RISED
A method designed for multi-environment agent training that addresses data selection and reward signal failure in joint multi-environment training. RISED employs a rubrics-driven sample selection mechanism combined with self-distillation to improve training stability and effectiveness when an agent is trained across diverse environments simultaneously.
Timeline (last 90 days)
RISED stands for Rubrics for Agentic Multi-Environment Selection and Self-Distillation.
RISED's Rubrics mechanism introduces more structured evaluation criteria to guide multi-environment sample selection, refining the selection granularity to the prompt-group level.
RISED introduces a self-distillation mechanism that allows the model to extract supervision signals from its own high-quality trajectories, providing an additional learning source for samples that cannot be learned through group-relative rewards.
RISED's self-distillation mechanism follows the same line of thinking as iterative self-improvement methods such as STaR and ReST, with the core idea of guiding the model's own progress using its successful examples.
Currently available public materials only disclose the motivation and overall framework of the RISED method; specific algorithmic implementation details, experimental environment configurations, and quantitative performance improvements have not yet been released.
All Facts (5)
RISED stands for Rubrics for Agentic Multi-Environment Selection and Self-Distillation.
50%UnverifiedRISED's Rubrics mechanism introduces more structured evaluation criteria to guide multi-environment sample selection, refining the selection granularity to the prompt-group level.
50%UnverifiedRISED introduces a self-distillation mechanism that allows the model to extract supervision signals from its own high-quality trajectories, providing an additional learning source for samples that cannot be learned through group-relative rewards.
50%UnverifiedRISED's self-distillation mechanism follows the same line of thinking as iterative self-improvement methods such as STaR and ReST, with the core idea of guiding the model's own progress using its successful examples.
50%UnverifiedCurrently available public materials only disclose the motivation and overall framework of the RISED method; specific algorithmic implementation details, experimental environment configurations, and quantitative performance improvements have not yet been released.
50%