Episode 0054: Silent Reasoning

Dispatches from the Edge | August 14, 2026

The Paper

BDH-CQ: In-Context Learning with Recurrent Latent Reasoning - arXiv: 2608.09888 - Authors: Pathway Research (Jan Chorowski et al.) - Published: August 10, 2026

What It Does

BDH-CQ is a 150-million-parameter model that cracks ARC-AGI-1 tasks through silent reasoning — iterative computation in a high-dimensional continuous latent space, without generating any intermediate tokens or chain-of-thought steps.

The architecture has two phases:

1. In-context learning: Demonstrations are ingested sequentially, each updating a recurrent memory state S. By the time all examples have been seen, S holds a compressed record of the transformation the task is specifying.

2. Latent reasoning: The query input is encoded using S, then processed through repeated applications of a learned function across a structured latent workspace H. The answer is decoded only after R rounds of silent computation.

Nothing between ingesting the context and producing the answer is ever projected through a discrete vocabulary. The model reasons without language.

The Result That Matters

29.5% pass@2 on ARC-AGI-1 at $0.00070 per task.

That is less than one-tenth of a cent. For comparison:

System	Score (pass@2)	Cost per task
BDH-CQ (150M)	29.5%	$0.00070
GPT 5.6 Luna (Low)	34.2%	$0.040 (pre-cut), ~$0.008 (post-cut)
HRM	Strong ARC score	$1.48
TRM	Strong ARC score	$1.76

BDH-CQ breaks through the cost-accuracy Pareto frontier that previously made "cheap reasoning" seem like a contradiction in terms. No prior system achieved this accuracy at equal or lower cost.

An independent black-box audit by co-authors from Bielik and New York University confirmed the 29.5% score without access to model weights.

What It Can and Cannot Do

The team ran ConceptARC — a 16-family organized benchmark — and controlled generalization experiments, systematically varying complexity to find where the model's capability profile has sharp edges.

Strengths: - ExtendToBoundary: 9/10 - FilledNotFilled: 9/10 - TopBottom2D: 9/10 - Color permutation binding: 96/96 held-out outputs, all densities tested - Boundary propagation: 48/48 out to distance 8 (no ceiling found) - Copying: 48/48 as the number of copies increases (no ceiling found)

Structured limits: - Ordering: Near-perfect through 5 objects, falls sharply (29/36 at 6, 8/24 at 7, 1/24 at 8). Failures involve constructing the wrong output dimensions entirely. - Nesting (containment depth): Near-perfect through 4, falls at depth 5 — but almost all cells are correct; errors are single containment decisions. Providing one depth-5 demonstration recovers from 19/24 to 24/24. The limitation is extrapolation from shallower context, not execution capacity.

Within-task consistency gap: 52/160 ConceptARC tasks have some correct test inputs but are not solved. Pass@2 pair accuracy (78%) is substantially higher than strict task accuracy (59%). The model can often produce the right answer for individual inputs but does not yet apply the inferred transformation consistently across an entire task.

The Architecture Behind It

BDH (Dragon Hatchling) is a post-Transformer architecture by Pathway Research built around: - High-dimensional positive activations - Low-rank communication - Recurrent associative state

"Brain-inspired but not brain-imitative" — it drew on what biological systems do well (local interaction, sparse activity, persistent state) without trying to simulate the biology.

BDH was previously used for a Sudoku solver (iteratively refining a visual state until puzzle constraints are satisfied). BDH-CQ extends that to open-ended visual reasoning from demonstrations.

Why Silent Reasoning Is Interesting

Chain-of-thought models reason through language — every intermediate state must be serialized into tokens, emitted autoregressively, then consumed again. This couples reasoning to sequential narration. Cost and latency grow with reasoning trace length.

Latent reasoning removes this constraint. Multiple hypotheses can coexist simultaneously in the continuous space without being committed to words. Theory (Zhu et al. 2025) shows continuous thoughts can expand multiple search frontiers in parallel — something token-by-token generation cannot do.

This is consistent with neuroscience evidence that conceptual reasoning and language involve partly separate systems in humans (Fedorenko and Varley 2016). You can reason without words. The architecture takes that seriously.

Scale and Future Work

• Scaling law experiments confirm Transformer-like laws from 1B to 600B parameters during BDH pretraining
• Architecture designed for 1T parameter scale with tensor sharding from the BDH base
• Next target: ARC-AGI-2 (the behavioral analysis in this paper becomes the development agenda)
• Future directions: language + math reasoning on BDH, hybrid systems combining latent reasoning with verbalized steps when communication or verification requires it

Code and Data

• Task generation tooling (training/eval infrastructure): github.com/pathwaycom/arc-task-gen
• Model weights: proprietary
• Training recipe: proprietary (data mixture is described; exact training procedure is not)

Related Work Mentioned

• ARC Prize Leaderboard
• Chollet 2019: On the Measure of Intelligence (ARC original paper)
• Coconut (Hao et al. 2024): Continuous thoughts in language models — prior work on latent reasoning
• HRM (Wang et al. 2025): Hierarchical reasoning model on ARC, $1.48/task
• Geiping et al. 2025: Recurrent-depth language models

---

Generated for DTF:FTL Episode 0054. Paper fetched August 14, 2026.