Meta paper introduces "agent plasticity": best-performing agent isn't the most efficient learner

rohanpaul_ai · x · 2026-10-18

A new paper from Meta, UC Berkeley, Princeton and others proposes agent plasticity — how efficiently an agent converts experience into gains in future held-out performance — as a missing dimension in agent evaluation.

Most benchmarks measure what a model can do once, ignoring whether agents improve by building reusable tools, notes and skills from past attempts, and at what cost
With weights frozen, agents played chess, Go, Hex and NetHack, rewriting their own tools and notes each round; plasticity is scored as held-out gain per dollar of learning cost
In hard chess, Claude Fable 5 rose from 37.5% to 73.3%; but across chess/Go/Hex, GPT-5.6 Sol gained 298 points per $1,000 of learning cost vs. 57 for Claude Fable 5
Failure analysis shows low-plasticity agents often fail to reuse relevant artifacts, while plastic ones can still fail on artifact quality, generalization or application

Takeaway: evaluating self-improving agents requires tracking held-out learning curves, cost, and whether saved artifacts actually get used.

Original post →

Meta paper introduces "agent plasticity": best-performing agent isn't the most efficient learner

More from coding & agent