Feed
BDH-CQ: In-Context Learning with Recurrent Latent Reasoning
A 150M-parameter model learns visual puzzles from examples and reasons silently, for a fraction of a cent.
57x
cheaper
Cost per ARC task
Reaches 29.5% pass@2 for $0.00070 per task
How it works
Reasoning costs tokens
Chain-of-thought models narrate every step as text. Long traces drive up tokens, latency and inference compute.
Memory plus latent loops
Demonstrations update a recurrent memory. The query is then solved by repeated hidden-state updates, decoding only the answer.
Cheap ARC solving
Scores 29.5% pass@2 on ARC-AGI-1 at $0.00070 per task. That beats the prior cost–accuracy frontier.
From the paper