BDH Architecture: Latent Reasoning and Research Directions

This concluding chapter explores how BDH could move AI beyond the Transformer era, enabling models that remember, adapt, and reason at a fraction of today’s inference cost.

What BDH establishes

BDH brings identifiable neurons and synapses, recurrent synaptic memory, local computation, and GPU-compatible sparse dynamics into a single construction. The design is grounded in formal equations with inspirations from the brain and intuitions from physical processes, and its empirical behavior is validated with several experiments on language tasks.

Rewriting attention as a linear operation in a high-dimensional, sparse, non-negative neuron space turns the memory of the model into an evolving matrix of neuron-to-neuron connections, updated by sparse outer products that take the form of Hebbian writes. This gives us a graph view of the model, in which neurons carry fast transient activity, synapses carry slower contextual state, and every interaction is local. Low-rank factorization with thresholding then connects that conceptual graph to the dense tensor operations GPUs execute efficiently, so the graph is never materialized in memory and still remains recoverable for analysis.

Further empirical analyses show that BDH masters language-based abilities while extending beyond language: it solves constraint-based tasks such as Sudoku through iterative latent reasoning and retains relevant information across long-context tasks such as BABILong.

There is a further consequence that matters more than we initially expected. Unlike other LLMs available today, BDH scales like a population of cooperating neurons rather than through globally coordinated computation. Its most remarkable effect is locality: important data sits close to where it is processed. This minimizes communication and alleviates one of the main bottlenecks for reasoning models during inference: memory-to-core bandwidth. It therefore opens the door to very high inference speeds on new generations of AI hardware, from dedicated chips to next-generation GPUs.

BDH-CQ: evidence from ARC-AGI-1

Pathway’s August 2026 BDH-CQ paper reports results for a reasoning system built on the BDH architectural family. It extends BDH with in-context learning and latent iterative reasoning for visual problem-solving.

The evaluated 150M-parameter configuration reaches 29.5% pass@2 at a computed inference cost of $0.0007 per task—less than one-tenth of a cent. This operating point breaks through the previously reported ARC-AGI-1 cost-accuracy Pareto frontier, establishing a new state of the art in benchmark cost efficiency.

A black-box audit conducted by Remigiusz Kinas of Bielik AI and Richard Zhong of New York University reproduced the score under a documented protocol without access to the model weights

An independent black-box audit conducted by co-authors from Bielik and New York University reproduced the deployed system’s 29.5% pass@2 score on the public ARC-AGI-1 evaluation set. The auditors evaluated the system under a documented protocol without access to model weights.

Latent reasoning in BDH-CQ

BDH-CQ uses distinct contextual-memory and reasoning states, described in Sections 3.2–3.3 of the BDH-CQ paper.

Contextual memory: Demonstrations are processed sequentially and update a recurrent state, acting as a short-term memory while the model’s learned parameters remain fixed during inference.
Reasoning workspace: The query and accumulated memory initialize a separate latent state, which is repeatedly updated before decoding a candidate solution.

Its recurrent memory continuously integrates new context, while a high-dimensional latent workspace can preserve and refine multiple hypotheses without verbalizing every intermediate step, enabling flexible, efficient reasoning without routing all computation through tokens. Models can work through a new problem without generating long, verbalized reasoning traces, and without requiring fine-tuning or retraining.

The paper also varies the level of latent reasoning used at inference. Pass@2 rises from 21% at LOW effort to 27% at MEDIUM and 29.5% at HIGH. This provides a test-time compute mechanism without decoding intermediate reasoning into language.

The bigger picture

At Pathway, we’re building the next generations of BDH models in order to unlock new features and use cases that are out of reach for current LLMs, at a fraction of the cost. Here are our concluding thoughts:

Transformers are a bad starting point to tackle the continual learning and reasoning challenges. Long-term memory and adaptation won’t come from patchy markdown notes that increase token consumption and governance burden.
BDH is a beautiful and efficient model for moving past the Transformer era. It is built for memory, continual learning, and reasoning to emerge from efficient brain-like local dynamics rather than from engineering patches.

Watch the research talks

This three-part architectural explainer, with an introduction and conclusion, also serves as a written companion to the Pathway research team's sessions on BDH from first principles. These sessions include presentations by Jan Chorowski, Adrian Kosowski, and other team members at Mila, Stanford, the Vector Institute, Harvard, and Sakana AI.

What comes next for BDH and Post-Transformer AI?