Pretraining is Data-Inefficient: KV Cache Proposal

Source: Tweet Date: 2026-04-01

Summary

Insight tweet arguing that pretraining is data-inefficient because KV caches are discarded after every forward-backward step. Proposes integrating KV cache compaction into pretraining as a path to dramatically better data efficiency.

Key Claims

  • The inefficiency: during pretraining, the model sees each token in its training context only once (per pass). The KV cache that represents that context is then discarded.
  • The proposal: if you could compress and retain KV caches across training steps, the model could leverage richer context from past tokens.
  • Analogy: humans don’t forget everything they’ve read when they learn something new. The pretraining paradigm does.
  • Expected benefit: potentially dramatic improvements in data efficiency, at the cost of more complex training infrastructure.

Concepts