Understanding Generalization and Forgetting in In-Context Continual Learning

ORID 68AMoK2YNk · tags icml2026-repro paper-68AMoK2YNk

#StatusPageArtifactClaim excerpt
1VERIFIED 2/201-decomposes-task-t-prediction-error-irreducibleartifactTheorem 4.3 decomposes the task-t prediction error into an irreducible error ter…
2VERIFIED 2/202-forgetting-arises-attention-reweighting-pastartifactTheorem 4.4 shows that forgetting arises from attention reweighting past tasks w…
3VERIFIED 2/203-theory-predicts-non-monotonic-error-curvesartifactThe theory predicts non-monotonic error curves with context length M, where long…
4VERIFIED 2/204-mean-misalignment-interference-persists-regardleartifactMean-misalignment interference persists regardless of context length, while vari…
5VERIFIED 2/205-experiments-gpt-2-architectures-confirm-predicteartifactExperiments on GPT-2 architectures confirm the predicted non-monotonic error cur…

Open logbook index · logbook.json