Stanford CS25: Transformers United V6 I Distinct Modes of Generalization from Parameters and Context

Fine-tuned models fail basic relational reversals

When trained on facts like 'X is Y's parent,' models cannot answer 'Who is X's child?' despite high accuracy on the forward relation.

In-context learning solves reversal instantly

Placing the same dataset in the model's context window yields 99% accuracy on reversal questions without any parameter updates.

Syllogistic reasoning shows similar gaps

Models trained on logical premises fail to deduce conclusions parametrically, yet succeed when the premises are provided in-context.

Pre-training from scratch doesn't fix the gap

Even when trained from scratch on datasets containing explicit reversals, models memorize forward relations perfectly but show zero generalization to held-out reversals in parameters.

Parametric learning binds to explicit form

Knowledge stored in weights consolidates across documents but remains tied to how information was explicitly stated during training.

Context preserves latent structural flexibility

Information in context maintains richer detail and relationships, allowing flexible manipulation and generalization at test time.

Statistical vs. symbolic generalization trade-off

Parametric learning extracts broad statistical patterns across documents, while in-context learning enables specific flexible reasoning on provided data.

Architecture dependence of the curse

The reversal curse is systematic in causal next-token prediction models but can be mitigated through bidirectional architectures or modified learning objectives.