BDH-CQ hits 29.5% on ARC-AGI-1 with only 150M parameters

PromptCube Intermediate 1h ago 282 views 1 likes 2 min read

The cost-accuracy trade-off in AGI benchmarks is finally shifting because BDH-CQ manages to tackle ARC-AGI-1 tasks for a fraction of a cent per run. What makes this interesting isn't just the efficiency, but the fact that it uses recurrent latent reasoning. Instead of the standard "Chain of Thought" where the model has to write out its reasoning in plain English—which consumes tokens and introduces linguistic noise—BDH-CQ does its thinking in a high-dimensional latent workspace.

Basically, it handles in-context learning by updating a recurrent memory when it sees a new task. When it comes time to solve the actual query, it iterates internally. The critical takeaway here is that intermediate reasoning states are never decoded into language. It's essentially "silent" reasoning.

How the latent workspace actually functions

Unlike typical LLM agents that rely on prompt engineering to force a step-by-step verbal breakdown, BDH-CQ integrates memory and inference into the same computational fabric. The process looks like this:

1. Memory Update: The model takes in demonstrations of a task it has never seen before. Instead of just storing these in a KV cache, it updates its recurrent memory.
2. Latent Iteration: The query is processed through iterative computation. It cycles through the latent space to refine the answer.
3. Direct Output: It jumps straight to the solution without verbalizing the "scratchpad" steps.

This architecture avoids the need for task identifiers or specific demonstration pairs during the training phase. Even more impressive is that no parameters are updated during inference; it's all happening through the recurrent state.

Performance and Efficiency

The numbers on this are wild when you consider the model size. A 150M-parameter setup—which is tiny compared to the behemoths we usually see—hit a 29.5% pass@2 on ARC-AGI-1.

Model Size: 150M parameters
ARC-AGI-1 Pass@2: 29.5%
Cost per task: $0.00070

Comparing this to larger models that try to brute-force AGI benchmarks with massive prompt windows, BDH-CQ proves that recurrent latent states can be way more efficient than token-heavy reasoning. It's a real-world example of how moving away from "thinking out loud" might actually lead to better generalization on abstract reasoning tasks.

If you're looking for a deep dive into the math, the full paper is available here: