Pathway's BDH-CQ Solves ARC-AGI-1 Puzzles Without Writing a Single Reasoning Token

Pathway's BDH-CQ Solves ARC-AGI-1 Puzzles Without Writing a Single Reasoning Token

Pathway, working with Bielik AI and New York University, has published BDH-CQ, a reasoning system that learns a new visual transformation from a handful of examples and solves it entirely inside a continuous latent state. The model never writes out a chain of thought. Demonstrations update a recurrent memory, and the query is answered through repeated transformations of a hidden vector that gets decoded only once, at the end, into the final grid.

On the public 400-task ARC-AGI-1 evaluation set, a 150-million-parameter configuration reaches 29.5% pass@2 at a computed cost of $0.0007 per task. According to the ARC Prize leaderboard data the authors cite, that combination of score and cost sits beyond every previously reported system's cost-accuracy curve. The paper also runs a set of controlled experiments, built after the model was frozen, that map out exactly which visual operations the system generalizes cleanly and which ones it cannot yet apply beyond the range covered by its demonstrations.

Memory first, reasoning second

BDH-CQ separates two jobs that chain-of-thought models usually collapse into one stream of tokens. A recurrent memory state absorbs each demonstration pair in sequence, so what the model has learned about the current task depends on the associations built up from earlier examples rather than a single compressed task vector. Once the demonstrations are ingested, the query input is encoded into a separate latent workspace, which the model then updates through several rounds of the same learned transformation before decoding an answer.

The architecture builds on Pathway's earlier Dragon Hatchling (BDH) model, a sequence architecture built around high-dimensional positive activations and a recurrent associative state. The authors describe that underlying design as "brain-inspired but not brain-imitative," and BDH-CQ extends the same recurrent machinery from language modeling and constraint satisfaction into learning an unseen transformation purely from in-context demonstrations, with no parameter updates and no task identifiers supplied at inference time.

The cost comparison

The headline efficiency claim rests on a comparison to other systems on the ARC Prize leaderboard. Recursive solvers like HRM and TRM report strong ARC accuracy but do so by fine-tuning on augmented copies of each evaluation task's own demonstration pairs and assigning a learned identity embedding to every puzzle before predicting an answer, which the paper notes puts a backward pass between a new task and its solution. The leaderboard prices that approach at $1.48 per task for HRM and $1.76 for TRM. BDH-CQ, by contrast, updates only its recurrent state at inference and reports roughly 0.85 GPU-seconds of compute per task on an H200, priced at $3 per GPU-hour.

Against frontier language models, the paper estimates BDH-CQ runs about 57 times cheaper than GPT-5.6 Luna (Low), which scores 34.2% at a reported $0.040 per task. After accounting for OpenAI's 80% public API price cut on July 30, 2026, a change the ARC Prize leaderboard had not yet reflected as of the paper's writing, the authors revise that gap down to roughly 11 times cheaper. That adjustment is a useful reminder that cost comparisons pulled from a leaderboard are only as current as the leaderboard's last price update, and the paper is transparent about the correction rather than quietly keeping the larger multiple.

A capability profile rather than a single score

The paper's more informative contribution is a breakdown of where the 29.5% comes from. On ConceptARC, which groups tasks into 16 hand-designed concept families of ten tasks each, strict task accuracy ranges from 9 out of 10 on families like ExtendToBoundary and FilledNotFilled down to 2 out of 10 on Copy and Order. With only ten tasks per family, the authors caution that the descriptive confidence intervals for the best and worst families overlap broadly, so the ranking is a profile rather than a reliable ordering.

A more durable finding is the gap between test-pair accuracy and strict task accuracy: 77.92% versus 59.38% across ConceptARC. Under pass@2, the system solved zero of a task's three test pairs on 13 tasks, one pair on 15 tasks, two pairs on 37 tasks, and all three on 95 tasks. In other words, 52 of 160 tasks got partial credit rather than a clean solve, meaning the model sometimes finds a transformation that works on some inputs of a task but not consistently across all of them, a pattern the authors treat as evidence against uniform rule application rather than as a scoring artifact.

Simple operators extrapolate, complex ones do not

To isolate individual operations from this aggregate picture, the authors built fresh ARC-style tasks after freezing the model and varied one dimension of difficulty at a time. Propagating a signal to a boundary and copying a motif to multiple anchor points both stay at 100% accuracy as the tested range expands, propagation from distance 2 to distance 8 and copying from one to four target sites, with no sign of hitting a ceiling.

Ordering and nesting behave differently. Sorting bars by height stays near-perfect through five objects, then drops to 29 of 36 outputs at length six, 8 of 24 at length seven, and 1 of 24 at length eight. Selecting cells inside nested containment frames holds up through depth four and falls to 29 of 36 at depth five. The failure modes differ in kind: at ordering length eight, only 3 of 24 outputs even have the right output dimensions, while at nesting depth five, all 36 outputs have correct dimensions and average better than 99.9% cell accuracy, with errors typically limited to a single misclassified containment relation.

Demonstration coverage explains most of the gap

The paper then asks whether these breakdowns reflect a hard capacity limit or simply a failure to extrapolate beyond what the demonstrations showed. Re-running the identical length-eight ordering and depth-five nesting test inputs, but adding one demonstration at the target complexity instead of stopping short of it, recovers nesting depth five from 19 of 24 to 24 of 24 exact outputs at pass@2. Ordering length eight improves from 0 of 24 to 13 of 24. The nesting failure, in particular, looks largely like a failure to extrapolate a demonstrated relation depth rather than an inability to execute it once shown, while long ordering keeps a partial bottleneck even with matched support.

Composition succeeds unevenly

A separate experiment tested whether operations learned individually still work when combined. Relocating a motif to a marked position, reflecting it, and rotating it are each solved on all 72 held-out outputs alone. Rotation composed with relocation is also solved on all 72. Reflection composed with relocation drops to 47 of 72, varying by motif color layout. A demonstration-defined color swap is learned atomically only within a motif family that has a fixed color layout, at 26 of 72, and composing it with relocation fails on all 72 held-out outputs. The authors note that two color-shuffled motif families solve the swap in isolation only 1 time in 24, which shows the fixed layout made that operation easier to infer in the first place and complicates attributing the composed failure to composition alone.

Ruling out an obvious shortcut

Because the evaluation format could let a system exploit a semantic task label or a batch grouped by concept rather than the actual grids, the authors reran ConceptARC with cryptographically opaque identifiers and mixed concept areas within batches. Aggregate performance barely moved: 96 of 160 opaque tasks solved versus 95 of 160 with semantic labels, and 374 of 480 test pairs correct under both conditions. The paired results showed exactly six semantic-only and six opaque-only successes, giving no directional signal that removing those cues changed anything. The individual candidates returned did shift between runs, with only 276 of 480 complete ordered candidate lists matching exactly, so the replication rules out this specific confound without establishing ConceptARC as an entirely fresh test of the model.

More reasoning steps cost more and score higher

The paper also trained a version of the model across three levels of latent reasoning effort. On the public evaluation set, LOW effort reaches 21% pass@2 at a 22% cost reduction relative to HIGH, MEDIUM reaches 27% at an 11% reduction, and HIGH reaches the full 29.5%. The relationship is monotonic across all three settings, and the authors frame the low absolute cost of the HIGH setting as room to spend on larger models or more reasoning steps while remaining competitive with existing systems on price.

What's still open

A few limits are worth naming plainly. The 29.5% headline figure and the cost estimate come from the paper's own authors, run on their own deployed system, using hardware time they measured themselves rather than a published API price. The paper describes an "independent" black-box audit that reproduced the 29.5% score, but the auditors, from Bielik AI and New York University, are themselves co-authors of this paper, so the audit confirms internal consistency of the reported number rather than constituting third-party replication by a group with no connection to the work.

The evaluation-set analysis in the appendix adds a similar caution about the mechanic-level breakdowns: solve rates on a generated, mechanic-stratified set range from 68.6% for flood fill down to 2.9% for gravity-and-stacking tasks, but hand-written comparison tasks for the same mechanics score very differently, 80 to 100% for gravity and 85 to 95% for counting, which the authors attribute to how the task generator instantiates a requested operation rather than to the difficulty of the operation itself. The public and generated mechanic rankings do not correspond reliably (rank correlation of 0.300 across 16 mechanics), underscoring that a generated benchmark's numbers describe the generator as much as the model.

Pathway says the architecture is designed to scale, citing early pretraining runs from 1 billion to 600 billion parameters that follow familiar scaling laws while preserving the latent-reasoning behavior described here, and lists ARC-AGI-2, Sudoku-style constraint satisfaction, and language and mathematical reasoning as the next domains for testing whether the same recurrent memory generalizes beyond visual grids.

---

David Borish is a journalist and analyst covering frontier AI, cybersecurity, and enterprise AI deployment economics, and author of the forthcoming book The Tony Hawk Paradox. More at davidborish.com.

www.tonyhawkparadox.com