When Should an In-Context Learner Expand Its Hypothesis Space?

Key points

• We treat this as a costly sequential decision: prediction failure must be turned into structural evidence, evidence into a value of expansion, and value into action.1
• Its solution shows that revision is a value boundary and not an evidence threshold: one history has different optimal actions under different prices, horizons and announced queries, the boundary between local repair and expansion is set by the inputs a rule predicts and a repair cannot cover, and belief in the richer family crosses long before the decision does.1
• Language models of three post-training lineages carry a failure-sensitive signal in their predictions that is not reflected in their revision decisions, and given the gain of expanding they read it without weighing it against price and horizon.1
• Controlled post-training of the meta-trained learners moves the prior and the sharpness of predictions, and neither moves the criterion.1

Sources (1)

• [1]When Should an In-Context Learner Expand Its Hypothesis Space?
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 7, 05:26 AM

  Learning systems adapt quickly inside a familiar family of models.

  We treat this as a costly sequential decision: prediction failure must be turned into structural evidence, evidence into a value of expansion, and value into action.

Extractive summary: sentences quoted from the sources.