New paper proposes 'agent plasticity' to measure whether self-improving agents learn

An arXiv paper introduces agent plasticity, a measure of how efficiently agents turn past experience into gains on held-out tasks.

If your agents write notes, skills or memory for future runs, a paper posted to arXiv on October 6, 2026 offers a way to check whether that loop pays off. Agent Plasticity: Measuring Self-Improvement Through Experience defines plasticity as the efficiency with which an agent converts experience into gains on future held-out performance, with learning cost counted. A post by elvis (@omarsar0) attributes the work to Meta Superintelligence Labs. We could not confirm that affiliation in the sources.

In the setup, agents distill past interactions into reusable artifacts that later instances inherit. An explainer from CCTest lists these artifacts as tools, skills, memory and textual knowledge. Performance is measured at each checkpoint on both training and held-out interactions.

The abstract reports three findings:

The abstract also describes how the loop fails. Low-plasticity agents often failed to reuse relevant artifacts they had available. More plastic agents sometimes reused the right artifacts and still failed.

The X post names specific models and dollar figures. The abstract does not, and we have not verified those numbers.

Much is still unclear: the exact formula, whether cost is measured in dollars or tokens, and the size of the transfer drop. We found no released code or evaluation harness.

One practitioner has already proposed a lightweight version for an existing agent. It tracks held-out gain per 1,000 learning tokens, plus the share of injected skills the agent actually used. The author warns that a single cycle on a small hold-out set is noise.

Sources