Can latent reasoning match chain-of-thought cost efficiency without verbalizing?

BDH-CQ combines two capabilities that chain-of-thought models normally get through one channel: learning a task from demonstrations, and having a workspace in which to compute the answer. In the abstract, inputs presented at inference time "continuously update the model's recurrent memory," and the model then "solves a query through iterative computation in a high-dimensional latent space, without verbalizing its intermediate reasoning." The headline result is that a 150M-parameter configuration reaches 29.5% pass@2 on the public ARC-AGI-1 evaluation set at a computed inference cost of $0.0007 per task, which the authors say breaks through the previously reported cost-accuracy Pareto frontier. The claim is about cost efficiency at an operating point, not about the highest accuracy on the benchmark.

The introduction gives the mechanism as a contrast. In autoregressive models, demonstrations "specify what to do, while generated intermediate tokens support the computation required to do it," and reinforcement learning on verifiable problems has made that pairing powerful while coupling reasoning to "serial narration." Every intermediate state must be "projected through a discrete vocabulary, emitted autoregressively, and consumed again before computation can continue," so token consumption, latency, and inference compute grow with the trace. Latent reasoning is the alternative regime: the model "repeatedly transforms its continuous hidden state and decodes only the answer." BDH-CQ assigns the two jobs to two places, with demonstrations written into recurrent memory and the computation done by iteration in latent space. The discussion adds that the "unusually low starting cost" leaves room to increase capacity and reasoning compute while staying competitive.

The controlled ARC-like interventions describe what that combination can express. Dense task-specific color mappings are applied to all 96 held-out outputs. Boundary propagation and copying extrapolate throughout the tested ranges, while ordering and nested containment show distinct limits. Matched demonstrations eliminate the depth-five nesting failures and substantially improve length-eight ordering, so "coverage in the in-context examples can determine whether a learned operation extrapolates." Composition depends more on representation: rotation composes with relocation on all 72 held-out outputs, reflection succeeds on 47 of 72 and varies sharply across motif families, and color swapping is not acquired reliably outside the original fixed layout.

Against the library, Can models reason without generating visible thinking tokens? surveys depth-recurrent, Heima, and Coconut-style models that scale reasoning without more output tokens. This paper adds a recurrent memory that the demonstrations write to at inference time, so the latent computation is conditioned on examples rather than only on the query, plus a cost figure on a named benchmark. Can models reason without generating visible thinking steps? raises the question of whether verbalized thinking is necessary; BDH-CQ answers only on the cost side, at small scale. The introduction's framing of tokens as a "computational workspace" fits the scaffolding reading in Do reasoning traces need to be semantically correct?, with the workspace moved into hidden state, though the excerpt runs no such test. Can neural memory modules scale language models beyond attention limits? also updates memory at test time, and the excerpt does not say how BDH-CQ's update rule compares.

The excerpt does not establish how the model is trained, how the memory update works, how many latent iterations a task uses, or which systems sit on the earlier Pareto frontier. The cost is "computed" rather than measured, and the excerpt gives no matched comparison with a verbalizing model of similar size. What it does support is narrower: recurrent latent reasoning conditioned on in-context demonstrations is viable at 150M parameters on ARC-AGI-1, at very low computed cost, and its failures on ordering and nesting track how well the demonstrations cover the operation. It does not show that latent reasoning beats verbalized reasoning on accuracy.