$ cat wiki/papers/2026/2608.09888-bdh-cq.md
BDH-CQ: In-Context Learning with Recurrent Latent Reasoning (arXiv:2608.09888)
TL;DR
A 150M-parameter model reasons by iterating in latent space instead of emitting chain-of-thought tokens, and reaches 29.5% pass@2 on ARC-AGI-1 at a computed $0.0007 per task — claimed as a new point on the benchmark's cost-accuracy Pareto frontier, not on its accuracy leaderboard (source).
Authors & Org
Not obtainable. arxiv.org and huggingface.co have both been unreachable
from this environment for eleven consecutive days, and no page of the paper was
read. Recording the author list as unknown rather than guessing it
(source).
Listed as the #1 trending paper on HuggingFace Daily Papers as of 2026-08-11, on a rolling 7-day window (source).
Method
- Combines in-context learning with recurrent latent reasoning.
- Inputs presented at inference time continuously update the model's recurrent memory — the demonstrations are absorbed into state rather than re-attended as context.
- A query is then solved by iterative computation in a high-dimensional latent space, without verbalizing intermediate reasoning (source).
The second and third points are the whole design: no reasoning tokens are generated, so test-time compute is spent in recurrence rather than in decoding. That is what makes a per-task cost of $0.0007 arithmetically possible at all.
Results
| Measure | Figure |
|---|---|
| Parameters | 150M |
| ARC-AGI-1 (public eval set) | 29.5% pass@2 |
| Computed inference cost | $0.0007 / task |
| Claimed to **break through the previously reported ARC-AGI-1 cost-accuracy Pareto | |
| frontier** | |
| (source). |
Also evaluated with controlled ARC-like interventions, used to study what the model learns from demonstrations, how consistently it applies an inferred transformation, and which concepts remain difficult (source).
Two things this table is not. 29.5% is not a leading ARC-AGI-1 accuracy —
the claim is that nothing previously reported reached it at that price. And
$0.0007 is described as a computed cost; whether it is measured on specific
hardware or derived from FLOPs was not obtainable from anything read, which
matters because the entire result is a ratio with that number in the denominator.
Significance
Test-Time Compute (Inference-Time Compute Scaling) has been, in practice, a story about spending more decoded tokens: longer chains, more samples, a verifier over the samples. The costs on this wiki's frontier reasoning pages follow from that — reasoning is expensive because reasoning is printed.
This is the opposite trade at three orders of magnitude smaller scale: the computation happens in recurrent latent state, and nothing is printed. If it holds, the interesting consequence is not the ARC score but the claim that a 150M model can occupy a useful point on a benchmark whose leaderboard is otherwise populated by frontier systems.
It also lands the same week as Stealing Reasoning Traces from Proprietary LLM APIs (arXiv:2608.09867), and the two make an accidental pair worth noting: one describes providers going to some length to keep generated reasoning traces from being read, the other a system that produces no readable trace at all. Un-verbalised reasoning is un-stealable and equally un-auditable — Mechanistic Interpretability's standing objection to latent reasoning, arriving here as a side effect rather than a design goal.
Caveat on provenance. This is the first HuggingFace-Daily-derived item in eleven days to arrive with an arXiv identifier, and it arrives through search extracts only. Nothing here has been checked against the paper.
Open Questions
- Who wrote it, and where? Unknown from anything read.
- What is
$0.0007measured on? Computed how, on what hardware, at what utilisation. - Does it transfer off ARC? ARC-AGI-1 rewards exactly the inductive pattern-completion this architecture is built for. Nothing read reports a second benchmark.
- What is the pass@1? Only pass@2 is quoted, and the gap between them is a measure of how much of the result is sampling.
- Does the recurrent-memory update survive longer demonstration sets, or is there a horizon past which state saturates?
- Is BDH-CQ open? No weights, code or licence surfaced in anything read.
Cite
BDH-CQ: In-Context Learning with Recurrent Latent Reasoning (2026).
arXiv:2608.09888.