Pathway Unveils BDH-CQ: A Breakthrough AI Reasoning Model That Bypasses Chain-of-Thought Tokens in Favor of Latent Recurrence

The landscape of artificial intelligence research is experiencing a profound paradigm shift away from traditional, token-heavy generation models toward internal, continuous cognitive processing. In a development that could fundamentally alter how large language and reasoning models scale, researchers at Pathway and their collaborators have introduced BDH-CQ, a novel reasoning architecture. Unlike contemporary autoregressive models that rely heavily on intermediate thought tokens—a technique popularized by systems like OpenAI’s o3-mini and DeepSeek R1—BDH-CQ performs in-context learning and problem-solving entirely within a recurrent latent space, decoding only the final answer upon completion.

This technical departure addresses a growing bottleneck in modern AI infrastructure. While chain-of-thought methodologies have dramatically improved logical, mathematical, and coding capabilities, they inherently tie computational depth to a serial, text-based channel. Every intermediate hypothesis, constraint check, and strategic pivot must be compressed into a token sequence, driving up latency, computational overhead, and financial costs. BDH-CQ rethinks this dynamic by shifting the computational heavy lifting directly into the model’s continuous hidden states.

Background Context and the Evolution of Latent Reasoning

To understand the significance of BDH-CQ, one must examine the limitations of current autoregressive paradigms. Standard reasoning models operate on a sequential feedback loop: the system reads a prompt, generates a token, appends it to the active context, and repeats the process. Although this mechanism allows humans and auxiliary software components to audit intermediate reasoning steps, it places a severe ceiling on efficiency. As reasoning traces grow longer to accommodate harder problems, token consumption and hardware demands skyrocket.

Zuzanna Stamirowska, Chief Executive Officer of Pathway, points to a deeper philosophical and structural flaw in treating language as the primary substrate for computation. According to Stamirowska, reasoning and natural language are related yet distinct phenomena. Traditional text-based architectures force models to constantly verbalize their intermediate computations while navigating constraints, maintaining alternative hypotheses, and converging on a solution.

Latent reasoning offers an alternative pathway. By repeatedly modifying a continuous hidden representation rather than spelling out thoughts in natural language, a model can simultaneously preserve partial hypotheses, explore multiple computational branches in parallel, and sidestep the serialization penalty of text. Prior academic and industrial efforts—such as Coconut and various looped or recurrent-depth Transformer architectures—have explored continuous representations, but they often struggled with flexible task acquisition. Most recurrent systems required intensive optimization or task-specific representations to adapt to unseen evaluation workloads. BDH-CQ bridges this gap by merging in-context learning capabilities with continuous latent reasoning.

Architecture of BDH-CQ: From Dragon Hatchling to Contextual Latent Workspace

BDH-CQ is built upon Pathway’s foundational Dragon Hatchling (BDH) architecture, which diverges sharply from standard autoregressive Transformers. While standard models rely heavily on dense attention maps and token prediction, BDH utilizes high-dimensional positive activations, low-rank communication, and a persistent recurrent associative state.

Information within a BDH network is represented as patterns of active features across a large internal space, governed by ReLU-based positive activations. Instead of permitting every feature to communicate densely with the entire network, the architecture employs lower-dimensional transformations to route data efficiently. Crucially, a persistent recurrent state enables the model to retain historical inputs and continuously update its internal memory without forcing every computational step to manifest as a generated token.

Previously, Pathway demonstrated the power of the BDH framework on a Sudoku benchmark consisting of 250,000 Extreme puzzles. While leading reasoning engines like OpenAI o3-mini and DeepSeek R1 reportedly scored zero on the specific evaluation set due to the complex constraints and lack of token-based scratchpads, the BDH-based system achieved a 97.4% solve rate by repeatedly refining a latent representation until all puzzle constraints were satisfied.

BDH-CQ extends this foundational design by introducing a dual-interface system that separates task acquisition from query execution. When presented with a novel task through demonstrations—such as abstract visual reasoning examples—the system processes input-output pairs sequentially. These examples do not alter the model’s underlying parameters; rather, they dynamically modify the recurrent contextual state, building task-specific associations in memory.

Once the demonstrations have shaped the recurrent memory, the actual query enters a separate latent workspace. The model conditions its computations on the contextual memory while iteratively updating the workspace over multiple cycles. Only after the latent computation reaches a terminal state does the system decode the final output into the target format.

Empirical Performance and Cost-Efficiency on ARC-AGI-1

To validate the architecture, the research team tested a 150-million-parameter BDH-CQ model on the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI-1), a widely recognized benchmark for abstract visual reasoning that resists brute-force memorization.

The results demonstrated a striking advantage in economic efficiency and compute scaling. The compact 150M-parameter BDH-CQ model achieved a 29.5% score (pass@2) on ARC-AGI-1 while consuming approximately 0.85 H200 GPU-seconds per task. Based on an estimated hardware operational cost of $3 per H200-hour, the computed inference cost came out to roughly $0.00070 per task. According to the researchers, this metric pushes the model well beyond the previously established cost-accuracy Pareto frontier for the benchmark.

Stamirowska attributes these efficiency gains directly to the internalization of memory and reasoning. By eliminating expensive token-based reconstruction steps and executing iterations within the hidden state, the system avoids the repetitive overhead that plagues traditional large-scale models.

However, granular analysis of the experimental data reveals specific operational boundaries. While the model excelled at certain few-shot adaptations—such as perfectly mastering novel color mappings and boundary propagation across held-out sequences—its performance degraded when scaling up complex relational properties. For instance, when tasked with ordering objects by height, accuracy remained high up to five objects before dropping sharply. Furthermore, compositional operations showed mixed reliability: while combinations of rotation and relocation achieved perfect scores on held-out tests, reflection combined with relocation saw notable drops in accuracy. Consistency also varied; on ConceptARC, the model scored 77.92% on individual test pairs, but that figure dropped to 59.38% when a task was judged successful only if all three test inputs were solved concurrently.

Implications for Agentic Engineering and Enterprise Workflows

Beyond academic benchmarks, the persistent-state mechanism embedded in BDH-CQ holds substantial implications for the future of AI agent engineering. Current enterprise agent stacks are heavily burdened by the stateless nature of traditional model calls. To maintain operational continuity, developers must build complex external scaffolds: retrieving conversation histories, maintaining external vector stores, summarizing historical contexts, compressing prompts, and orchestrating external tools.

Stamirowska highlights the operational toll this takes on software engineers. Every step an agent executes often requires re-reading and re-summarizing expansive contexts, resulting in latency and token usage that scale upward with task complexity simply because the system must repeatedly reconstruct what it has already processed.

By transitioning from context reconstruction to persistent intelligence, architectures like BDH-CQ could shift a portion of this burden from the external agent harness into the model itself. Agents could maintain task-specific associations and experiential states across multiple interactions rather than relying entirely on exhaustive prompt engineering. While retrieval-augmented generation (RAG) and external factual databases would still be required for accessing live, authoritative data, internal persistent states could drastically reduce the infrastructure dedicated to reminding models of information they have already encountered.

Comparative Analysis Within the Recurrent Reasoning Ecosystem

BDH-CQ joins a rapidly expanding cohort of recurrent and loop-based architectures challenging the supremacy of standard Transformers, including the Hierarchical Reasoning Model (HRM) and the Tiny Recursive Model (TRM). While HRM and TRM have demonstrated formidable capabilities on puzzles like ARC and Sudoku, their adaptation pipelines typically rely on optimization phases, where demonstration pairs are augmented during training and unseen tasks require backward-pass adaptation before evaluation.

In contrast, BDH-CQ integrates adaptation directly into the inference phase. By keeping model weights frozen while modifying recurrent memory via in-context demonstrations, BDH-CQ mirrors the intuitive interaction patterns developers expect from general-purpose large language models. Furthermore, while looped-Transformer variants—such as those linked to exploratory architectures like OpenAI’s GPT-6 Astra—focus primarily on parameter-efficient depth and test-time compute via shared Transformer blocks, BDH-CQ is engineered from the ground up to decouple task memory acquisition from query execution.

Despite these advancements, historical hurdles associated with recurrent neural networks—such as optimization difficulties and challenges with massively parallelized distributed training—remain points of technical vigilance. Nevertheless, the research team reports Transformer-like pretraining scaling behaviors for BDH-CQ models ranging from 1 billion to 600 billion parameters. Utilizing inherited tensor-sharding patterns from the base BDH framework, the architecture demonstrates technical viability for massive, trillion-parameter deployments without encountering fundamental scaling penalties.

Commercial Roadmap and Future Outlook

As the artificial intelligence community digests these findings, immediate access to the technology remains strictly controlled. Pathway has elected not to open-source the model weights, complete training code, or precise update recipes for BDH-CQ. Instead, the company intends to commercialize the architecture through a managed service model.

According to Stamirowska, the commercial rollout will transition from collaborative design partnerships to a targeted research-preview application programming interface (API), followed by a wider beta release and eventual general availability. This strategy allows enterprise developers to leverage the architectural efficiencies of latent reasoning without exposing proprietary training methodologies.

Future research will test whether the efficiencies observed in visual reasoning and constraint satisfaction translate seamlessly to natural language processing, complex mathematical theorem proving, agentic software coding, and extended enterprise workflows. While purely latent systems excel at internal computation, industry leaders predict a hybrid future. Rather than eliminating natural language entirely, latent reasoning and text-based generation are expected to work in tandem—allowing models to perform heavy logical calculations invisibly in the background while outputting human-readable text only when communication, verification, or external tool execution is required.