Can latent reasoning close the scaling gap with explicit chain-of-thought?

The paper starts from a scaling problem: latent CoT methods "already match CoT accuracy" on small backbones like GPT-2, but "underperform explicit CoT beyond 1B parameters, and the gap widens with scale." Its answer is LOTUS, which the authors call, "to our knowledge, the first latent-CoT method to bridge the gap to explicit CoT at the 3B scale." On Llama-3.2-3B-Instruct it closes the in-domain gap on GSM8K, surpasses CoT on the out-of-domain average, and cuts thought-phase latency by 2.5×. The abstract reports a 2.5× to 6.9× cut across CoT formats "from compact math expressions to natural language."

The recipe is deliberately simple. A looped padded Transformer processes K latent blocks in parallel for R iterations, and each latent position gets a cross-entropy loss on its gold CoT-step token, "similar to explicit CoT supervision." The cost argument is that explicit CoT of length N needs N sequential model evaluations, while latent blocks condense many steps into fewer evaluations. Ablations reportedly show that the looped backbone, the parallel gold CoT supervision, and sufficient block width and loop depth are each necessary. A representation analysis adds that the latents are "transparent": the gold CoT is recoverable by direct readout, they put graded probability on unseen but valid reasoning chains rather than one memorized trace, and the step and answer losses contribute complementary structure.

Against the vault's latent-reasoning thread, this is a scaling result more than a new mechanism. It contrasts with Can models reason without generating visible thinking tokens?, which stresses avoiding specialized training data, because LOTUS trains against the gold CoT tokens. It sits beside Can reasoning be learned during pretraining rather than after?, where looping is baked into pretraining. LOTUS instead attaches supervised looping to an existing 3B instruct model. Both report latents that stay legible against the output. The tractability worry in Can continuous thoughts have tractable likelihoods for sampling and scoring? bears on it too: LOTUS also leans on explicit CoT as its teacher. The excerpt does not say whether it keeps sampling or likelihood scoring over latents. The readout result gives partial support to Can models reason without generating visible thinking steps?, with a qualification. Reasoning need not be decoded token by token, but these latents are trained to be readable as verbal steps.

The excerpt leaves a lot open. It gives no accuracy figures, baselines, or margins, and it does not name the out-of-domain benchmarks. It does not say which CoT format yields the 2.5× end of the latency range and which the 6.9× end, and it does not test scales above 3B, though the gap is said to widen with scale. The authors flag that evaluation is on math benchmarks, as in prior latent work, and that transfer to other domains "remains an open direction." At the strength the evidence allows, the claim is that the latent-versus-explicit gap is not fixed beyond 1B, at least for math and with gold CoT supervision available. Whether latent reasoning can stand apart from verbal traces is not addressed.