Agent Plasticity: Measuring Self-Improvement Through Experience
The paper proposes “agent plasticity” as a way to measure how efficiently agents turn experience into better held-out performance.
The authors test self-improvement in a controlled setup where agents carry forward reusable artifacts from past interactions. They find frontier models improve in sharply different ways, even when given comparable chances to learn. Training-regime gains only partly transfer out of distribution, and the strongest final performer is not always the most efficient learner. Low-plasticity agents often fail to reuse relevant artifacts; higher-plasticity agents can still fail because the artifacts are weak, poorly generalized, or poorly applied. HF Daily Papers' note
The authors test self-improvement in a controlled setup where agents carry forward reusable artifacts from past interactions. They find frontier models improve in sharply different ways, even when given comparable chances to learn. Training-regime gains only partly transfer out of distribution, and the strongest final performer is not always the most efficient learner. Low-plasticity agents often fail to reuse relevant artifacts; higher-plasticity agents can still fail because the artifacts are weak, poorly generalized, or poorly applied. HF Daily Papers' note
score 6