Reinforcement Learning (RL)

Google has successfully demonstrated a technique to cultivate magic states with much higher fidelity on its Willow quantum chip. This advancement addresses the massive resource overhead typically required for error-corrected quantum computations.

MAI-Gym is a new open-source platform that unifies the evaluation and reinforcement learning training of GUI agents across multiple operating environments. It addresses the challenges of scalability and task diversity by enabling automated application synthesis and massive parallel execution.

This technical analysis investigates whether reasoning tokens in advanced LLMs act as computational state or simple linguistic sequences. It highlights the growing discrepancy between visible chain-of-thought traces and the actual internal logic of frontier models.

Researchers have developed a new fine-tuning technique called ConfSFT that enhances the efficiency of reasoning models by teaching them to predict their own confidence. This approach reduces computational costs by up to 25% without needing explicit instructions to shorten reasoning traces or specialized stopping mechanisms at inference time.

Researchers have developed FlyBy, a framework that allows small language models to overcome knowledge limitations by selectively querying stronger external models. This approach optimizes test-time computation by distinguishing between problems solvable through more thought and those requiring new information.

The researchers present RIDE, a technique that improves language model distillation by extrapolating RL-induced residuals in the hidden-state representation space rather than the logit space. This approach avoids the noise and attenuation caused by the language-model head, allowing student models to match or surpass their RL-trained teachers.

The article presents a new method called 'Retrieve-for-Train' to solve the computational bottleneck in complex AI search tasks. By using offline reinforcement learning and diffusion models, it enables efficient query decomposition without the need for expensive test-time computation.

This technical paper introduces the T1 reinforcement learning method to solve stability issues in long-horizon agent tasks. It specifically focuses on mitigating token drift and routing instabilities in large-scale Mixture-of-Experts models.

The G-CARL framework addresses the challenge of medical factuality in automated report interpretation by verifying individual claims against authoritative medical knowledge. It utilizes a dual-verification process and clinician-weighted checklists to ensure responses are both medically accurate and contextually relevant.

This technical excerpt provides a formal mathematical proof for the stability and convergence properties of the Co-RL framework. It specifically shows how collaborative dynamics enlarge the basin of attraction for correct outcomes in multi-agent reinforcement learning.

This research presents OPD2, a novel distillation technique that focuses on the learning trajectory of reasoning models by using a delta signal between teacher and base models. Experimental results prove it outperforms standard on-policy distillation across multiple complex reasoning domains.

The Seed framework addresses the limitations of sparse rewards in agentic reinforcement learning by extracting reusable skills from completed interaction trajectories. It utilizes on-policy distillation to provide dense, token-level supervision that evolves alongside the policy's capabilities.

This technical paper proposes a new distillation paradigm called Direct-OPD to improve weak-to-strong generalization in reasoning models. By distilling only the policy shift induced by reinforcement learning, the method avoids the capacity limitations inherent in traditional teacher-student imitation.