Research

Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization

Internalizes document knowledge for retrieval-free question answering while recovering general model capabilities.

Result: Across eight dataset-model settings, domain QA accuracy improves by an average 3.6 percentage points over Vanilla SFT.

All four reported metrics improve in seven of eight settings across CC and CCI and four model families. Mean general performance across IFEval, MMLU, and MSBench improves by 12.1 percentage points on average. The Inject, Align, and Recover stages separate knowledge injection, QA alignment, and general-capability recovery; some alternative methods win individual general metrics.

Wnuan: Staged Post-Training for Question Answering over Proprietary Enterprise Knowledge

Presents a three-stage post-training pipeline for question answering over proprietary enterprise documents and reports both domain results and general-capability trade-offs.

Result: On the 707-question WnuanBench, acceptable-answer rate rises from 52.76% before adaptation to 91.51% after RL; the general-benchmark average decreases by 5.17 points.

ChartWalker: Benchmarking the Cross-Chart RAG Task with Hierarchical Knowledge Graphs

Presents a framework and benchmark for cross-chart retrieval-augmented generation, together with an agentic baseline for the task.

RAFT: Data Refinement and Adaptive Distillation for Domain Fine-Tuning with Alleviated Forgetting

Describes a two-stage domain fine-tuning framework that combines model-compatible supervision construction with answer-conditioned on-policy distillation.

Result: Across three backbones and five domains, average domain accuracy improves by 23.2% over standard SFT, with relative improvements of 18.2% on MS-Bench and 10.2% on IFEval.

MechVQA: Benchmarking and Enhancing Multimodal LLMs on Comprehensive Mechanical Drawing Understanding

Introduces the MechVQA mechanical-drawing benchmark and the MechVL domain-specialized multimodal model, with public code, benchmark data, and model resources.

Result: MechVL-4B-RL scores 84.85 on MechVQA Total, 7.57 points above Gemini-3-Pro-Preview (77.28) in the paper’s evaluation.

MechVQA contains 3.3K drawings and 21K QA pairs. Table 2 of the paper reports Gemini-3-Pro-Preview as the strongest evaluated closed-source baseline. Total is the benchmark total score, distinct from the unweighted average over ten subtasks.

Rethinking Supervised Fine-Tuning: Emphasizing Key Answer Tokens for Improved LLM Accuracy

Proposes a two-stage supervised fine-tuning scheme that adds an optimization stage focused on key answer tokens after conventional fine-tuning.

SPAR: Scholar Paper Retrieval with LLM-based Agents for Enhanced Academic Search

Introduces a multi-agent framework for scholarly retrieval and the SPARBench evaluation dataset with expert-reviewed relevance labels.

Result: F1 reaches 0.3843 on AutoScholar and 0.3015 on SPARBench: relative gains of 56.9% and 23.0% over the best baselines in Table 1.

The respective baselines are PaSa (F1 0.2449) and Google-ArXiv (F1 0.2451). Relative gain is (SPAR F1 − baseline F1) / baseline F1; the absolute F1 differences are 0.1394 and 0.0564. Systems use their native search sources rather than a shared retrieval corpus.

SciSage: A Multi-Agent Framework for High-Quality Scientific Survey Generation

Introduces a multi-agent scientific survey generation framework and SurveyScope, an evaluation resource covering computer-science survey topics.

MoSLD: An Extremely Parameter-Efficient Mixture-of-Shared LoRAs for Multi-Task Learning

Proposes a mixture-of-shared-LoRAs architecture for parameter-efficient multi-task learning.

CareBot: A Pioneering Full-Process Open-Source Medical Language Model

Documents an open medical language model developed across continued pre-training, supervised fine-tuning, and preference alignment stages.

CCI3.0-HQ: a large-scale Chinese dataset of high quality designed for pre-training large language models

Describes a 500GB Chinese pre-training corpus produced with a two-stage filtering pipeline and released with data-quality resources.

Aquila-Med LLM: Pioneering Full-Process Open-Source Medical Language Models

Describes a bilingual medical language model, domain training data, and training stages spanning continued pre-training, supervised fine-tuning, and preference optimization.

Image2Katex Formula Image Recognition

Implements image-to-LaTeX recognition for printed and handwritten mathematical formulas.

CHINESE-OCR Natural-Scene Chinese Text Detection and Recognition

Implements a historical natural-scene Chinese text detection and recognition pipeline using CTPN, CRNN, and CTC-based components.