# Reinforcement Learning Optimization

MMaDA is a novel class of multimodal diffusion foundation models, excelling in various domains such as text reasoning, multimodal understanding, and text-to-image generation.

ReasonGen-R1 is an autoregressive image generation model that integrates chain-of-thought reasoning. It enhances the logic and quality of image generation through SFT and RL.

The Thinkless framework is a learnable framework that enables large models to adaptively choose between short reasoning or long-chain reasoning based on task complexity and their own capabilities.

UI-R1-E-3B is an efficient GUI positioning model fine-tuned based on Qwen2.5-VL-3B-Instruct, specializing in visual question-answering tasks, particularly excelling at locating and identifying operational elements in user interface screenshots.

Llama-3.1-Nemotron-Nano-8B-v1 is an inference model based on Meta Llama-3.1-8B-Instruct, enhanced through post-training to improve reasoning capabilities, human chat preferences, and task execution.

An improved multimodal vision-language model based on Qwen2.5-VL-72B-Instruct, excelling in multiple visual reasoning benchmarks

A vision-language model optimized based on Qwen2.5-VL-72B-Instruct, excelling in multiple visual reasoning benchmarks

Meta Llama 3.1 series multilingual large language model, featuring 8B parameters, optimized for multilingual conversational use cases, supporting 8 languages.

RM-R1 is a training framework for reasoning reward models (ReasRM), which evaluates candidate answers by generating scoring criteria or reasoning traces, providing explainable judgments.

A medical reasoning model fine-tuned based on Qwen/Qwen2.5-7B-Instruct, excelling in multiple medical QA benchmarks

Skywork-VL-Reward-7B is a 7B-parameter multimodal reward model based on the Qwen2.5-VL-7B-Instruct architecture, enhanced with a value head structure for training reward models.

A code-reasoning large language model fine-tuned based on DeepSeek-R1-Distilled-Qwen-1.5B, utilizing distributed reinforcement learning technology to extend long-context processing capabilities

TinyLLaVA-Video-R1 is a small-scale video reasoning model based on the traceable training model TinyLLaVA-Video. It significantly enhances reasoning and thinking abilities through reinforcement learning and exhibits the emergent property of 'epiphany moments'.

DeepCoder-14B-Preview is a code generation model developed based on DeepSeek-R1-Distill-Qwen-14B, focusing on solving verifiable programming problems.

A code reasoning large model fine-tuned based on DeepSeek-R1-Distilled-Qwen-1.5B, utilizing distributed reinforcement learning technology to enhance long-context processing capabilities

Quasar-3.0-7B is the distilled version of the upcoming 400B Quasar 3.0 model, showcasing the early strength and potential of the Quasar architecture.

Quasar-3.0-Max is a 7B parameter distilled model provided by SILX INC, showcasing the early potential of the Quasar architecture with innovative TTM training process and reinforcement learning techniques.

VARGPT-v1.1 is a visual autoregressive unified large model, enhanced through iterative instruction tuning and reinforcement learning, capable of performing both visual understanding and generation tasks.

VARGPT-v1.1 is a vision autoregressive unified large model enhanced through iterative instruction tuning and reinforcement learning, supporting vision understanding and generation tasks.

UI-R1 is a vision-language model enhanced by reinforcement learning for GUI agent action prediction, built upon Qwen2.5-VL-3B-Instruct.

R1-AQA is an audio question answering model based on Qwen2-Audio-7B-Instruct, optimized through Group Relative Policy Optimization (GRPO) algorithm, achieving state-of-the-art performance in the MMAU benchmark.

Light-R1-14B-DS is a 14B-parameter math SOTA model trained with reinforcement learning, excelling in AIME24/25 and GPQA benchmarks.

The first multimodal reasoning model to reproduce 'Aha moments' and increased response length on just a 2B model with unsupervised fine-tuning

TRL is a reinforcement learning library based on the Transformer architecture for training and fine-tuning language models.

A vision-language model based on Qwen2.5-VL-3B-Instruct, enhanced with VLM-R1 reinforcement learning, focusing on referring expression comprehension tasks.

A text-to-graph information extraction model based on Qwen-2.5-0.5B, jointly trained through reinforcement learning (GRPO) and supervised learning.

CycleResearcher is an automated research system based on reinforcement learning and iterative feedback, specifically trained for machine learning research, covering fields such as computer vision and natural language processing.

This is a generative model specifically designed for search query rewriting, employing a sequence-to-sequence architecture and reinforcement learning framework to produce diverse and relevant query rewrites.

This model is a merge of AIDC-ai-business/Luban-13B and Open-Orca/OpenOrca-Platypus2-13B, forming a 13-billion-parameter large language model based on the Llama 2 architecture.

This is a reinforcement learning agent based on the PPO algorithm, specifically trained for the LunarLanderContinuous-v2 environment, capable of controlling the lunar lander for smooth landing.

Korean dialogue summarization model based on BART architecture, trained by the 'Alaggung Dalaggung' team in the 2021 Hunminjeongeum Korean Speech & Natural Language AI Competition