Claims Browser

GLM-5.3-Flash (Ox Alpha) impressed everyone except GDM vaguepoasters

"GLM-5.3-Flash (aka Ox Alpha) impressing everyone (except GDM vaguepoasters)"

GLM-5.3-Flash is a natively multimodal model with 1M-token context window, 320B total parameters / 18B active parameters, released under MIT License

"GLM-5.3-Flash as a natively multimodal model with a 1M-token context window, 320B total parameters / 18B active parameters, released under the MIT License"

GLM-5.3-Flash outperforms GLM-5.2 at every effort level and is on par with Claude Opus 4.8 on coding according to Z.ai's internal benchmark

"claiming on its internal benchmark that it outperforms GLM-5.2 at every effort level and is on par with Claude Opus 4.8 on coding"

GLM-5.3-Flash may be the best intelligence-per-dollar option according to community response

"more substantive claims that the model may now be the best intelligence-per-dollar option"

Open Chinese labs are converging on similar architecture choices around linear attention, sparse attention, residual path design, and Muon

"posts arguing that open Chinese labs are converging on similar architecture choices around linear attention, sparse attention, residual path design, and Muon"

GLM-5.3-Flash is stronger on practical code/agentic workflows than on broad real-world factual knowledge

"GLM-5.3-Flash may be much stronger on practical code/agentic workflows than on broad real-world factual knowledge"

The subsampled randomized Hadamard transform (SRHT) achieves coordinate-wise accuracy guarantees for overconstrained L2 regression with O(ε^-2 d^(1+Θ(√(log log n/log d)))) rows

"Price, Song and Woodruff initiated the study of this problem and showed that the subsampled randomized Hadamard transform (SRHT) with $O(ε^{-2} d^{1+Θ(\sqrt{\log\log n/\log d})})$ rows achieves this guarantee."

A prior work by Song, Ye, Yin and Zhang claiming to improve row count to O(ε^-2 d log^3 n) has a flawed proof that relies on an independence assumption that does not hold in general

"A subsequent work of Song, Ye, Yin and Zhang claimed to improve the row count to $O(ε^{-2}d\log^3 n)$. Unfortunately, their proof relies on an independence assumption that does not hold in general, and we exhibit an explicit instance on which it fails."

A new fast, dense randomized transform combining Hadamard flattening, random permutation, and balanced Gaussian pooling achieves L∞ guarantees with O(ε^-2 d log d) rows

"To achieve a truly nearly-linear-in-$d$ row count, we introduce a new fast, dense randomized transform, which combines a randomized Hadamard flattening, a random permutation, and balanced, disjoint Gaussian pooling. Conditioned on the Hadamard-and-permutation stage, the sketched problem becomes an exact Gaussian regression in which the noise is independent of the entire sketched design; this conditional independence is exactly what the earlier argument was missing. Our sketch yields the $\ell_\infty$ guarantee with $m=O(ε^{-2}d\log d)$ rows"

The new randomized transform can compute the sketched pair (SA, Sb) in O(nd + ε^-2 d^4) time using one Hadamard pass

"uses one Hadamard pass with a padded internal dimension $N=\widetilde{O}(n+ε^{-2}d^3)$, and is efficient to apply: the sketched pair $(SA, Sb)$ can be computed in $O(Nd\log N)=\widetilde{O}(nd+ε^{-2}d^4)$ time."

Discrete diffusion models have become a strong, widely adopted class of generators for sequence data

"Discrete diffusion models have become a strong, widely adopted class of generators for sequence data"

Training-free steering of discrete diffusion models toward downstream rewards is increasingly important

"steering them toward a downstream reward at inference time, without any retraining, is increasingly important"

Current guided proposal methods have a weakness in estimating gradients from single noisy samples

"the guided proposal estimates its gradient from a single noisy sample"

Current search methods have a weakness in resampling particles at fixed temperature that ignores reward distribution across denoising steps

"the search then resamples particles at a fixed temperature that ignores how rewards spread across each denoising step"

GRAS method achieves best training-free reward across regulatory DNA and protein design, outperforming prior training-free methods and matching or surpassing reward-fine-tuned models

"across regulatory DNA and protein design it attains the best training-free reward, outperforming prior training-free methods and matching or surpassing a reward-fine-tuned model"

GRAS remains effective even for non-differentiable rewards

"it remains effective even for non-differentiable rewards"

Rao-Blackwellized reveal lowers estimator variance for differentiable rewards and leave-one-out baseline works for non-differentiable rewards

"we lower the estimator variance with a Rao-Blackwellized reveal for differentiable rewards and a leave-one-out baseline for non-differentiable ones"

The Descriptive-Complexity Information Criterion (DCIC) achieves selection consistency through approximation-error separation without requiring RIP-type conditions, even under sub-Weibull noise

"Under sub-Weibull noise, we establish selection consistency through approximation-error separation without relying on RIP-type conditions"