BDH-CQ: In-Context Learning with Recurrent Latent Reasoning
Björn Engdahl, Adrian Kosowski, Jan Chorowski, Zuzanna Stamirowska, Przemysław Uznański, Junlin Jiang, Rohan Phadke, Remigiusz Kinas, Richard Zhong
cs.NE, cs.AI, cs.LG, stat.ML
2026-08-11
BDH-CQ folds in-context learning into recurrent memory and reasoning into latent-space iteration; a 150M model hits 29.5% pass@2 on ARC-AGI-1 at $0.0007 per task, breaking the prior cost-accuracy Pareto frontier.
Chain-of-thought reasoning makes a model pay for thinking in tokens: every intermediate state must be projected through a discrete vocabulary, emitted one token at a time, and read back in before the next step runs. That is expensive in tokens, latency, and inference compute. Two lines of work that could offset this have stayed separate: chain-of-thought models learn flexibly from context but depend on generating tokens, while recursive solvers that iterate in latent space need task-specific optimization. BDH-CQ puts demonstration-driven memory updates and iterative latent reasoning into one model, and never decodes the reasoning into language.
BDH-CQ is built on Dragon Hatchling (BDH), a post-Transformer sequence model with high-dimensional positive activations, low-rank communication, and a recurrent associative state. The key design is two state vectors with different jobs:
St and Hr are conceptually distinct: one stores the rule learned from demonstrations, the other holds the working computation for this query. Intermediate states are never decoded into tokens, so thinking and speaking are separated. Reasoning compute is adjustable through the iteration count R.
The headline numbers are on the public ARC-AGI-1 set (400 tasks):
| Config | Metric | Result |
| 150M params | pass@2 | 29.5% [95% Wilson 25.24, 34.15] |
| same | pass@1 | 24.25% (97/400) |
| same | cost per task | $0.00070 (0.85 H200 GPU-seconds) |
This point sits outside the previously reported cost-accuracy Pareto frontier: on the leaderboard at the time, no plotted system reached at least this accuracy at equal or lower cost, and an independent third party reproduced the 29.5%. On the more controlled ConceptARC (160 tasks), strict-task pass@2 is 59.38% and test-pair accuracy 77.92%.
The behavioral picture is mixed. Propagation (distances 2 to 8) and copying (1 to 4 targets) are 48/48 perfect, and color binding is 24/24 with 2 to 8 simultaneous bindings, a substantial contextual capacity. Extrapolation breaks, though: ordering falls to 29/36, 8/24, 1/24 at lengths 6/7/8, and nesting drops at depth 5. Add one demonstration at the test complexity, however, and depth-5 nesting recovers from 19/24 to 24/24 while length-8 ordering recovers from 0/24 to 13/24, so much of the extrapolation failure is really a coverage gap in the examples. Stratified by mechanic over 1,131 generated tasks, flood fill is highest at 68.6% and gravity and stacking lowest at 2.9%. More reasoning effort helps: HIGH 29.5%, MEDIUM 27%, LOW 21%.
For practitioners this is a workable form of reasoning that does not burn tokens: 150M parameters, less than a tenth of a cent per task, hitting a leaderboard-worthy cost-efficiency point on a benchmark as scrutinized as ARC-AGI-1. It folds latent multi-step thinking and learning a rule from a few examples into one set of recurrent weights, with reasoning depth as a knob. Against systems that use hundreds of billions of parameters and long chain-of-thought to chase the top of the leaderboard, this is a cheaper alternate path. The caveat is that 29.5% is far from the ARC-AGI-1 top; the selling point is cost, not the leaderboard crown.
The authors are unusually forthright. Consistency is the main problem: on ConceptARC, test-pair accuracy is 77.92% but strict-task accuracy only 59.38%, an 18.5-point gap meaning the same transformation is often not applied reliably to all of a task's inputs, and 52 of 160 tasks had one or two correct test inputs but were not solved as tasks. Composition is representation-dependent: reflection composed with relocation gets 47/72, but color swap composed with relocation gets 0/72. Extrapolation has hard edges: ordering hits an execution bottleneck while depth-5 nesting is a failure to extrapolate the demonstrated relation depth. Conditional rule selection drops from 100% to 56.7% (p=9 times ten to the minus ninth); when a test parameter value never appears in the demonstrations, 0 of 120 tasks are solved; three-panel union (from 2 to 3 panels) collapses from 65% to 2.5%. Results are specific to ARC-like visual reasoning, and the 150M model leaves a large budget to scale capacity.