The digital preservation of ancient Chinese literature requires robust capabilities spanning layout analysis and text recognition. This paper presents a comprehensive framework addressing two fundamental challenges: (1) Layout Element Analysis (Task B) for detecting page elements (text, image, book_edge, seal) amidst degradation, nested structures, and extreme class imbalance; and (2) Text Recognition (Tasks A & C) for end-to-end transcription of printed and handwritten classical documents. For layout analysis, we propose a dual-modality solution. The Closed Modality formulates this as a sequence-to-sequence problem using Vision-Language Models (VLMs), introducing spatial discretization tokenization and a Frequency-Aware Sequential Curriculum Learning framework with dynamic memory replay. The Open Modality presents HistLayout-DETR, a set prediction architecture integrating an Augmented Morphological Encoder and a Polygon Boundary Refinement head. For text recognition, we formulate OCR as a domain-constrained visual language generation task using Qwen2.5-VL with LoRA fine-tuning. We employ structured prompts encoding reading order and Traditional Chinese character preservation across domains. Extensive experiments on the EvaHan 2026 dataset validate our framework’s superiority. In layout analysis, our curriculum-guided paradigm achieves a Macro F1 of 0.7992 and mAP@[.5:.95] of 0.5438. In text recognition, we achieve CERs of 0.0271 on printed and 0.0433 on handwritten texts.

While large language models (LLMs) have shown excellent capabilities in language understanding, text generation and many other tasks, they still struggle in complex multi-step reasoning problems such as mathematical reasoning. In this paper, through a newly proposed arithmetical puzzle problem, we show that the model can perform well on multi-step reasoning tasks via fine tuning on high-quality synthetic data. Experiments with the open-llama-3B model on three different test datasets show that not only the model can reach a zero-shot pass@1 at 0.44 on the in-domain dataset, it also demonstrates certain generalization capabilities on the out-of-domain datasets. Specifically, this paper has designed two out-of-domain datasets in the form of extending the numerical range and the composing components of the arithmetical puzzle problem separately. The fine-tuned model have shown encouraging performance on these two far more difficult tasks with the zero-shot pass@1 at 0.33 and 0.35 correspondingly.