idk you already covered it or not, I’m proposing that the current edit mechanism—which relies on score-based rollouts to accept or reject changes—could be enhanced by specifically prioritizing or weighting the inclusion of high-density semantic 'leading words.'
Since SkillOpt aims to reach an optimum in text space, it might be beneficial to guide the optimizer to favor terminology that carries strong, well-defined associations for the target model. This could help the system converge faster or arrive at more 'steerable' skills that require less overall token count to achieve the same task performance. Perhaps we could introduce a 'semantic density' heuristic into the reward function to encourage the emergence of these high-influence phrases during the training loop.
idk you already covered it or not, I’m proposing that the current edit mechanism—which relies on score-based rollouts to accept or reject changes—could be enhanced by specifically prioritizing or weighting the inclusion of high-density semantic 'leading words.'
Since SkillOpt aims to reach an optimum in text space, it might be beneficial to guide the optimizer to favor terminology that carries strong, well-defined associations for the target model. This could help the system converge faster or arrive at more 'steerable' skills that require less overall token count to achieve the same task performance. Perhaps we could introduce a 'semantic density' heuristic into the reward function to encourage the emergence of these high-influence phrases during the training loop.