Announcement

Agent Plasticity is proposed to measure how efficiently AI agents improve on held-out environments

Researchers study agents that turn experience into reusable artifacts; they say the top performer need not learn most efficiently.

18 Sources, 1d ago, first seen 1d ago

TLDR

Researchers propose Agent Plasticity as held-out performance gain per unit of learning cost, meant to measure how efficiently agents learn from experience without weight updates. In tests across Chess, Go and Hex, they report Claude Fable 5 reached the highest fitted performance while GPT-5.6 Sol learned about five times more efficiently. They also found that frequently reusing learned artifacts did not guarantee improvement.

Combined views

16.1K

18 Sources, first seen 1d ago

52 likes12 comments38 saves95 reposts

Announcement

Agent Plasticity is proposed to measure how efficiently AI agents improve on held-out environments

Researchers study agents that turn experience into reusable artifacts; they say the top performer need not learn most efficiently.

18 Sources, 1d ago, first seen 1d ago

TLDR

Today's Rank

—

Not ranked yet

Combined views

16.1K

18 Sources, first seen 1d ago

52 likes12 comments38 saves95 reposts

Today's AI benchmarks measure what an agent has learned. The researchers behind Agent Plasticity want to measure something different: how efficiently an agent can turn experience into reusable tools or memory, then improve on unseen tasks without changing its model weights.

Featured Source

Researcher Anirudh Goyal described Agent Plasticity as the gain in held-out performance per unit of learning cost, measured until performance saturates. Across Chess, Go and Hex, the team reports Claude Fable 5 reached the highest fitted performance, while GPT-5.6 Sol learned roughly five times more efficiently — evidence, they argue, that raw capability and learning efficiency are not the same.

Reuse alone did not guarantee better results. In the team's account, four models used learned artifacts in 94% to 98% of relevant decisions but improved by sharply different amounts. Among stronger improvers, 83% to 99% of remaining failures occurred while an artifact was already in use.

One chess example shows how the process can alter later behavior. After hitting a 300-ply limit and scoring zero, Claude Fable 5 changed its engine to value draws near that limit; at a later checkpoint, the researchers say it held a draw until Stockfish erred, then delivered mate.

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.