Tag
Google DeepMind introduces SkillSmith, which treats model weights as an additional modality that LLMs can natively reason over, enabling instruction-steered parametric synthesis for composing skills at inference time. The approach outperforms text-only and weight-only baselines.
This paper introduces SkillSmith, an LLM augmented to reason over both prefix weights and textual knowledge, enabling instruction-steered synthesis of new parametric skills that outperform text-only and weight-only baselines.
This paper introduces AdaPrefix-GRPO, a method that adaptively controls the length of correct solution prefixes provided to a model during GRPO training, maintaining a 50% success rate to maximize gradient signal. It significantly improves accuracy on hard math reasoning problems while reducing computational cost.