Berkeley & DeepMind paper: Abstract Token Curriculum lets models build internal scratchpads without CoT

rbhar90 · x · 2026-09-24

A paper from Berkeley and DeepMind challenges our reliance on Chain of Thought: we burn enormous compute and human effort hand-crafting step-by-step reasoning tokens just to nudge transformers toward the right answer.

Instead, the paper introduces an Abstract Token Curriculum (ATC) — the model is never forced to emit human-readable tokens, but is fed problems through steadily harder distributions, forcing it to invent its own continuous internal scratchpads to bridge the gap.

On parity learning they prove that single-layer softmax attention naturally gravitates toward intermediate representations, showing this latent reasoning mechanism emerges on its own.

Original post →

More from Research

Research channel →