Four arXiv preprints report efficiency gains in training and quantum sampling

The papers touch the resources that set training and inference economics — compute time, memory, communication and measurement shots — rather than introducing new model capabilities.

By Adrian Cole

November 2, 2026 at 3:35 AM EST2 min read

Key takeaways (3)

1.FARO's authors report matching a hand-designed optimizer, plus a 0.0136-nat average loss gain across 6 seeds, at no added time or memory.
2.The quantum diffusion authors report comparable quality with 303 shots per sample versus 1.6×10^6, on an eight-dimensional Gaussian mixture.
3.The decentralized method is said to halve communication versus gradient tracking; its convergence guarantee covers the strongly convex setting.

Key takeaways (3)

1.FARO's authors report matching a hand-designed optimizer, plus a 0.0136-nat average loss gain across 6 seeds, at no added time or memory.
2.The quantum diffusion authors report comparable quality with 303 shots per sample versus 1.6×10^6, on an eight-dimensional Gaussian mixture.
3.The decentralized method is said to halve communication versus gradient tracking; its convergence guarantee covers the strongly convex setting.

Four arXiv preprints posted on 2 November 2026 report methods aimed at efficiency, from optimizer budget allocation and decentralized communication to quantum measurement shots and diffusion-model training.

Each paper claims a gain against a specific resource: a fixed per-block update budget, communication per round, shots per sample, or training speed. All figures are the authors' own, taken from abstracts.

The abstracts frame each problem as a cost. The FARO paper says training a large language model is compute-bound and that current recipes split work between optimizers by hand under a single global update budget (arXiv:2610.40858 1). The decentralized-optimization authors say earlier variance-reduced ADMM methods need stepsizes that shrink with the square root of network size and the cube of the number of local updates. The quantum authors say suppressing measurement noise through repeated shots raises the cost of a quantum neural network.

FARO reallocates a frozen per-block budget using equal risk contribution, a risk-parity rule from portfolio theory. On a GPT-style model trained on FineWeb, the authors say FARO-Adaptive reached 3.2780 nats at the full schedule, matching NorMuonAndAdam's 3.2794, which they call the best hand-designed recipe. They report that the ERC stage lowered clean-data loss at each of 6 seeds, by 0.0136 nats on average, at no added time or memory. That is a better result at equal cost, not a lower cost. The abstract gives no parameter count or total compute, and says the source code is available.

The decentralized paper proposes a primal-dual algorithm that, the authors say, exchanges one variable per round, halving communication cost 4 compared with gradient-tracking-based methods. They report a stepsize independent of network size and inversely proportional to the number of local updates rather than its cube. To the best of their knowledge, they say, they establish the first linear convergence guarantee in expectation for decentralized optimization with multiple local updates under stochastic gradients in the strongly convex setting. The abstract describes the numerical experiments only as showing advantages in speed and accuracy.

The quantum diffusion paper uses the noise measurement produces anyway to replace part of the noise a diffusion sampler would inject (arXiv:2610.41721 3). On an eight-dimensional Gaussian mixture, the authors report comparable generation quality with 303 shots per sample against a high-shot reference of 1.6×10^6, about 5300× fewer. They say latent diffusion experiments on MNIST further validate the mechanism, but the abstract gives no shot count for that case.

The diffusion-acceleration paper says methods such as REPA and dispersive regularization work at least in part by reorganizing the AdaLN-Zero residual gate, rather than solely through the feature structures they impose (arXiv:2610.40914 7). The authors propose a teacher-free radial objective and say it improves FID over dispersive loss on ImageNet 256×256 with SiT-B/2. The abstract reports that one model and configuration, and gives no FID values or training-time figures.

None of the four abstracts states peer-review status, and they report no independent replication or third-party evaluation.

Sources10 documents

1. arxiv.org

arxiv.org/abs/2610.40858

4 checked claims

2. arxiv.org

arxiv.org/abs/2610.41269

3. arxiv.org

arxiv.org/abs/2610.41721

3 checked claims

4. arxiv.org

arxiv.org/abs/2610.41927