#GRPO

'Agent0 co-evolves a curriculum agent and an executor from the same base LLM, using sandboxed Python tool calls and ambiguity-aware RL to improve math and general reasoning without external data.'

'Meta's DreamGym synthesizes environment interactions as text using a reasoning experience model and grounded replay memory, cutting real rollouts and boosting RL performance across web benchmarks.'

'Gelato-30B-A3B converts screenshots and text instructions into precise click coordinates, achieving state of the art results on GUI grounding benchmarks and improving end-to-end agent success over GTA1-32B.'

'Early Experience uses agent-generated future states as supervision to train language agents without rewards or large demo sets, outperforming imitation learning across eight benchmarks and improving RL initialization.'

'Alibaba open-sourced Tongyi DeepResearch-30B-A3B, a 30B-parameter MoE LLM tailored for long-horizon, tool-augmented web research with 128K context and strong benchmark scores.'

'R-Zero is a co-evolutionary framework where a Challenger creates hard problems and a Solver learns from them, enabling zero-data training that boosts reasoning accuracy across benchmarks.'

'Rubrics as Rewards (RaR) introduces a reinforcement learning approach that uses structured rubrics as reward signals, improving language model training in complex domains like medicine and science.'

MMSearch-R1 introduces a reinforcement learning framework that enables large multimodal models to perform efficient, on-demand searches by learning when and how to retrieve relevant information, significantly improving accuracy and reducing search overhead.

Meta and NYU developed a semi-online reinforcement learning method that balances offline and online training to enhance large language model alignment, boosting performance in both instruction-based and mathematical tasks.

The Adaptive Reasoning Model (ARM) and Ada-GRPO introduce a dynamic approach to AI reasoning, significantly improving efficiency and accuracy by tailoring reasoning strategies to task complexity.

QwenLong-L1 introduces a structured reinforcement learning approach enabling large language models to excel at long-context reasoning tasks, achieving state-of-the-art results on multiple benchmarks.

RLV introduces a unified framework that integrates verification into value-free reinforcement learning for language models, significantly improving reasoning accuracy and computational efficiency on mathematical reasoning benchmarks.

PrimeIntellect has launched INTELLECT-2, a 32-billion parameter reasoning model trained using decentralized asynchronous reinforcement learning, achieving superior performance and open-sourcing the full codebase.

'Alibaba’s ZeroSearch framework leverages reinforcement learning and simulated document generation to train language models for retrieval without relying on costly real-time search APIs, achieving performance comparable to or better than Google Search.'

USC researchers introduce Tina, a family of compact reasoning models that leverage LoRA and reinforcement learning to deliver strong multi-step reasoning performance at a fraction of typical training costs.