output postprocessing

To address the error-prone and inefficient manual rewriting of data transformation logic upon JSON Schema evolution, this paper proposes a type-directed, top-down program synthesis approach for automatically generating semantics-preserving JSON Schema converters. Our method integrates type inference, semantic constraint modeling, a rewrite system, and intermediate representation (IR)-driven code generation to guarantee lossless data transformation and formal verifiability. It natively supports complex nested schemas and synthesizes correct, efficient, and human-readable Python and JavaScript conversion code. We evaluate our approach on real-world API configuration schemas and healthcare data integration scenarios, demonstrating its safety—via formal guarantees and empirical validation—its practical utility in industrial settings, and its generalizability across diverse schema evolution patterns. Experimental results confirm high accuracy, robustness to structural changes (e.g., field additions, type refinements, nested object restructuring), and scalability to large, deeply nested schemas.

To address the challenge of acquiring invoice data under stringent privacy regulations and high manual annotation costs, this paper proposes an end-to-end synthetic data generation framework. First, OCR precisely extracts textual content and spatial layout from real invoices. Second, a large language model generates semantically coherent and format-compliant replacement text. Third, image inpainting techniques render the new text at original positions with consistent font, size, and styling, yielding photorealistic, layout-aligned synthetic invoice images alongside corresponding JSON-structured annotations. This is the first method to jointly preserve layout fidelity, typographic consistency, and semantic validity in synthetic invoice generation. By eliminating reliance on sensitive real-world data, it enables scalable, diverse, and fully annotated invoice dataset augmentation—thereby significantly improving the robustness and accuracy of document intelligence models.

This study investigates the robustness of small language models (SLMs) in generating structured outputs for open-ended attribute-value extraction from clinical notes—particularly under privacy-sensitive conditions where output serialization format critically affects parseability. Method: We systematically evaluate JSON, YAML, and XML parsing success rates across multiple SLM sizes using fine-grained prompt engineering on real-world clinical documents, analyzing impacts of model scale, prompt design, and document length/type on output stability. Contribution/Results: JSON achieves significantly higher parsing success than YAML or XML. Increasing model parameters and optimizing prompts improve robustness, yet challenges persist for long documents and specific note types (e.g., operative reports). This work identifies, for the first time, characteristic failure modes of serialization formats in clinical NLP and proposes evidence-based guidelines for format selection and prompt optimization tailored to privacy-preserving environments—providing empirical foundations for deploying lightweight medical LMs in practice.

This work addresses the low compliance and lack of systematic evaluation of constrained decoding techniques under realistic, complex constraints—particularly JSON Schema. We introduce JSONSchemaBench, the first large-scale structured generation benchmark comprising 10K real-world JSON schemas, and propose a multidimensional evaluation framework assessing compliance, constraint coverage, and output quality across six state-of-the-art methods. Our analysis reveals, for the first time, that compliance rates drop by over 40% for existing frameworks when handling nested, recursive, and conditional constraints. We further identify XGrammar and Outlines as achieving the best trade-off between inference efficiency and generation quality. The benchmark and evaluation suite are fully open-sourced, filling a critical gap in systematic evaluation for structured generation and advancing constrained decoding toward higher reliability and stronger generalization.

Pretraining data faces a “data wall”: high-quality text is scarce, and conventional filtering discards over 99% of web-crawled data. Method: This paper proposes REWIRE—a novel framework that systematically recovers low-quality, previously filtered web text and transforms it into high-fidelity training samples via large language model–guided controllable rewriting, balancing factual accuracy and lexical/semantic diversity. REWIRE integrates multi-stage quality assessment, hybrid data-ratio optimization, and DCLM—a multi-scale evaluation protocol. Contribution/Results: On 1B/3B/7B models, REWIRE achieves average gains of 1.0–2.5 percentage points across 22 downstream tasks. Using only original + rewritten data, it outperforms models trained on twice the volume of raw web data. Notably, 82% of newly generated tokens originate from originally discarded documents, effectively alleviating both data scale and quality bottlenecks.

This work demonstrates that structured output formats—such as JSON and XML—significantly degrade the accuracy of open-source large language models on reasoning and writing tasks, primarily because format instructions interfere with the models’ natural reasoning processes. To address this “format tax,” the authors propose a general strategy that decouples reasoning from formatting, either by first generating content freely and then reformatting it, or by integrating an expanded chain-of-thought approach within a single generation pass. Experiments across six open-source models, four structured formats, and diverse task types show that this method substantially recovers lost performance and narrows the gap with closed-source counterparts. Notably, state-of-the-art closed-source models exhibit minimal sensitivity to such formatting constraints.

This study investigates the impact of output format on the performance of large language models in single-turn code generation tasks and its interaction with model identity. Through controlled experiments across four open-source projects, three prominent models—Doubao, DeepSeek, and Qwen—were evaluated using three output formats: JSON Patch, unified diff, and full file, resulting in 4,013 test cases. The findings reveal that output format significantly affects success rates, with no universally optimal format; instead, each model exhibits distinct format preferences—for instance, Doubao achieves a 94% success rate with JSON Patch, while DeepSeek performs best (66%) with unified diff. This work is the first to demonstrate a strong interaction effect between output format and model identity, proposing model-specific output strategies and design principles for tools aimed at preventing format misuse.

This study addresses the challenge of natural language–driven simulation model discovery by systematically investigating the impact of data representation, Transformer-based embedding models, and reranking strategies on retrieval performance. By constructing multimodal model metadata and leveraging standard information retrieval metrics, the work presents the first quantitative evaluation of open-source embedding models for this task. Experimental results demonstrate that the proposed approach achieves strong performance in recall@5 and nDCG@5, with reranking substantially enhancing effectiveness on complex queries. These contributions establish the first benchmark framework for AI-enabled model reusability, composability, and interoperability in simulation model retrieval.

Existing benchmarks inadequately assess the quality of structured outputs from large language models in multi-source scenarios, as they focus either on structural compliance or value correctness within a single modality. This work proposes the first cross-modal, source-agnostic evaluation framework for structured generation, which uniformly converts inputs from text, images (OCR-processed PDFs), and audio (AMI meeting transcripts) into textual contexts, constrains model outputs via JSON Schema, and constructs a complex, realistic dataset through multi-hop question answering. Evaluation of 21 state-of-the-art models across seven metrics reveals near-perfect structural compliance but markedly lower value accuracy—83.0%, 67.2%, and 23.7% for text, image, and audio sources, respectively—highlighting that extracting structured information from long, multi-source contexts remains a significant challenge.

This work addresses the lack of intrinsic validation mechanisms for extraction fidelity in existing intelligent document processing (IDP) systems, which often leads to silent error propagation. The authors propose “Reconstruction as Verification” (RaV-IDP), a novel framework that introduces reconstruction as a core component of the document processing pipeline. Specifically, a dedicated reconstructor renders extracted entities back into visual form and compares them against the original document regions to produce an unsupervised fidelity score. The approach integrates a vision-enhanced GPT-4.1 fallback strategy, a fidelity-threshold-triggered verification mechanism, and a multi-stage evaluation benchmark to enable quantifiable assessment of extraction quality. Experimental results demonstrate that RaV-IDP effectively filters low-quality outputs and significantly enhances the robustness of downstream tasks. The code is publicly released.