Paper deep dive
Beyond Exact Match: Task-Aware GRPO for Cross-Domain PCBA Visual Question Answering
Jia Li, Li Dai, Peng Jia, Zhenzhen Hu, Chee Seng Chan, Bingkun Bao, Richang Hong
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 9/21/2026, 4:12:56 AM
Summary
This paper proposes a multimodal reasoning framework for cross-domain Printed Circuit Board Assembly (PCBA) Visual Question Answering (VQA). It addresses domain shifts between standards-derived and real-world imagery and heterogeneous output spaces (choice vs. counting) by introducing Task-Aware Group Relative Policy Optimization (GRPO). The method utilizes unified instruction formats, verified Chain-of-Thought (CoT) reasoning traces, and task-specific reward functions (semantic, distance-aware, and format rewards). The system achieves an Overall Score of 83.24 on the PCBA Standard-to-Real Grand Challenge leaderboard.
Entities (6)
Relation Signals (6)
Jia Li → affiliatedwith → Hefei University of Technology
confidence 99% · Jia Li Affiliation: Hefei University of Technology, Hefei, China
Paper → publishedat → ACM International Conference on Multimedia
confidence 98% · Proceedings of the 34th ACM International Conference on Multimedia; November 10–14, 2026
Task-Aware GRPO → usedfor → PCBA Visual Question Answering
confidence 98% · We propose a multimodal reasoning framework for cross-domain PCBA visual question answering... We further introduce Task-Aware Group Relative Policy Optimization (GRPO)
Task-Aware GRPO → achievesscoreon → PCBA Standard-to-Real Grand Challenge
confidence 97% · The proposed system achieves an Overall Score of 83.24 on the official PCBA Standard-to-Real Grand Challenge leaderboard
Task-Aware GRPO → usesrewardtype → semantic rewards
confidence 95% · integrating multi-component semantic rewards for choice-based questions
Task-Aware GRPO → usesrewardtype → distance-aware rewards
confidence 95% · distance-aware rewards for counting questions
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:In automated Printed Circuit Board Assembly (PCBA) inspection, standards-guided decisions require systems to jointly reason over fine-grained visual cues, component semantics, and manufacturing knowledge. Although large vision-language models (VLMs) provide a promising foundation, their deployment is hindered by the domain shift between standards-derived samples and real-world production-line imagery, together with heterogeneous output spaces spanning choice-based and numerical counting tasks. To address these challenges, we propose a multimodal reasoning framework for cross-domain PCBA visual question answering. The framework converts standards-derived, real-world, and auxiliary PCB-domain data into a unified instruction format and constructs verified reasoning traces aligned with visual evidence, question semantics, candidate options, and ground-truth answers. We further introduce Task-Aware Group Relative Policy Optimization (GRPO), which moves beyond exact-match supervision by integrating multi-component semantic rewards for choice-based questions, distance-aware rewards for counting questions, and an auxiliary format reward for valid outputs. During inference, answer-option semantic consistency correction, self-consistency voting, and multi-model arbitration are combined to improve prediction robustness. The proposed system achieves an Overall Score of 83.24 on the official PCBA Standard-to-Real Grand Challenge leaderboard, demonstrating the effectiveness of task-aware reward design and robust inference for cross-domain PCBA visual question answering.
Tags
Links
- Source: https://arxiv.org/abs/2609.21276v1
- Canonical: https://arxiv.org/abs/2609.21276v1
Trouble viewing inline? Open PDF directly →
Full Text
46,996 characters extracted from source content.
Expand or collapse full text
Beyond Exact Match: Task-Aware GRPO for Cross-Domain PCBA Visual Question Answering Conference: Proceedings of the 34th ACM International Conference on Multimedia; November 10–14, 2026; Rio de Janeiro, BrazilProceedings of the 34th ACM International Conference on Multimedia (M ’26), November 10–14, 2026, Rio de Janeiro, BrazilDOI: 10.1145/3767308.3836634ISBN: 979-8-4007-2213-4/2026/11CCS: Computing methodologies Artificial intelligence Jia Li Affiliation: Hefei University of Technology, Hefei, China email: jiali@hfut.edu.cn , Li Dai Affiliation: Hefei University of Technology, Hefei, China email: 2025110483@mail.hfut.edu.cn , Peng Jia Affiliation: Hefei University of Technology, Hefei, China email: 2020214631@mail.hfut.edu.cn , Zhenzhen Hu Note: Corresponding author. Affiliation: Hefei University of Technology, Intelligent Interconnected Systems Laboratory of Anhui Province, Hefei, China email: zzhu@hfut.edu.cn , Chee Seng Chan Affiliation: Universiti Malaya, Kuala Lumpur, Malaysia email: cs.chan@um.edu.my , Bingkun Bao Affiliation: Hefei University of Technology, Hefei, China email: bingkunbao@hfut.edu.cn and Richang Hong Affiliation: Hefei University of Technology, Hefei, China email: hongrc.hfut@gmail.com © c Abstract. In automated Printed Circuit Board Assembly (PCBA) inspection, standards-guided decisions require systems to jointly reason over fine-grained visual cues, component semantics, and manufacturing knowledge. Although large vision-language models (VLMs) provide a promising foundation, their deployment is hindered by the domain shift between standards-derived samples and real-world production-line imagery, together with heterogeneous output spaces spanning choice-based and numerical counting tasks. To address these challenges, we propose a multimodal reasoning framework for cross-domain PCBA visual question answering. The framework converts standards-derived, real-world, and auxiliary PCB-domain data into a unified instruction format and constructs verified reasoning traces aligned with visual evidence, question semantics, candidate options, and ground-truth answers. We further introduce Task-Aware Group Relative Policy Optimization(GRPO), which moves beyond exact-match supervision by integrating multi-component semantic rewards for choice-based questions, distance-aware rewards for counting questions, and an auxiliary format reward for valid outputs. During inference, answer–option semantic consistency correction, self-consistency voting, and multi-model arbitration are combined to improve prediction robustness. The proposed system achieves an Overall Score of 83.24 on the official PCBA Standard-to-Real Grand Challenge leaderboard, demonstrating the effectiveness of task-aware reward design and robust inference for cross-domain PCBA visual question answering. Keywords: PCB Assembly Inspection; Cross-Domain VQA; Chain-of-Thought Reasoning; Answer-Type-Specific Rewards; Multimodal Large Language Models †c-license: by Figure 1. Illustration of cross-domain PCBA visual question answering, which combines standards-derived and real-world factory data and evaluates multimodal reasoning across factuality, cause and handling, quantitative reasoning, and attribute reasoning tasks. 1. Introduction Industrial visual inspection is essential in Printed Circuit Board Assembly (PCBA), where subtle defects may cause functional failures, reliability degradation, and costly rework (Petkov and Ivanova, 2024). In Surface Mount Technology (SMT) lines, Automated Optical Inspection (AOI) is increasingly expected to support standards-aligned decisions by connecting fine-grained visual evidence with component semantics, process rules, and handling actions (Goti, 2025; Jessurun et al., 2023). Although existing industrial inspection benchmarks and PCB-oriented datasets have advanced anomaly localization, defect detection, and board-level visual understanding (Bergmann et al., 2019; Zou et al., 2022; Jiang et al., ; Fan et al., 2025; Tang et al., 2019; Lv et al., 2024; Mineo et al., 2025), most remain perception-oriented and rarely assess whether models can integrate visual observations with manufacturing knowledge to produce actionable decisions. This gap motivates a unified multimodal reasoning framework that combines visual perception, domain knowledge, and structured decision-making. To move beyond perception-centric inspection, models should be evaluated on their ability to connect visual evidence with manufacturing knowledge in decision-oriented scenarios. The PCBA Standard-to-Real Challenge formulates this problem as cross-domain visual question answering (Chen et al., 2026), combining standards-derived and real-world production-line samples across diverse inspection-oriented question families. This setting requires models to generalize from standards-derived samples to visually diverse production-line images while handling heterogeneous answer spaces and task-specific output constraints. Recent large vision-language models have demonstrated strong multimodal instruction-following and reasoning capabilities (Li et al., 2023; Liu et al., 2023; Bai et al., 2025; Chen et al., 2024a), while Chain-of-Thought(CoT) reasoning and reinforcement-based optimization have shown promising performance in complex reasoning and verifiable problem solving (Wei et al., 2022; Zhang et al., 2023; Shao et al., 2024; Liu et al., 2025). Nevertheless, directly applying general-purpose VLMs or generic reasoning optimization pipelines to PCBA VQA remains suboptimal, because valid predictions must be consistent not only with visual evidence, but also with option semantics, numerical targets, and required answer formats. To address these challenges, we propose a task-aware multimodal framework for PCBA VQA. We unify heterogeneous data sources and question types into a consistent multimodal instruction format and construct verified Chain-of-Thought data aligned with visual evidence, question semantics, candidate options, and final answers. Building on these structured data, we introduce Task-Aware GRPO with answer-type-specific rewards: multi-component semantic rewards for choice-based questions, distance-aware rewards for counting questions, and an auxiliary format reward for valid and parseable responses. During inference, semantic consistency correction, self-consistency voting, and multi-model arbitration are combined to improve robustness. Our method ranks second on the official PCBA Standard-to-Real Challenge leaderboard with an Overall Score of 83.2483.24. Our main contributions are summarized as follows: • We develop a task-structured data construction pipeline that unifies heterogeneous PCBA samples into a consistent multimodal instruction format and produces verified CoT data. • We introduce Task-Aware GRPO with semantic rewards for choice-based reasoning, distance-aware rewards for counting, and format regularization for valid outputs. • Our system ranks second on the PCBA Challenge leaderboard with an Overall Score of 83.2483.24, demonstrating its effectiveness for cross-domain manufacturing inspection VQA. Figure 2. The overall framework consists of two main stages: task-structured data construction and Task-Aware GRPO. In the first stage, standards-derived data, real-world data, and PCB-Bench data are converted into a unified multimodal instruction format, followed by task-specific CoT construction and verification. In the second stage, the policy model is optimized with GRPO using task-aware rewards, including choice-based rewards, counting rewards, and format rewards. 2. Related Work Industrial Visual Inspection and PCBA Defect Understanding. Industrial inspection has been widely studied through anomaly detection and defect localization. Benchmarks such as MVTec AD, MVTec LOCO AD, MVTec AD 2, and Real-IAD extend industrial anomaly detection beyond normal-only defect spotting to settings involving logical anomalies, realistic conditions, and diverse industrial variations (Bergmann et al., 2019; Bergmann et al., 2022; Heckler-Kram et al., ; Wang et al., 2024). Representative methods have evolved from distribution modeling and self-supervised tasks to reconstruction-, distillation-, and memory-based detectors, including PaDiM, DRAEM, PatchCore, Reverse Distillation, and EfficientAD (Defard et al., 2021; Zavrtanik et al., 2021; Roth et al., 2022; Deng and Li, 2022; Batzner et al., 2024). PCB-specific datasets and surveys, such as HRIPCB, the Scientific Data PCB defect dataset, and PCB defect detection reviews, further adapt this paradigm to board-level inspection (Huang et al., 2020; Lv et al., 2024; Chen et al., 2023). However, these works mainly predict anomaly existence, location, or category. PCBA instead requires component semantics, defect reasoning, counting, attribute understanding, and standard-guided decisions, motivating a unified multimodal reasoning framework beyond standalone defect detection. Multimodal VQA and Vision-Language Models for Industrial Reasoning. Visual question answering has progressed from generic image-question answering to structured and knowledge-intensive multimodal reasoning. VQA, GQA, OK-VQA, TextVQA, and DocVQA establish open-ended visual QA, compositional reasoning, external knowledge, reading, and grounding requirements (Antol et al., 2015; Hudson and Manning, 2019; Marino et al., 2019; Singh et al., 2019; Mathew et al., 2021). ScienceQA, MathVista, and MMMU further push multimodal models toward scientific, mathematical, and expert-level reasoning (Lu et al., 2022; Lu et al., 2024; Yue et al., 2024). Meanwhile, vision-language models have advanced from CLIP-style contrastive pretraining to BLIP-2, LLaVA, Qwen2.5-VL, and InternVL-style scaling, improving instruction following, OCR, grounding, and high-resolution perception (Radford et al., 2021; Li et al., 2023; Liu et al., 2023; Bai et al., 2025; Chen et al., 2024a). Nevertheless, industrial PCBA inspection differs from generic VQA: PCB-Bench focuses on PCB design and routing reasoning, whereas PCBA Standard-to-Real emphasizes cross-domain generalization from manufacturing standards to real inspection images with mixed answer spaces and tiny visual cues (Li et al., 2026; Chen et al., 2026). Chain-of-Thought Reasoning and Reinforcement Optimization. Recent reasoning studies show that exposing intermediate steps improves complex problem solving. Chain-of-Thought prompting, zero-shot CoT, STaR, and Tree-of-Thoughts show that explicit reasoning traces help decomposition, planning, and self-correction (Wei et al., 2022; Kojima et al., 2022; Zelikman et al., 2022; Yao et al., 2023). This idea also extends to multimodal reasoning, where Multimodal-CoT and visual CoT prompting align visual evidence with semantic decisions, and step-level verification highlights the value of supervising reasoning processes (Lightman et al., 2024; Zhang et al., 2023; Chen et al., 2024b). In parallel, large-model optimization has progressed from RLHF and PPO-based policy learning to direct preference optimization and reasoning-oriented reinforcement learning (Ouyang et al., 2022; Schulman et al., 2017; Rafailov et al., 2023). DeepSeekMath introduces GRPO for scalable verifiable reasoning, while DeepSeek-R1 and Visual-RFT show that reinforcement fine-tuning is practical for multimodal reasoning (Shao et al., 2024; Guo et al., 2025; Liu et al., 2025). However, most CoT and RL pipelines assume homogeneous targets, whereas PCBA inspection mixes multiple-choice recognition, binary defect judgment, counting, and standard-aware decisions, requiring task-aware reasoning and optimization. 3. Method 3.1. Problem Formulation and Overview Given one or more images and a natural-language question, the PCBA task requires a multimodal model to predict an answer. It involves standards-based knowledge, real-world visual recognition, attribute reasoning, and quantitative reasoning, requiring stable reasoning in complex scenarios. Based on their output formats, we categorize questions into three types: multiple-choice, binary judgment, and numerical counting. Choice and binary questions correspond to multi-class and two-class judgments and require selecting a valid option from visual evidence and question semantics; counting questions identify target objects and return an integer. Our framework consists of three components. First, to reduce format shift across heterogeneous sources and question types, we organize all samples into a unified multimodal instruction format and construct verified CoT-format reasoning data, providing a consistent training basis for supervised fine-tuning and reasoning optimization. Second, to address heterogeneous answer spaces in PCBA VQA, we introduce Task-Aware GRPO with answer-type-specific reward functions. It combines hierarchical rewards for choice-based reasoning, distance-aware rewards for counting reasoning, and format compliance rewards, enabling reinforcement optimization to jointly improve answer correctness, semantic consistency, numerical accuracy, and instruction-following ability. Third, during inference, we combine semantic consistency correction, self-consistency estimation, and multi-model arbitration to improve robustness. 3.2. Task Structured Data Construction Unified Multimodal Instruction Format. PCBA VQA contains heterogeneous question types that differ in question formulations and output constraints. Directly mixing these samples may introduce format shift and make the model overly sensitive to superficial prompt patterns. To mitigate this issue, we convert samples from different sources and task categories into a unified multimodal instruction format. For choice-based questions, the output is normalized to a single valid option letter; for counting questions, the output is normalized to an integer without additional explanations or redundant units. This design has two benefits. First, it provides a consistent training interface for standards-derived knowledge understanding, defect judgment, and quantitative reasoning. Second, it constrains the output space, thereby improving instruction following and providing a stable basis for reward design. Task Specific CoT Construction and Verification. To enhance intermediate reasoning on complex PCBA questions, we construct CoT-format reasoning data for subsequent reasoning-oriented optimization. Each CoT sample consists of a concise reasoning process and a final answer. The reasoning process is expected to focus on visual evidence, question semantics, or standards-related cues directly relevant to the current sample, while the final answer follows the required answer format. During CoT construction, we use a strong vision-language model and provide it with the image, the question, candidate options when available, and the ground-truth answer. This setting helps keep the generated reasoning process consistent with the final answer and reduces hallucinated explanations. For different answer spaces, we adopt task-specific generation instructions: choice-based samples emphasize option-semantic alignment, while counting samples emphasize numerical consistency. Furthermore, we introduce a verification step to improve the reliability of CoT data. The verification process checks whether the output follows the required CoT structure, whether the final answer matches the ground-truth answer, whether the reasoning content is grounded in the visual evidence and question semantics, and whether the reasoning conclusion is consistent with the candidate option or integer answer. Samples that fail verification are regenerated or filtered out. Through this process, we obtain reliable and parseable reasoning data for subsequent Task-Aware GRPO optimization. 3.3. Task-Aware GRPO To further improve performance on PCBA VQA tasks, we apply Group Relative Policy Optimization (GRPO) (Shao et al., 2024). Given an input sample x, the old policy πθold _ _old samples a group of responses oii=1G\o_i\_i=1^G. Each response is evaluated by a reward function, yielding a scalar reward ri=R(oi,x)r_i=R(o_i,x). The relative advantage of the i-th response is computed by normalizing rewards within the sampled group: (1) A^i=ri−mean(rjj=1G)std(rjj=1G)+ϵ, A_i= r_i-mean(\r_j\_j=1^G)std(\r_j\_j=1^G)+ε, where ϵε is a small constant for numerical stability. This group-relative normalization removes the need for a separate value model and encourages the policy to favor responses that receive higher rewards than other responses sampled for the same input. The policy is optimized with a clipped objective: (2) GRPO(θ)=x,oii=1G[1G∑i=1GSi−βDKL(πθ(⋅|x)∥πref(⋅|x))],J_GRPO(θ)=E_x,\o_i\_i=1^G [ 1G _i=1^GS_i-β D_KL ( _θ(·|x)\| _ref(·|x) ) ], where (3) Si=min(ρiA^i,clip(ρi,1−ϵc,1+ϵc)A^i),S_i= ( _i A_i,clip( _i,1- _c,1+ _c) A_i ), and (4) ρi=πθ(oi|x)πθold(oi|x). _i= _θ(o_i|x) _ _old(o_i|x). Here, ρi _i denotes the policy ratio, ϵc _c is the clipping threshold, πref _ref denotes the reference policy, and β controls the strength of KL regularization. 3.3.1. Task-Aware Reward Modeling Since PCBA VQA contains heterogeneous answer spaces, a single generic reward is insufficient for optimizing different question types with distinct evaluation criteria. We therefore design task-aware rewards for choice-based and counting tasks, together with a format reward to encourage valid and instruction-following outputs. For each sampled response oio_i, the overall reward is defined as: (5) R(oi,x)=Rtask(oi,x)+λfmtRfmt(oi,x),R(o_i,x)=R_task(o_i,x)+ _fmtR_fmt(o_i,x), where RtaskR_task denotes the task-specific reward selected according to the question type, RfmtR_fmt denotes the format reward, and λfmt _fmt controls the strength of the format constraint. RtaskR_task is defined as: (6) Rtask(oi,x)=Rchoice(oi,x),choice-based tasks,Rcount(oi,x),counting tasks.R_task(o_i,x)= casesR_choice(o_i,x),&choice-based tasks,\\ R_count(o_i,x),&counting tasks. cases Choice-Based Reward. Choice-based tasks include both multiple-choice questions and binary judgment questions. Let a^i a_i and a denote the predicted and ground-truth option labels for response oio_i, respectively. Let e^i e_i denote the semantic conclusion extracted from the reasoning trace, while eae_a and ea^ie_ a_i denote the option texts corresponding to the ground-truth label and the predicted label. We define three reward components: (7) Racc(oi,x)=1,if a^i=a, otherwise 0,Rsem(oi,x)=1,if e^i≃ea, otherwise 0,Rcon(oi,x)=1,if e^i≃ea^i, otherwise 0. \ array[]lR_acc(o_i,x)=1,&if a_i=a, otherwise 0,\\ R_sem(o_i,x)=1,&if e_i e_a, otherwise 0,\\ R_con(o_i,x)=1,&if e_i e_ a_i, otherwise 0. array . where ≃ denotes semantic equivalence. Here, RaccR_acc measures final answer correctness, RsemR_sem measures semantic alignment with the ground-truth option text, and RconR_con measures consistency between the predicted option and the reasoning conclusion. The final choice-based reward is computed as: (8) Rchoice(oi,x)=αaccRacc+αsemRsem+αconRcon,R_choice(o_i,x)= _accR_acc+ _semR_sem+ _conR_con, where αacc _acc, αsem _sem, and αcon _con are weighting coefficients. This design encourages the model to align option labels with their semantic meanings rather than relying only on superficial label patterns. Counting Reward. For counting tasks, exact-match rewards are overly sparse, since near-correct numerical predictions should be distinguished from large counting errors. Let y^i y_i and y denote the predicted and ground-truth counts for response oio_i, respectively. We define the absolute error and, when y≠0y≠ 0, the relative error as: (9) di=|y^i−y|,δi=di|y|(y≠0).d_i=| y_i-y|, _i= d_i|y| (y≠ 0). To provide smoother supervision, we assign fixed rewards to exact and near-correct predictions, while applying error-dependent decay to larger deviations. Specifically, we define absolute-error and relative-error decay terms as: (10) Rabs(di,y)=0.6exp(−di−2sabs(y)),Rrel(δi)=0.6(1−δiτrel)+2,R_abs(d_i,y)=0.6 (- d_i-2s_abs(y) ), R_rel( _i)=0.6 (1- _i _rel )_+^2, where (z)+=max(z,0)(z)_+= (z,0), sabs(y)s_abs(y) controls the absolute-error decay scale, and τrel _rel denotes the relative-error tolerance. The counting reward is then defined as: (11) Rcount(oi,x)=1.0,di=0,0.8,di=1,0.6,di=2,min(Rabs(di,y),Rrel(δi)),di>2,y≠0,Rabs(di,0),di>2,y=0.R_count(o_i,x)= cases1.0,&d_i=0,\\ 0.8,&d_i=1,\\ 0.6,&d_i=2,\\ (R_abs(d_i,y),R_rel( _i) ),&d_i>2,\ y≠ 0,\\ R_abs(d_i,0),&d_i>2,\ y=0. cases When y=0y=0, only the absolute-error decay is used to avoid undefined relative error. Compared with an exact-match reward alone, this design provides denser supervision for near-correct counts and penalizes large absolute and relative errors. By producing more informative reward differences within sampled response groups, it yields more stable relative advantages for GRPO training. Format Reward. To preserve output validity, we introduce a format reward RfmtR_fmt. This reward checks two aspects: whether the final answer conforms to the answer space specified by the input x, and whether the response follows a complete reasoning-answer structure. Let va(oi,x)∈0,1v_a(o_i,x)∈\0,1\ indicate whether the answer in response oio_i is valid, and let vs(oi)∈0,1v_s(o_i)∈\0,1\ indicate whether the reasoning-answer structure is complete. The format reward is defined as: (12) Rfmt(oi,x)=1.0,va(oi,x)=1andvs(oi)=1,0.5,va(oi,x)=1andvs(oi)=0,0.2,va(oi,x)=0andvs(oi)=1,0,va(oi,x)=0andvs(oi)=0.R_fmt(o_i,x)= cases1.0,&v_a(o_i,x)=1\ and\ v_s(o_i)=1,\\ 0.5,&v_a(o_i,x)=1\ and\ v_s(o_i)=0,\\ 0.2,&v_a(o_i,x)=0\ and\ v_s(o_i)=1,\\ 0,&v_a(o_i,x)=0\ and\ v_s(o_i)=0. cases This auxiliary reward encourages instruction-following responses. Moreover, its graded piecewise design provides non-binary feedback for partially valid outputs, which reduces reward sparsity and improves the stability of reinforcement learning optimization. 3.4. Confidence-Calibrated Robust Inference During inference, we apply a lightweight calibration strategy to reduce answer inconsistency and sampling instability. For choice-based tasks, if the reasoning conclusion is semantically aligned with one candidate option but the answer corresponds to another option, we correct the answer to the semantically matched option. We further estimate confidence through self-consistency voting. Given K sampled answers z^(k)k=1K\ z^(k)\_k=1^K, the majority-vote answer and its confidence are defined as: (13) z^∗ z^* =argmaxv|k∣z^(k)=v, 1≤k≤K|, = *arg\,max_v |\k z^(k)=v,\ 1≤ k≤ K\ |, cvote c_vote =1Kmaxv|k∣z^(k)=v, 1≤k≤K|. = 1K _v |\k z^(k)=v,\ 1≤ k≤ K\ |. If cvotec_vote exceeds a confidence threshold, we adopt z^∗ z^* as the answer. Otherwise, the sample is treated as low-confidence and further resolved by multi-model arbitration, where predictions from multiple vision-language models are aggregated by majority voting. This strategy improves inference robustness by combining semantic consistency, sampling consistency, and model-level agreement. 3.5. Training Details We use Qwen3.5-27B (Team, 2026) as the backbone model and adopt LoRA (Hu et al., 2022) for parameter-efficient fine-tuning. Training is conducted in two stages. In the first stage, supervised fine-tuning is performed on unified multimodal instruction data to align the model with PCB-domain knowledge and task formats. In the second stage, Task-Aware GRPO is applied to task-structured CoT data to further improve reasoning performance across heterogeneous PCBA VQA tasks, including choice-based and counting questions. Table 1. Comparison with the top five teams on the final PCBA Standard-to-Real Grand Challenge leaderboard. Our method is highlighted in gray and boldface. All scores are reported as percentages. Rank Submission Overall Attribute Comp. Type Count Comp. Count Pin Defect Exist. Defect Type Mount Side Standard Know. 1 syds3sxhdjs 83.82 100.00 95.08 98.96 99.55 81.25 85.25 87.50 66.00 2 Ours 83.24 100.00 90.16 98.87 98.11 82.47 84.43 87.50 66.00 3 travis111 81.33 100.00 95.08 98.99 99.52 80.85 79.51 91.67 60.00 4 gmehong 80.11 100.00 96.72 98.96 98.13 76.92 77.05 91.67 61.00 5 dkhonker 77.52 100.00 93.44 98.72 97.58 77.08 73.77 75.00 58.00 Table 2. Ablation studies on the public evaluation set, evaluating backbone scale, auxiliary data, and GRPO reward design. All scores are reported as percentages, and the best result within each ablation group is shown in boldface. Setting Overall Attribute Comp. Type Count Comp. Count Pin Defect Exist. Defect Type Mount Side Standard Know. Backbone Scale and Auxiliary Data Qwen3.5-4B 82.66 80.25 73.80 97.81 98.45 94.82 81.15 87.25 66.63 Qwen3.5-9B 83.99 98.75 83.90 97.84 97.66 93.91 77.75 84.50 68.75 Qwen3.5-27B 86.84 100.00 89.00 98.35 98.69 94.78 82.50 84.75 72.38 Qwen3.5-27B + PCB-Bench 86.93 99.50 90.10 97.55 98.89 95.09 82.05 84.50 72.69 GRPO Reward Design SFT 86.93 99.50 90.10 97.55 98.89 95.09 82.05 84.50 72.69 GRPO w/ format reward 87.01 96.75 90.10 98.02 98.19 94.91 82.65 84.75 73.25 GRPO w/ task-aware reward 88.72 97.50 92.60 98.35 98.89 95.71 86.15 88.75 73.63 Ours 88.75 97.50 92.50 98.37 98.92 95.62 86.35 89.00 73.63 4. Experiments 4.1. Dataset and Evaluation Protocol We use the PCBA SMT Dataset from the PCBA Challenge (Chen et al., 2026) and incorporate PCB-Bench (Li et al., 2026) as auxiliary data. The PCBA SMT Dataset covers four task categories: Defect Cause & Handling, Perception-Level Factuality, Quantitative Reasoning, and Attribute Reasoning. The training set contains 8,2008,200 samples, including 1,6001,600 standards-derived samples and 6,6006,600 real-world samples, while the public test set contains 8,2008,200 samples. We additionally use 2,1932,193 samples from PCB-Bench. After unified format conversion, the SFT and CoT training set contains 10,39310,393 samples, comprising 9,5189,518 choice-based questions and 875875 counting questions. For CoT construction, Qwen3.6 (Qwen Team, 2026) serves as the teacher VLM, while GPT-5.4 (OpenAI, 2026), Qwen3.6 (Qwen Team, 2026), and Gemini (Gemini Team, 2025) are used as verifier VLMs. We follow the evaluation protocol of the PCBA Challenge (Chen et al., 2026). Choice-based tasks are evaluated using accuracy. For defect-existence questions, accuracy and F1-score are averaged to account for potential class imbalance. Counting tasks are evaluated using exact accuracy and mean absolute error (MAE), where MAE is converted into a normalized higher-is-better score and averaged with accuracy. The Overall Score is computed as the unweighted mean of all question-family scores. Let StestS_test and ShiddenS_hidden denote the aggregated scores on the public test split and hidden split, respectively, and the final score is defined as Sfinal=0.8Stest+0.2ShiddenS_final=0.8S_test+0.2S_hidden. Both the public test set and hidden test set are used only for evaluation. 4.2. Experimental Setup All experiments are conducted on 44 H100 GPUs. Training follows a two-stage pipeline consisting of supervised fine-tuning and GRPO-based reasoning optimization. In the GRPO stage, the format reward weight is set to λfmt=0.1 _fmt=0.1. For choice-based rewards, the component weights are set to αacc=0.70 _acc=0.70, αsem=0.20 _sem=0.20, and αcon=0.10 _con=0.10. In the SFT stage, we set the LoRA rank to 1616, LoRA alpha to 3232, the learning rate to 5×10−55× 10^-5, and train the model for 22 epochs. In the GRPO stage, we set the LoRA rank to 88, LoRA alpha to 3232, the learning rate to 5×10−75× 10^-7, the rollout group size to G=8G=8, the sampling temperature to 1.01.0, and train the model for 11 epoch. During inference, we first perform one greedy decoding pass and then generate K=8K=8 stochastic samples for self-consistency voting. The sampling configuration uses temperature =0.5=0.5 and top-p=0.9p=0.9, with the voting confidence threshold set to 0.550.55. 4.3. Comparison with Competitors We compare our method with the top-performing teams on the PCBA Standard-to-Real Challenge test set. As shown in Table 1, our method ranks second overall with an Overall Score of 83.2483.24, only 0.580.58 points behind the first-place solution. Notably, our method achieves the best performance on Defect Existence with a score of 82.4782.47, suggesting that the proposed task-aware optimization and confidence-calibrated inference strategy improve the robustness of defect-existence recognition. We also obtain perfect performance on Attribute Reasoning with a score of 100.00100.00, competitive results on Component Counting and Pin Counting with scores of 98.8798.87 and 98.1198.11, respectively, and a tied-best score of 66.0066.00 on Standard Knowledge. These results show that our method maintains strong robustness across diverse PCBA VQA tasks, including perception-level factuality, quantitative reasoning, and standards-derived knowledge reasoning. Overall, the second-place result validates the effectiveness of our task-structured data construction, task-aware reward design, and confidence-calibrated inference strategy. 4.4. Ablation Study Backbone Scale and Auxiliary Data. We first evaluate the contributions of backbone capacity and auxiliary PCB-domain data. As shown in Table 2, scaling the backbone from 4B to 27B consistently improves the Overall Score from 82.6682.66 to 86.8486.84. The improvement is particularly evident on Component Type, Attribute Reasoning, and Standard Knowledge, suggesting that larger vision-language backbones provide stronger fine-grained visual recognition and domain knowledge modeling capabilities. The 27B model also improves Defect Type recognition, which requires distinguishing subtle local patterns under dense PCB layouts. We further introduce PCB-Bench as auxiliary PCB-domain data. Compared with 27B trained with the base data setting, adding PCB-Bench slightly improves the Overall Score from 86.8486.84 to 86.9386.93. The gains are concentrated in Component Type, Defect Existence, Pin Counting, and Standard Knowledge. This indicates that auxiliary PCB-domain data improves domain coverage, while the official PCBA training set remains the primary source of task-specific supervision. GRPO Reward Design. We then ablate the reward design used in the GRPO stage, using the SFT model as the baseline. As reported in Table 2, applying GRPO using only a simple format reward marginally improves the Overall Score from 86.9386.93 to 87.0187.01, indicating that sparse exact-match feedback provides limited supervision for heterogeneous PCBA VQA tasks. In contrast, replacing the format reward with the proposed task-aware reward increases the Overall Score to 88.7288.72, with clear gains on Component Type, Defect Type, and Mount Side. This suggests that modeling different answer spaces with task-specific rewards is beneficial for both fine-grained recognition and reasoning-oriented questions. Adding the format reward further improves the Overall Score to 88.7588.75. Although the improvement over the task-aware reward alone is modest, the full reward design achieves the best performance on counting-related metrics, Defect Type, and Mount Side, showing that format regularization helps maintain stable and parseable outputs during reinforcement optimization. 5. Conclusion This study shows that cross-domain PCBA visual question answering is challenged not only by the visual gap between standards-derived and real-world production data, but also by the mismatch between generic exact-match objectives and heterogeneous answer spaces. By combining task-structured data, verified CoT supervision, and Task-Aware GRPO, our framework provides more informative optimization signals for both choice-based and counting questions. The ablation results indicate that answer-type-specific reward modeling contributes most of the performance improvement, while format regularization provides a smaller but consistent gain by maintaining valid and parseable outputs. Together with an Overall Score of 83.24 on the PCBA Standard-to-Real Grand Challenge, these findings support the effectiveness of moving beyond exact-match supervision in industrial multimodal reasoning. More broadly, our results suggest that PCBA inspection-oriented VQA systems benefit when their optimization objectives reflect the semantic and numerical structures of the underlying decisions rather than treating all outputs as homogeneous labels. This finding further emphasizes the importance of task-aware optimization for reliable industrial multimodal reasoning systems. Acknowledgements. This work was supported in part by National Natural Science Foundation of China under Grant No.U23B2031, and in part by the Fundamental Research Funds for the Central Universities under Grant No. PA2025IISL0110. The computations were performed on the High-Performance Computing (HPC) Platform of Hefei University of Technology. References Antol et al. (2015) S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. L. Zitnick, and D. Parikh Vqa: visual question answering. In Proceedings of the IEEE international conference on computer vision, p. 2425–2433. Cited by: §2. Bai et al. (2025) S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y. Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y. Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin Qwen2.5-vl technical report. External Links: 2502.13923, Link Cited by: §1, §2. Batzner et al. (2024) K. Batzner, L. Heckler, and R. König Efficientad: accurate visual anomaly detection at millisecond-level latencies. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, p. 128–138. Cited by: §2. Bergmann et al. (2022) P. Bergmann, K. Batzner, M. Fauser, D. Sattlegger, and C. Steger Beyond dents and scratches: logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision 130 (4), p. 947–969. Cited by: §2. Bergmann et al. (2019) P. Bergmann, M. Fauser, D. Sattlegger, and C. Steger MVTec ad–a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 9592–9600. Cited by: §1, §2. Chen et al. (2026) T. Chen, R. Xu, T. Chen, L. Lin, Y. Chiu, R. Lin, F. Oppenheimer, Y. Chuang, and W. Cheng PCBA standard-to-real challenge: cross-domain vqa for real-world manufacturing inspection. In Proceedings of the ACM Multimedia 2026 Grand Challenges, Note: To appear External Links: Link Cited by: §1, §2, §4.1, §4.1. Chen et al. (2023) X. Chen, Y. Wu, X. He, and W. Ming A comprehensive review of deep learning-based pcb defect detection. IEEE Access 11, p. 139017–139038. Cited by: §2. Chen et al. (2024a) Z. Chen, W. Wang, Y. Cao, Y. Liu, Z. Gao, E. Cui, J. Zhu, S. Ye, H. Tian, Z. Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271. Cited by: §1, §2. Chen et al. (2024b) Z. Chen, Q. Zhou, Y. Shen, Y. Hong, Z. Sun, D. Gutfreund, and C. Gan Visual chain-of-thought prompting for knowledge-based visual reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38, p. 1254–1262. Cited by: §2. Defard et al. (2021) T. Defard, A. Setkov, A. Loesch, and R. Audigier Padim: a patch distribution modeling framework for anomaly detection and localization. In International conference on pattern recognition, p. 475–489. Cited by: §2. Deng and Li (2022) H. Deng and X. Li Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 9737–9746. Cited by: §2. Fan et al. (2025) L. Fan, D. Fan, Z. Hu, Y. Ding, D. Di, K. Yi, M. Pagnucco, and Y. Song Manta: a large-scale multi-view and visual-text anomaly detection dataset for tiny objects. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 25518–25527. Cited by: §1. Gemini Team (2025) Gemini Team Gemini: a family of highly capable multimodal models. External Links: 2312.11805, Link Cited by: §4.1. Goti (2025) A. Goti Automated optical inspection (aoi) based on ipc standards. Cited by: §1. Guo et al. (2025) D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Bi, et al. Deepseek-r1: incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948. Cited by: §2. [16] L. Heckler-Kram, J. Neudeck, U. Scheler, R. König, and C. Steger The mvtec ad 2 dataset: advanced scenarios for unsupervised anomaly detection. arxiv 2025. arXiv preprint arXiv:2503.21622. Cited by: §2. Hu et al. (2022) E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, et al. Lora: low-rank adaptation of large language models.. Iclr 1 (2), p. 3. Cited by: §3.5. Huang et al. (2020) W. Huang, P. Wei, M. Zhang, and H. Liu HRIPCB: a challenging dataset for pcb defects detection and classification. The Journal of Engineering 2020 (13), p. 303–309. Cited by: §2. Hudson and Manning (2019) D. A. Hudson and C. D. Manning 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. 6700–6709. Cited by: §2. Jessurun et al. (2023) N. Jessurun, O. P. Dizon-Paradis, J. Harrison, S. Ghosh, M. M. Tehranipoor, D. L. Woodard, and N. Asadizanjani FPIC: a novel semantic dataset for optical pcb assurance. ACM Journal on Emerging Technologies in Computing Systems 19 (2), p. 1–21. Cited by: §1. [21] X. Jiang, J. Li, H. Deng, Y. Liu, B. Gao, Y. Zhou, J. Li, C. Wang, and F. Zheng Mmad: a comprehensive benchmark for multimodal large language models in industrial anomaly detection, 2025. URL https://arxiv. org/abs/2410.09453. Cited by: §1. Kojima et al. (2022) T. Kojima, S. S. Gu, M. Reid, Y. Matsuo, and Y. Iwasawa Large language models are zero-shot reasoners. Advances in neural information processing systems 35, p. 22199–22213. Cited by: §2. Li et al. (2026) J. Li, L. Chen, B. Yang, J. Zhu, Y. Wang, Y. Ma, and M. Yang PCB-bench: benchmarking llms for printed circuit board placement and routing. In The Fourteenth International Conference on Learning Representations, Cited by: §2, §4.1. Li et al. (2023) J. Li, D. Li, S. Savarese, and S. Hoi Blip-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, p. 19730–19742. Cited by: §1, §2. Lightman et al. (2024) H. Lightman, V. Kosaraju, Y. Burda, H. Edwards, B. Baker, T. Lee, J. Leike, J. Schulman, I. Sutskever, and K. Cobbe Let’s verify step by step. In International Conference on Learning Representations, Vol. 2024, p. 39578–39601. Cited by: §2. Liu et al. (2023) H. Liu, C. Li, Q. Wu, and Y. J. Lee Visual instruction tuning. Advances in neural information processing systems 36, p. 34892–34916. Cited by: §1, §2. Liu et al. (2025) Z. Liu, Z. Sun, Y. Zang, X. Dong, Y. Cao, H. Duan, D. Lin, and J. Wang Visual-rft: visual reinforcement fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 2034–2044. Cited by: §1, §2. Lu et al. (2024) P. Lu, H. Bansal, T. Xia, J. Liu, C. Li, H. Hajishirzi, H. Cheng, K. Chang, M. Galley, and J. Gao Mathvista: evaluating mathematical reasoning of foundation models in visual contexts. In International Conference on Learning Representations, Vol. 2024, p. 23439–23554. Cited by: §2. Lu et al. (2022) P. Lu, S. Mishra, T. Xia, L. Qiu, K. Chang, S. Zhu, O. Tafjord, P. Clark, and A. Kalyan Learn to explain: multimodal reasoning via thought chains for science question answering. Advances in neural information processing systems 35, p. 2507–2521. Cited by: §2. Lv et al. (2024) S. Lv, B. Ouyang, Z. Deng, T. Liang, S. Jiang, K. Zhang, J. Chen, and Z. Li A dataset for deep learning based detection of printed circuit board surface defect. Scientific Data 11 (1), p. 811. Cited by: §1, §2. Marino et al. (2019) K. Marino, M. Rastegari, A. Farhadi, and R. Mottaghi Ok-vqa: a visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, p. 3195–3204. Cited by: §2. Mathew et al. (2021) M. Mathew, D. Karatzas, and C. Jawahar Docvqa: a dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, p. 2200–2209. Cited by: §2. Mineo et al. (2025) R. Mineo, A. Sorrenti, R. Faro, G. Mineo, F. Cancelliere, and A. Faro PCB-said: a low-cost camera-based dataset for few-shot smd assembly inspection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 1351–1357. Cited by: §1. OpenAI (2026) OpenAI Introducing GPT-5.4. Note: OpenAI External Links: Link Cited by: §4.1. Ouyang et al. (2022) L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems 35, p. 27730–27744. Cited by: §2. Petkov and Ivanova (2024) N. Petkov and M. Ivanova Printed circuit board and printed circuit board assembly methods for testing and visual inspection: a review. Bulletin of Electrical Engineering and Informatics 13 (4), p. 2566–2585. Cited by: §1. Qwen Team (2026) Qwen Team Qwen3.6-27B: flagship-level coding in a 27B dense model. External Links: Link Cited by: §4.1. Radford et al. (2021) A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, p. 8748–8763. Cited by: §2. Rafailov et al. (2023) R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn Direct preference optimization: your language model is secretly a reward model. Advances in neural information processing systems 36, p. 53728–53741. Cited by: §2. Roth et al. (2022) K. Roth, L. Pemula, J. Zepeda, B. Schölkopf, T. Brox, and P. Gehler Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 14318–14328. Cited by: §2. Schulman et al. (2017) J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347. Cited by: §2. Shao et al. (2024) Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. Li, Y. Wu, et al. Deepseekmath: pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300. Cited by: §1, §2, §3.3. Singh et al. (2019) A. Singh, V. Natarajan, M. Shah, Y. Jiang, X. Chen, D. Batra, D. Parikh, and M. Rohrbach Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 8317–8326. Cited by: §2. Tang et al. (2019) S. Tang, F. He, X. Huang, and J. Yang Online pcb defect detector on a new pcb defect dataset. arXiv preprint arXiv:1902.06197. Cited by: §1. Team (2026) Q. Team Qwen3.5: accelerating productivity with native multimodal agents. External Links: Link Cited by: §3.5. Wang et al. (2024) C. Wang, W. Zhu, B. Gao, Z. Gan, J. Zhang, Z. Gu, S. Qian, M. Chen, and L. Ma Real-iad: a real-world multi-view dataset for benchmarking versatile industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 22883–22892. Cited by: §2. Wei et al. (2022) J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35, p. 24824–24837. Cited by: §1, §2. Yao et al. (2023) S. Yao, D. Yu, J. Zhao, I. Shafran, T. Griffiths, Y. Cao, and K. Narasimhan Tree of thoughts: deliberate problem solving with large language models. Advances in neural information processing systems 36, p. 11809–11822. Cited by: §2. Yue et al. (2024) X. Yue, Y. Ni, K. Zhang, T. Zheng, R. Liu, G. Zhang, S. Stevens, D. Jiang, W. Ren, Y. Sun, et al. Mmmu: a massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 9556–9567. Cited by: §2. Zavrtanik et al. (2021) V. Zavrtanik, M. Kristan, and D. Skočaj Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, p. 8330–8339. Cited by: §2. Zelikman et al. (2022) E. Zelikman, Y. Wu, J. Mu, and N. Goodman Star: bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems 35, p. 15476–15488. Cited by: §2. Zhang et al. (2023) Z. Zhang, A. Zhang, M. Li, H. Zhao, G. Karypis, and A. Smola Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923. Cited by: §1, §2. Zou et al. (2022) Y. Zou, J. Jeong, L. Pemula, D. Zhang, and O. Dabeer Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. In European conference on computer vision, p. 392–408. Cited by: §1.