LLM Paper Radar: Automated Tracking of Cutting-Edge Research on LLM Inference Optimization

LLM Paper Radar is an automated paper tracking tool focused on the field of LLM inference optimization. It aims to solve the dilemma of researchers screening high-value papers in the era of information overload. It scans the latest arXiv papers daily, uses AI-driven screening mechanisms and structured summary generation to help users quickly grasp field dynamics and improve information acquisition efficiency.

The LLM field is developing rapidly, with dozens of related papers published on arXiv daily. Traditional methods like manual browsing and RSS subscriptions struggle to handle the massive amount of information. Researchers urgently need tools for intelligent screening and in-depth interpretation, and LLM Paper Radar was created to address this problem.

LLM Paper Radar is maintained by the AMD Zhaolin team and is an open-source tool. Its core positioning is "Daily LLM Inference Optimization Paper Summaries". Its workflow is fully automated—from paper crawling, relevance scoring to summary generation—without manual intervention, providing researchers with screened paper briefings daily.

Data Collection Layer

Crawls paper metadata (title, abstract, authors, etc.) from the cs.CL category of arXiv via API.

Intelligent Screening Layer

Establishes a multi-level scoring mechanism; only papers with a relevance score ≥7 proceed to the next step (e.g., on May 12, 2026, 97 papers were scanned and only 3 passed).

Summary Generation Layer

Uses Claude Sonnet 4.6 to generate structured summaries including research objectives, methodological innovations, experimental results, and practical significance.

Output Presentation Layer

Outputs in Markdown format, including metadata such as arXiv ID, date, authors, tags, links, and community feedback.

Topic Focus and Precise Screening

Identifies high-value directions, such as structured pruning and knowledge distillation (Qwen3-Next compression case), KV cache intelligent eviction, dynamic inference, and speculative decoding.

Community Signal Integration

Integrates likes and comment counts from Hugging Face Daily Papers as quality references.

Historical Tracking and Indexing

Maintains a complete paper index (INDEX.md) to support tracing the development of topics.

• Automated intelligence collection: Replaces manual work for time-consuming information aggregation, improving efficiency.
• Value of intelligent screening: Strict standards (3 out of 97) ensure output quality and reduce information overload.
• Standardization of structured summaries: A unified framework facilitates paper comparison and quick understanding; AI enables large-scale standardization.

• Researchers: Daily briefings maintain field sensitivity and help discover relevant progress in a timely manner.
• Engineers and architects: Understand the latest optimization technologies, evaluate applicability, and avoid reinventing the wheel.
• Technical decision-makers: Track technical trends and assist in formulating technical roadmaps and investment decisions.

Current Limitations

AI summaries may miss details; screening thresholds have judgment biases; key papers require reading the original text.

Future Expansion Directions

Multi-source data integration (conference papers, OpenReview, etc.), personalized recommendations, in-depth analysis (code evaluation, reproducibility checks), interactive exploration (topic clustering, citation networks).