MDSKILL
A multi-view distillation method for self-evolving agents that splits skill memory into three views—state perception, action planning and execution control—and uses On-Policy Self-Distillation to internalize complementary experience into shared policy parameters.
Timeline (last 90 days)
The paper "MDSKILL: Learning to Internalize Skill Memory through Multi-View Distillation for Efficient Dynamic Agent Self-Evolution" proposes a multi-view distillation scheme to address the noise and entanglement problems in the process of internalizing skill memory.
MDSKILL splits skill memory by function into three complementary views: state perception, action planning, and execution control.
MDSKILL uses dynamic metrics to filter high-quality skills, builds view-specific reasoning branches plus an integrated reasoning branch, and uses OPSD to transfer and consolidate complementary experiential knowledge across branches.
MDSKILL enables agents to complete skill internalization and continuous self-evolution without a stronger external teacher model.
MDSKILL is evaluated on three agent benchmarks, compared against pure RL methods, skill-augmented methods, and self-distillation baselines that do not rely on a stronger external teacher model.
Without adding extra computational overhead, MDSKILL improves SOTA performance on ALFWorld and WebShop by up to 7.1% and 8.1%, respectively.
MDSKILL's code has been made public in an anonymous repository (anonymous.4open.science/r/MDSKILL-CF53).
MDSKILL proposes a "lighter" self-evolution path: instead of relying on a stronger external teacher or requiring extra compute, it improves distillation efficiency by reorganizing the structure of existing skill memory.
Whether the three-view division applies to more types of tasks and how stable the dynamic filtering metrics are in more complex environments remain to be verified by future work.
All Facts (9)
The paper "MDSKILL: Learning to Internalize Skill Memory through Multi-View Distillation for Efficient Dynamic Agent Self-Evolution" proposes a multi-view distillation scheme to address the noise and entanglement problems in the process of internalizing skill memory.
50%UnverifiedMDSKILL splits skill memory by function into three complementary views: state perception, action planning, and execution control.
50%UnverifiedMDSKILL uses dynamic metrics to filter high-quality skills, builds view-specific reasoning branches plus an integrated reasoning branch, and uses OPSD to transfer and consolidate complementary experiential knowledge across branches.
50%UnverifiedMDSKILL enables agents to complete skill internalization and continuous self-evolution without a stronger external teacher model.
50%UnverifiedMDSKILL is evaluated on three agent benchmarks, compared against pure RL methods, skill-augmented methods, and self-distillation baselines that do not rely on a stronger external teacher model.
50%UnverifiedWithout adding extra computational overhead, MDSKILL improves SOTA performance on ALFWorld and WebShop by up to 7.1% and 8.1%, respectively.
50%UnverifiedMDSKILL's code has been made public in an anonymous repository (anonymous.4open.science/r/MDSKILL-CF53).
50%UnverifiedMDSKILL proposes a "lighter" self-evolution path: instead of relying on a stronger external teacher or requiring extra compute, it improves distillation efficiency by reorganizing the structure of existing skill memory.
50%UnverifiedWhether the three-view division applies to more types of tasks and how stable the dynamic filtering metrics are in more complex environments remain to be verified by future work.
50%