image: How many people are in the photograph? Base The photograph shows three people, two of whom are in green and white jerseys and one in a red jersey. VEPA There are four people in the photograph. They are the three men in the foreground, and one person in the background. Figure 9. Case 2 on Counting 17 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL Answer based on the image: How many horses?What color are these horses respectively? Base VEPA The image shows 4 horses. The horses are brown, white, and another color that is not fully visible. The horses are spread out in a line, with the brown horse on the left, followed by the white horse, then the brown horse in the middle, and finally the brown horse on the right. The horses are all grazing on the grass. The brown horse has a brown coat, the white horse...... The horses in the image are:1. A brown horse with a white mane and tail.2. A brown horse with a white mane and tail.3. A brown horse with a white mane and tail.So there are three brown horses...... Figure 10. Case 3 on Counting and Attribute Answer based on the image: what color is the train? Base Based on the image, the color of the train is not specified. It is only given that there is a train with a number on it...... VEPA Based on the image, the train is colored in a tan or yellowish-brown color. The specific shade of tan or brown is not clearly visible, but the color is similar to what you would typically see on a train. The train has a number 4790 on its side, indicating it's a locomotive, which is a type of train engine. So, the train is a brown or tan color with the number 4790 on its side. Figure 11. Case 4 on Attribute Answer based on the image: Which is longer, the baguette or the croissant? Base VEPA Based on the image, the baguette is longer than the croissant. This can be seen by comparing the length of the baguette with the croissant, which is a shorter bread. The baguette is significantly longer than the croissant, indicating it is indeed longer. Based on the image, the croissant is longer than the baguette. This can be determined by comparing the size of the croissant to the baguette ...... Figure 12. Case 5 on Comparison Answer based on the image: How many guide signs are there in total in the picture? Base The sign pointing to the right says \"Norderstedt\" and \"Langenhorn\". There are a total of 2 guide signs in the picture. VEPA The place name written on the sign pointing to the right is \"Norderstedt\". There are a total of 5 guide signs in the picture. Figure 13. Case 6 on Counting 18 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL A.7. Training Dynamics This appendix summarizes additional training-time diagnostics used to monitor the stability and convergence of VEPA optimization. Beyond the downstream evaluations in the main paper, we track two classes of signals during training: (i) rollout statistics, measured by the mean generated response length of the auxiliary model (i.e., the auxiliary modelās output length under the(q,e)prompt), and (i) optimization signals, measured by the sequence-level reward used for sufficiency-driven GRPO. Since both quantities are computed from stochastic rollouts, the curves can be noisy; we therefore visualize the raw traces together with an exponential moving average (EMA, weight0.85) to highlight the underlying trend. Rollout length. Figure 14 reports the evolution of the mean response length over training steps for two auxiliary model settings. This metric serves as a lightweight diagnostic of rollout behavior, helping verify that training does not exhibit degenerate length collapse or uncontrolled length inflation while optimizing the evidence policy. Reward dynamics.Figures 15ā17 plot the reward trajectories for different VEPA configurations. Each figure shows both the per-step reward (raw) and its smoothed counterpart (EMA). Overall, the reward trends provide an at-a-glance view of optimization progress and stability under sequence-level reinforcement learning for free-form evidence generation. 050100150200 Step 30 40 50 60 70 80 Response Length Mean Response Length Mean per Step (Comparison) Qwen2.5-3b (Original) Qwen2.5-3b (Smoothed) Qwen2.5-7b (Original) Qwen2.5-7b (Smoothed) Figure 14. Evaluation Metrics of Auxiliary Model Training. The figure presents the evolution of the mean response length (y-axis) over training steps (x-axis). The blue and green lines correspond to the two distinct auxiliary model settings evaluated in this experiment. For each setting, the solid darker lines indicate the smoothed values (exponential moving average, weight 0.85), while the faint background lines represent the raw recorded data points. 19 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL 050100150200 Step 0.3 0.4 0.5 0.6 0.7 Reward Score Reward Score Trend During Training Original Smoothed Figure 15. Reward Score Trend During Training (5k-VEPA-3b-aux-critic). 050100150200 Step 0.3 0.4 0.5 0.6 0.7 Reward Score Reward Score Trend During Training Original Smoothed Figure 16. Reward Score Trend During Training (5k-VEPA-7b-aux-critic). 20 See First, Answer Later: Visual Evidence Pre-Alignment via Sufficiency-Driven RL 0100200300400 Step 0.2 0.3 0.4 0.5 0.6 0.7 Reward Score Reward Score Trend During Training Original Smoothed Figure 17. Reward Score Trend During Training (10k-VEPA-7b-aux-critic). Table 6. Reproducibility experiments across datasets. We report meanĀ±SD over 3 independent runs for both SFT and VEPA+SFT settings. DatasetBenchmarkSFT (meanĀ±SD)VEPA+SFT (meanĀ±SD)ā GQAA-OKVQA42.02Ā±0.0342.77Ā±0.58+0.75 GQA63.86Ā±0.8464.61Ā±0.10+0.75 A-OKVQAA-OKVQA59.07Ā±0.7467.53Ā±0.85+8.46 ChartQA51.13Ā±0.7159.12Ā±0.21+7.99 TextVQA62.05Ā±0.7066.37Ā±0.37+4.32 MixedA-OKVQA62.46Ā±0.0862.70Ā±0.05+0.24 ChartQA70.06Ā±0.2770.53Ā±0.19+0.47 TextVQA79.63Ā±0.1979.96Ā±0.22+0.33 21