From: Can Temporary Context Become Stable Parameters?

Can you design a benchmark that separates genuine retention from prompt-conditioned reconstruction?

Many evaluations test performance while leaving the relevant context available, which cannot distinguish memory from computation. A removal-and-recall protocol makes persistence experimentally visible.

Action

Test performance immediately, after context deletion, after interference, and after a long delay using matched held-out probes.

Read the full exploration

What else is in this exploration

3 evidence blocks3 perspectives2 visualizations8 media resources4 rabbit holes

evidence

Large-language-model parameters are static and read-only during inference.

evidence

For standard transformer architectures, the effect of in-context learning cannot be converted exa...

evidence

In-context learning can adapt a model's output during inference without explicit gradient updates...

Continue exploring

Can Temporary Context Become Stable Parameters?

Evidence, perspectives, rabbit holes, and more