---
type: analysis
slug: research-bdh-cq-is-a-cheap-arc-in-context-latent-solver-not-a-proven-long-horizon-win
title: "BDH-CQ is a cheap ARC in-context latent solver, not a proven long-horizon win"
description: "Pathway's 150-million-parameter BDH-CQ reports 29.5% pass@2 on public ARC-AGI-1 by writing demonstrations into recurrent memory and iterating in a latent workspace instead of verbal chain-of-thought. Builders should treat the in-context latent design as real and the headline Pareto number as not audit-ready: matched recurrent baselines, official leaderboard listing, and internally consistent cost"
published_at: 2026-10-26T02:00:00.000Z
topics: ["memory","architecture","design","research"]
source_urls: ["https://arxiv.org/abs/2608.09888","https://pathway.com/research/introducing-bdh-cq","https://github.com/pathwaycom/arc-task-gen","https://arcprize.org/leaderboard","https://arcprize.org/leaderboard/community","https://arcprize.org/results/openai-gpt-5-6-luna","https://arcprize.org/results/openai-gpt-5-6-luna-2026-07-30","https://pathway.com/blog/pathway-150m-model-breaks-arc-agi-1-cost-efficiency-frontier","https://arcprize.org/blog/hrm-analysis","https://arxiv.org/abs/2510.04871","https://arxiv.org/abs/2412.06769","https://github.com/pathwaycom/bdh","https://pathway.com/research/beyond-transformers-sudoku-bench"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
BDH-CQ is Pathway's 150-million-parameter system, posted to arXiv on 10 August 2026, that writes sparse visual demonstrations into a recurrent memory and then iterates a separate latent workspace before decoding only answer grids. Most authors are at Pathway; Remigiusz Kinas (Bielik AI) and Richard Zhong (NYU) are listed co-authors. The evaluation target is public ARC-AGI-1, and Pathway's research post presents 29.5% pass@2 at a computed $0.00070 per task as a new cost-efficiency frontier versus general-purpose chain-of-thought models, especially GPT-5.6 Luna (Low).

Builders care because the design promises extra reasoning compute without verbal traces, growing key-value caches, or test-time gradient steps. If that holds on long tasks against other recurrent and latent methods, architecture becomes a real lever beside scale. The paper keeps dimensions, update rules, and the training recipe proprietary, reports conflicting costs for the same score, and defers language, math, and ARC-AGI-2. The open question is whether this is a durable long-horizon win or a specialized public-eval ARC operating point.

## What we checked

- **Does BDH-CQ deliver a durable accuracy win on long tasks against comparable recurrent and latent-reasoning baselines, and how do compute, latency, and interpretability change?** partly confirmed. On public ARC-AGI-1, a 150M BDH-CQ reports 29.5% pass@2 with frozen weights and extra compute spent as latent iterations (LOW 21%, MEDIUM 27%, HIGH 29.5%). That is not a matched win against recurrent solvers: TRM reports 45% on ARC-AGI-1 with a 7M transductive pipeline, HRM scores 32% semi-private at $1.48 per task, and BDH-CQ has no published language or math long-horizon numbers. Latency is only GPU-seconds with conflicting dollar conversions; BDH-CQ interpretability is not shown because update rules are proprietary and the evaluator returns grids, not traces. [1][2][9][10][11]
- **The reported ARC-AGI-1 Pareto point of 29.5% pass@2 at $0.00070 per task is internally consistent and independently verified.** refuted. Section 5 prices the default 29.5% point at $0.00070 from 0.85 H200 GPU-seconds at $3 per hour, but Section 6.6 prices STANDARD effort at $0.00265246 for the same 118/400 score. BDH-CQ is absent from the official and community ARC Prize leaderboards opened here. The named black-box auditors are listed co-authors, and their protocol reports were not found as public documents. [1][2][4][5]
- **BDH-CQ infers unseen ARC transformations from demonstrations at inference without test-time training or evaluation-task identities, unlike HRM and TRM.** partly confirmed. The paper states that evaluation-task demonstration pairs and task identifiers are excluded from training and that no parameters are updated at inference. An opaque-identifier ConceptARC rerun is essentially unchanged (96/160 versus 95/160 tasks). Those are author-reported protocol facts: weights are unreleased, so the exclusion cannot be audited, whereas ARC Prize showed HRM's ARC score is largely test-time training on evaluation-task demos plus puzzle embeddings. [1][9][10]

## What we found

- The evaluated interface is explicit even though the implementation is not. Demonstrations update a recurrent state St = U_theta(St-1, Dt) with frozen parameters; after K demonstrations a query is encoded as H0 = E_theta(x*, SK), refined by Hr+1 = F_theta(Hr, SK), and decoded as yhat = G_theta(HR). Intermediate states are not verbalized. The paper states that neither task identifiers nor evaluation-task demonstration pairs participate in training. Dimensions and exact update rules remain proprietary, so the equations are a system contract, not a recipe a third party can train. [1][2]
- On the 400-task public ARC-AGI-1 set the 150M configuration reports 97/400 pass@1 (24.25%) and 118/400 pass@2 (29.50%), with Wilson interval [25.24, 34.15] on pass@2. ConceptARC strict-task pass@2 is 95/160 with semantic IDs and 96/160 with opaque IDs. Table 5 shows latent-effort scaling: LOW 21%, MEDIUM 27%, HIGH 29.5% pass@2, with reported cost reductions of 22% and 11% relative to HIGH. Training is described as privately curated ARC-style data plus public ARC-AGI-1 training, RE-ARC, ConceptARC, ARC-Heavy, and ARC-GEN100K. This is a self-reported public-eval result, not an ARC Prize verified semi-private score. [1][3]
- The headline cost does not survive the paper's own later table. Section 5 converts approximately 0.85 H200 GPU-seconds at $3 per H200-hour into $0.00070 per task for the 29.5% point. Section 6.6 reports STANDARD effort at $0.00265246 per task for the same 118/400 score, and MIN effort at $0.00088399 for 111/400. BDH-CQ does not appear on the official ARC Prize leaderboard or the community leaderboard pages opened in this session. Pathway says a black-box audit reproduced 29.5%, but names the auditors as listed co-authors; the cited protocol reports were not found as public documents. [1][2][4][5]
- The comparison to GPT-5.6 Luna is only half verified. ARC Prize's 9 July 2026 result page lists Luna Low at 34.2% on ARC-AGI-1 with API prices of $1 input and $6 output per million tokens. A 30 July page confirms an 80% Luna price cut to $0.20 / $1.20 and reports Luna Max at $0.07 per task on ARC-AGI-1 Semi-Private; Low after the cut is listed at 39.5%, without a Low per-task dollar figure on that page. The paper's $0.040 Low cost, and therefore the 57x and 11x cheapness ratios, were not found on the ARC Prize pages opened here. Luna is a generalist API model; BDH-CQ is a 150M ARC-specialized visual solver. [1][6][7][8]
- Controlled probes argue against a uniform long-task win. Propagation and copying stay at 48/48 held-out outputs in the tested range, but ordering collapses from near saturation at length five to 1/24 at length eight, and nesting falls at depth five. Matched demonstrations recover nesting (24/24) more than ordering (13/24). ConceptARC shows an 18.5-point gap between pair pass@2 (77.92%) and strict-task pass@2 (59.38%): 52 of 160 tasks are only partially solved. Composition is operation-dependent (rotation plus relocation 72/72, reflection plus relocation 47/72, color-swap plus relocation 0/72). These are structured failure modes, not evidence that extra latent depth solves long or compositional tasks. [1]
- Comparable recurrent ARC systems are not beaten on a matched protocol. ARC Prize's HRM analysis reports 32% on ARC-AGI-1 Semi-Private at $1.48 per task, finds that a same-size transformer nearly matches HRM, and attributes most of the score to an outer refinement loop plus training on evaluation-task demonstrations and puzzle identities rather than few-shot context. The Tiny Recursive Model paper reports 45% on ARC-AGI-1 and 8% on ARC-AGI-2 with 7M parameters under that transductive recursive setup. Coconut, a Transformer latent-thought method, reports gains on GSM8k and planning-heavy ProsQA with fewer generated tokens than verbal CoT, but BDH-CQ has not published language-task numbers. The architectural contrast (in-context memory versus test-time training) is real; an accuracy win on long tasks is not shown. [9][10][11][1]

## What it means

Do not replace a production chain-of-thought stack with BDH-CQ. The useful idea is the split between a recurrent demonstration memory and a latent query workspace, plus the claim of no test-time training. You cannot train or serve that system from public artifacts: BDH-CQ weights, dimensions, and recipe are proprietary, and the GitHub BDH repository is the parent language architecture, not the evaluated ARC pipeline. If you need an open recurrent ARC baseline you can actually run, TRM and HRM remain the reproducible references, with the caveat that their ARC scores are transductive.

Treat dollar Pareto claims as self-reported hardware estimates mixed with API prices. Recompute any comparison at a single metric (GPU-seconds or API dollars), and wait for an official ARC Prize listing on semi-private tasks. For product risk, latent recurrence removes the inspectable verbal trace; BDH-CQ currently returns grids only, so you lose both CoT auditability and open-weight mechanistic access. Copy the in-context-plus-latent interface in open models if you want the research bet; do not budget latency or cost from the $0.00070 figure until Pathway reconciles Section 5 with Section 6.6.

## Still open

- Will ARC Prize verify BDH-CQ on semi-private ARC-AGI-1 or ARC-AGI-2, and at which cost metric?
- How is $0.00070 in Section 5 reconciled with $0.00265246 STANDARD in Section 6.6 for the same 118/400 score?
- How does BDH-CQ compare to TRM, HRM, Coconut, and recurrent-depth models under a shared in-context protocol and shared hardware accounting?
- Do language, mathematical, and ARC-AGI-2 results exist at a disclosed scale, or only the paper's 1B-to-600B pretraining claim?

## Sources

1. [BDH-CQ: In-Context Learning with Recurrent Latent Reasoning](https://arxiv.org/abs/2608.09888) (primary)
2. [Reasoning at a fraction of the compute](https://pathway.com/research/introducing-bdh-cq) (primary)
3. [pathwaycom/arc-task-gen](https://github.com/pathwaycom/arc-task-gen) (primary)
4. [ARC Prize Leaderboard](https://arcprize.org/leaderboard)
5. [ARC-AGI Community Leaderboard](https://arcprize.org/leaderboard/community)
6. [GPT-5.6 Luna - ARC-AGI Results](https://arcprize.org/results/openai-gpt-5-6-luna)
7. [GPT-5.6 Luna 2026-07-30 - ARC-AGI Results](https://arcprize.org/results/openai-gpt-5-6-luna-2026-07-30)
8. [Pathway's 150M-Parameter Model Breaks the ARC-AGI-1 Cost-Efficiency Frontier](https://pathway.com/blog/pathway-150m-model-breaks-arc-agi-1-cost-efficiency-frontier) (primary)
9. [The Hidden Drivers of HRM's Performance on ARC-AGI](https://arcprize.org/blog/hrm-analysis)
10. [Less is More: Recursive Reasoning with Tiny Networks](https://arxiv.org/abs/2510.04871)
11. [Training Large Language Models to Reason in a Continuous Latent Space](https://arxiv.org/abs/2412.06769)
12. [pathwaycom/bdh](https://github.com/pathwaycom/bdh) (primary)
13. [Pathway's BDH solves Sudoku Extreme with 97.4% accuracy](https://pathway.com/research/beyond-transformers-sudoku-bench) (primary)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Memory](https://newruntime.com/topics/memory.md): Explore the memory topic hub.
- topic hub: [Architecture](https://newruntime.com/topics/architecture.md): Explore the architecture topic hub.
- topic hub: [Design](https://newruntime.com/topics/design.md): Explore the design topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)