BDH-CQ: In-Context Learning with Recurrent Latent Reasoning

Imagine you're developing an AI that needs to solve complex queries by reasoning through information over time. Traditional models often struggle with maintaining context and adapting their reasoning as new inputs come in, leading to issues like losing track of important details or failing to apply learned transformations consistently. This is what's called context drop or reasoning inconsistency. These challenges are particularly evident in tasks that require iterative thinking or when the model needs to build on previous knowledge without explicit prompts.

To address these limitations, BDH-CQ introduces a model that continuously updates its memory during inference, allowing it to reason through queries in a high-dimensional latent space without verbalizing its intermediate steps. This approach leverages in-context learning, where the model learns from examples presented during the task, and combines it with a recurrent memory mechanism that helps it keep track of ongoing reasoning. The evaluation on the ARC-AGI-1 set shows that a 150M-parameter version of this model achieves a pass rate of 29.5% at a low inference cost, breaking new ground in cost-efficiency compared to previous benchmarks. For builders, this means a more effective way to create AI systems that can handle complex reasoning tasks in real-world applications.

The model introduces a novel approach to reasoning by integrating in-context learning with recurrent memory.

Evaluation on a public benchmark with controlled interventions supports the model's claims, though further validation may be needed.