Today's AI Research: Fastest-Growing Projects — May 18, 2026
This week, the AI Research space continues to see significant activity across various subdomains, with a particular focus on multimodal learning and autonomous agents. The growth of repositories like PaperGuru-Benchmark underscores the importance of evaluating long-horizon capabilities in large language models (LLMs) through rigorous benchmarks.
PaperGuru-AI's PaperGuru-Benchmark repository stands out this week with a growth score of 17.15, accumulating 217 stars. The project focuses on developing lifecycle-aware memory for LLM agents to enhance their long-term reasoning abilities, achieving high scores in both the PaperBench and SurveyBench evaluations. Its impressive peer-reviewed acceptances at prestigious conferences like FSE/ICML/TOSEM/AEI/ICoGB highlight its significance in advancing research methodologies.
fkyah3's opencode-yg is another repository that has gained traction this week, with a growth score of 12.48 and 37 stars. This project presents a research fork of the original opencode to demonstrate language anchoring, which ensures LLMs consistently think in a specified language—in this case, Chinese—with over 95% compliance verified through rigorous testing.
Matrix-agent's awesome-agentic-world-modeling repository has a growth score of 6.50 and 216 stars. This project provides foundational resources and insights into agentic world modeling, covering aspects like capabilities, laws, and beyond for developing intelligent agents capable of understanding complex environments and making informed decisions.
AutoMedBench's AutoMedBench repository is growing steadily with a score of 5.87 and 26 stars. The initiative aims to establish benchmarks for medical autonomous research, enabling AI agents to autonomously conduct research in the medical domain efficiently and effectively.
huangrh99's AlphaGRPO project has garnered attention this week, boasting a growth score of 5.17 and 50 stars. This work presents an official implementation from ICML2026 that unlocks self-reflective multimodal generation within unified multimodal models through decompositional verifiable reward mechanisms.
XIAO4579's PRISM repository has seen considerable activity, with a growth score of 4.04 and 74 stars. The project explores pre-alignment techniques in black-box on-policy distillation for multimodal reinforcement learning (RL), moving beyond simple skill transfer to integrate diverse modalities more effectively.
Hedlen's Awesome-Multimodal-Intelligence repository has garnered 2.61 growth score points with 46 stars, showcasing a comprehensive collection of resources related to multimodal intelligence research. This includes VLMs, VLAs, world models, and embodied AI, focusing on the next generation of intelligent agents from perception to decision-making.
kokolerk's TCOD project has grown with a score of 1.80 and 39 stars. The repository explores temporal curriculum in on-policy distillation for multi-turn autonomous agents, aiming to enhance their learning efficiency through structured training schedules.
RockeyCoss's LeapAlign_Code repository, growing steadily at a rate of 1.79 points with 35 stars, introduces LeapAlign, a method that post-trains flow matching models at any generation step by constructing two-step trajectories for improved model robustness and versatility.
Finally, earleensarellano35823414097's WorpGPT-Latest-2026-AllPrompts repository has seen minor growth with a score of 1.38 points and 40 stars. This project focuses on developing comprehensive red-teaming frameworks to test LLM robustness against adversarial prompt engineering and jailbreak vectors, ensuring the security and reliability of AI models in real-world applications.
These repositories collectively reflect the dynamic nature of AI research, showcasing advancements in multimodal learning, autonomous agent capabilities, medical research automation, and model robustness testing.
PaperGuru-AI's PaperGuru-Benchmark repository stands out this week with a growth score of 17.15, accumulating 217 stars. The project focuses on developing lifecycle-aware memory for LLM agents to enhance their long-term reasoning abilities, achieving high scores in both the PaperBench and SurveyBench evaluations. Its impressive peer-reviewed acceptances at prestigious conferences like FSE/ICML/TOSEM/AEI/ICoGB highlight its significance in advancing research methodologies.
fkyah3's opencode-yg is another repository that has gained traction this week, with a growth score of 12.48 and 37 stars. This project presents a research fork of the original opencode to demonstrate language anchoring, which ensures LLMs consistently think in a specified language—in this case, Chinese—with over 95% compliance verified through rigorous testing.
Matrix-agent's awesome-agentic-world-modeling repository has a growth score of 6.50 and 216 stars. This project provides foundational resources and insights into agentic world modeling, covering aspects like capabilities, laws, and beyond for developing intelligent agents capable of understanding complex environments and making informed decisions.
AutoMedBench's AutoMedBench repository is growing steadily with a score of 5.87 and 26 stars. The initiative aims to establish benchmarks for medical autonomous research, enabling AI agents to autonomously conduct research in the medical domain efficiently and effectively.
huangrh99's AlphaGRPO project has garnered attention this week, boasting a growth score of 5.17 and 50 stars. This work presents an official implementation from ICML2026 that unlocks self-reflective multimodal generation within unified multimodal models through decompositional verifiable reward mechanisms.
XIAO4579's PRISM repository has seen considerable activity, with a growth score of 4.04 and 74 stars. The project explores pre-alignment techniques in black-box on-policy distillation for multimodal reinforcement learning (RL), moving beyond simple skill transfer to integrate diverse modalities more effectively.
Hedlen's Awesome-Multimodal-Intelligence repository has garnered 2.61 growth score points with 46 stars, showcasing a comprehensive collection of resources related to multimodal intelligence research. This includes VLMs, VLAs, world models, and embodied AI, focusing on the next generation of intelligent agents from perception to decision-making.
kokolerk's TCOD project has grown with a score of 1.80 and 39 stars. The repository explores temporal curriculum in on-policy distillation for multi-turn autonomous agents, aiming to enhance their learning efficiency through structured training schedules.
RockeyCoss's LeapAlign_Code repository, growing steadily at a rate of 1.79 points with 35 stars, introduces LeapAlign, a method that post-trains flow matching models at any generation step by constructing two-step trajectories for improved model robustness and versatility.
Finally, earleensarellano35823414097's WorpGPT-Latest-2026-AllPrompts repository has seen minor growth with a score of 1.38 points and 40 stars. This project focuses on developing comprehensive red-teaming frameworks to test LLM robustness against adversarial prompt engineering and jailbreak vectors, ensuring the security and reliability of AI models in real-world applications.
These repositories collectively reflect the dynamic nature of AI research, showcasing advancements in multimodal learning, autonomous agent capabilities, medical research automation, and model robustness testing.