中文

This repository is dedicated to using AI to optimize papers, making it easier for researchers to check the strengths, weaknesses, and improvement suggestions of manuscripts.

Quick use (web): please visit Online Demo. The following figure is a usage preview:

This repository will be updated continuously. Welcome to use this repository, raise issues, and submit pull requests to help improve the prompt templates and website features, thereby helping the community improve paper quality and acceptance probability.

Please see: Review example of "Deep Residual Learning for Image Recognition".

• Reverse Prompt Engineering (Included): Infer prompts from excellent results generated by large models. The excellent results here come from the AI review at AAAI2026.
• Few-Shot Prompting (Included): In the web UI, select "Prompt + Examples (Few-Shot)" under "Prompting mode". The system will load templates from Prompts/ and provide the examples Examples/review_in_Resnet.md and Examples/review_in_Verified.md as examples to the model to improve results. See few-shot prompting for background.
• Chain-of-Thought Prompting (Planned): Allow the model to explicitly perform step-by-step reasoning before answering to improve accuracy and logicality for complex tasks.

• Use VLM to accurately extract content from PDF (Included): The accuracy of PDF content extraction directly affects review effectiveness, so a VLM-based accurate PDF content extraction function has been integrated. This function has been added to the web version (click "Accurately extract PDF content" on the website). It is optional. Note this feature may be unstable; if extraction fails, disable the accurate extraction and the original method will still work.
• Scientific image quality evaluation (Planned): Plan to use VLM to evaluate images in PDFs and give improvement suggestions to help users better optimize figures.

See here. We welcome everyone to provide feedback and help optimize these prompts to better serve the community.

• [08/12/2025] Deployed the online demo to Cloudflare; the previous GitHub Pages version (available on branches) is no longer used.
• [16/11/2025] Added the function to accurately extract PDF content using VLM.
• [21/10/2025] Added Few-Shot Prompting to the web UI, see here.
• [18/10/2025] Updated review example samples, see here. These examples were generated using Few-Shot Prompting techniques.
• [06/10/2025] The web version added a Quick Try feature (no API required) and allows users to set Deepseek's API.
• [02/10/2025] Updated the web AI review functionality.
• [20/09/2025] Strengthened math symbol and formula checks in the prompt.
• [14/09/2025] Added the review example for "Deep Residual Learning for Image Recognition".
• [14/09/2025] Optimized the prompt so the AI generates more detailed Strengths, Weaknesses, and Suggestions as a secondary list.