In this paper we derive a Probably Approximately Correct (PAC)-Bayesian error bound for partially observed linear time-invariant (LTI) stochastic dynamical systems in state-space form with inputs and sub-Gaussian noise. Such bounds are widespread in machine learning, and they are useful for characterizing the predictive power of models learned from finitely many data points. The bound derived in this paper relates the expectation of prediction errors with the prediction error generated by the model on the data used for learning. In addition, we show that it can also be used to derive bounds for the parameter estimation error. In turn, this allows us to provide finite-sample error bounds for the prediction error and parameter estimation error for a wide class of system identification algorithms. Furthermore, as LTI systems are a sub-class of recurrent neural networks (RNNs), these error bounds could be a first step towards PAC-Bayesian bounds for RNNs.
We study zeroth-order optimization of nonconvex functions with the aid of directional hints, which are cheap but potentially inaccurate approximations of the true gradient direction, given by linear subspaces at each iteration. To leverage these hints adaptively while maintaining robustness to their quality, we introduce Control-Variate Zeroth-Order Descent (CV-ZOD), a new framework that refines the classical zeroth-order gradient estimator with a control variate that can be set based on the directional hints. We first show that the oracle algorithm that optimally sets the reference vector and step size at each iteration achieves a convergence rate that interpolates between the first-order $O(1/T)$ rate and the zeroth-order $O(d/T)$ rate, depending on the quality of the hints along the trajectory. We then develop a practical variant of CV-ZOD that achieves the same oracle guarantee up to logarithmic factors, without any prior knowledge of the hint quality. We validate the method empirically on simulation-based scientific optimization tasks, demonstrating sustained progress on nonconvex landscapes where zeroth-order descent is slower and existing guided methods stall as guidance deteriorates.
We study the training dynamics of multiclass logistic regression on high-dimensional mixture models with a large number of classes and establish precise scaling laws governing the cross-entropy risk under gradient-based optimization. We show that learning proceeds sequentially across classes, from most to least frequent. When the class priors follow a power law distribution, the risk dynamics decompose into three phases: an initial plateau until the first class is learned, a power-law decay regime during which sequential learning occurs, and a final convergence regime. We then analyze how model capacity interacts with optimization under a fixed compute budget. When the effective dimension is restricted via projection onto leading principal components, the risk decomposes into a capacity term (a power law in the retained dimension) and an optimization term (a power law in training time). Optimizing this tradeoff yields a compute-optimal scaling law for logistic regression, with explicit prescriptions for model size and training time as functions of compute. These results extend theoretical scaling laws from linear regression to multiclass classification, while connecting to empirical scaling laws observed in large-scale neural networks.
Foundation models for ground states in spin-1/2 systems are a promising method for problems ranging from quantum chemistry to identifying new phase diagrams. Nearly all such models are currently pure-states that condition on the Hamiltonian's parameters, whose Monte Carlo samples give energy estimates according to the variational principle. In this contribution, we show that this representation is topologically obstructed. For any gapped Hamiltonian family whose ground-state bundle is non-trivial, every continuous normalized state-vector model has zero fidelity with the ground state at some parameter value in the Hamiltonian family. For that value, the energy is at least one spectral gap, $\Delta$, with an $O(\Delta)$ gap in an open-neighbourhood of that point. We show that this is a sufficient no-go also in the case of degenerate ground-state manifolds, time dynamics, and periodic systems with mixed space-time topology, demonstrating these obstructions on one- and two-qubit systems. We discuss how this causes a spike in the fidelity susceptibility, giving a numerical signature of a phase-transition where there is none. Finally, we show that operator-valued models canonically avoid these obstructions and preserve topological information, implying a structural necessity in representation for foundation neural quantum states.
Activation steering promises modular control of language model behavior: a behavior such as politeness corresponds to a direction in a model's activations, and adding that direction while it generates should switch the behavior on and leave everything else alone. It does not. We ask what decides which other behaviors move, and by how much, and find that on average it points to the model rather than the behavior being steered. Steers mostly lower what the unsteered answer shows and raise a set of behaviors the model falls back on (on Qwen2.5-7B, for example, chiefly what the judge scores as refusal, sycophancy, and poeticism). Three results across 24 behaviors and ten instruction-tuned models support this, every effect read off the generated text by a language-model judge rather than off a probe. That readout matters: on Qwen2.5-7B all 24 behaviors are linearly decodable, but only 20 change what the model writes. First, on average, directions carrying no behavioral content, of a real steer's size and layer, move the same behaviors in a similar order as real steers do, clearly on five models, weakly on five. Second, much interference runs one way, so it cannot be an overlap between two directions: steering profanity makes the model toxic, while steering toxicity leaves profanity untouched. Third, with a behavior held out entirely, geometry measured on the others explains little of the interference it takes part in. The one-way structure and the failure of geometry hold on all ten models. Base models share the one-way structure, yet only one of four has refusal or sycophancy among the five behaviors steers induce most, so post-training appears to install them. Neither geometry nor random directions predicts one pair of behaviors well. For single-layer additive difference-of-means steering, this implies that disentangling behavior directions cannot by itself make steering modular.
Post-hoc safety training (RLHF, DPO) is the dominant way to align large language models, yet jailbreaks (Zou et al., 2023), fine-tuning attacks (Qi et al., 2024), and activation-space edits (Arditi et al., 2024) keep recovering the behaviors it was meant to remove. We give this fragility one geometric explanation and follow it into pretraining. We measure the safety update $\Delta = W_{safe} - W_{base}$ against the curvature of the model's capabilities (the empirical Fisher of a capability loss). Across five model families, post-hoc safety lands in a suppression regime: $\Delta$ is nearly orthogonal to the capability directions, and its small in-subspace part concentrates on a few high-curvature ones. The update is thin but sharp, a refusal gate laid over intact capabilities rather than erasure of them. A kernel-immobility lemma explains why such an update can only mask a capability, not remove it, and a little benign fine-tuning switches the gate off: 100 benign examples cut the AdvBench refusal of Qwen-2.5-7B-Instruct and Llama-3-8B-Instruct by 35 to 38 pp. Following the account into pretraining, a pretraining-checkpoint sweep of OLMo-2-1B (Team OLMo et al., 2024) shows the features that refusal attaches to emerging in a sharp transition between 1B and 63B pretraining tokens. We then use the account constructively: models trained from scratch with safety co-training spread continuously across pretraining reach 87 to 98% AdvBench refusal that the same attack erodes by only 2 to 14 pp at every scale from 410M to 6.9B, at a small cost on short-answer capability probes; a windowed schedule of equal total safety weight installs no refusal. Persistence, not timing, decides whether refusal survives. It does not keep output safer: by a harmful-output classifier, the attack raises harmful completions by 8 to 12 pp after post-hoc safety and by 11 to 27 pp after co-training.
In situ adjoint training extracts parameter gradients directly from measurement, but has so far been limited to reciprocal or restricted systems. We introduce the physical counterpart of the policy gradient theorem: a stochastic-adjoint gradient estimator that lifts these constraints by trading reciprocity for nondegenerate diffusion. As validation, we train a nonlinear resonator network, whose own dynamics supply the policy, against antagonistic temporal modulations with gradients from measured stochastic trajectories alone, without finite differences or a separate adjoint experiment.
UAV image collections contain spatially and temporally related frames, yet semantic-segmentation benchmarks commonly split them at image level. Such splitting can place samples from one acquisition in both model development and testing, obscuring transfer to a genuinely new survey. Using the 734-sample WeedyRice-RGBMS-DB, we fix a 124-image target-acquisition test set and compare two protocols with identical train, validation, and test counts: target-held-out, which excludes the target acquisition from development, and target-exposed, which admits its remaining images. SegFormer-B0 is evaluated with RGB, four-band multispectral (MS), and seven-channel RGB+MS input over two fixed-split seeds. RGB is strongest under complete acquisition holdout ($0.7317\pm0.0201$ IoU), whereas RGB+MS becomes strongest after target exposure ($0.7822\pm0.0269$). A fixed-split U-Net/ResNet18 replication confirms positive exposure gains for all three inputs, but retains RGB as the best modality under both protocols. Acquisition exposure therefore increases measured performance across both evaluated backbones, while its effect on modality ranking is architecture-dependent. A supplied-split audit reveals strong near-sequential dependence, and corruption tests show that early fusion is substantially more sensitive to RGB--MS displacement than to moderate radiometric scaling. These results support acquisition-aware same-test evaluation as a necessary complement to ordinary image-level splitting in multimodal UAV benchmarks. The code and supporting the findings of this study will be publicly released upon acceptance of the paper.
Catheterisation image processing requires segmentation models that are fast, accurate and explainable. While most of the existing studies usually focus on binary segmentation, there is a recent demand for simultaneous segmentation of multiple structures found in catheterization scenes. In this study, a dual-part MLBNet architecture is designed with multi-lateral encoder blocks and multi-head decoder branches for class-aware segmentation in cardiovascular catheterization scenes. Lateral branches in the encoder enable repeated feature extraction to learn diverse shared representations, while multiple decoder heads are used to introduce class-skewed branches that specialize in different structural properties in catheterization scenes. To analyze the performances of the dual-part MLBNet architecture, several multi-class segmentation angiogram data obtained during cardiovascular catheterization in phantom models, synthetic human-simulated aorta, and animal model are used for model training and evaluation. Results obtained showed the dual-part models could effectively separate guidewire, catheter, vessels and background pixels to their classes of memberships with high probability. The results demonstrate that all models were able to distinguish the dominant background class from foreground structures with high overall accuracy.
Temporal integration gives continuous-time recurrent networks memory, but in deep stacks it also delays bottom-up signals and attenuates top-down errors. We compensate for this delay by making each layer's bottom-up input prospective, and we develop Recursive Quadrature Filters (RQFs), biologically motivated temporal band-pass filters that are a special case of diagonal state-space models (SSMs). Starting from an energy model, we derive the RQF dynamics and show that each RQF is a band-pass filter whose learnable parameters control its tuning frequency and bandwidth. We then make each layer's bottom-up input prospective using a parameter-free two-tap update that leaves the recurrent transition and parallel scan unchanged. We extend this correction to general diagonal SSMs and show that it mitigates depth-dependent gradient attenuation when temporal gradients are truncated, i.e., spatial-only backpropagation. We evaluate the performance of RQFs, S5, and ORGaNICs (a nonlinear gated RNN with built-in attentional modulation and recurrent normalization) trained using full backpropagation through time (BPTT) and spatial-only backpropagation. Under full BPTT, the prospective variants of these models match or outperform their non-prospective controls in every configuration. A non-residual width-32 six-layer RQF achieves 96.09% accuracy on raw-audio Speech Commands with 31.9k parameters; a width-64 six-layer RQF achieves 83.36% on the 16,384-step Path-X task. We also introduce selective RQFs (sRQFs), which make each filter's input gain and, optionally, its bandwidth and tuning frequency input-dependent while retaining the parallel scan. With the input gain alone, the same six-layer model achieves 86.31% on Path-X; with all three input-dependent, a four-layer prospective sRQF solves Selective Copying (99.83%).
전선과 케이블 같은 변형 가능한 선형 객체는 얇고 변형성이 크며 자기 가림이 빈번해 분할하기 어렵고, 실제 장면에서 대규모 인스턴스 수준 주석을 확보하는 비용도 큽니다. 기존 자료는 제한된 설정에서 케이블 추적이나 의미 분할(semantic segmentation)에 집중하거나, 물리적으로 근거한 전선 변형 없이 시각적으로 그럴듯한 이미지만 생성합니다. 본 연구는 32,000장의 RGB 이미지, 인스턴스 마스크, 깊이 맵, 그리고 주석이 포함된 보완적 실세계 테스트 세트를 갖춘 전선 인스턴스 분할용 합성 데이터셋 WireSeg-32k를 제안합니다. 이 데이터셋 생성을 위해 Cosserat rod 역학과 photorealistic Isaac Sim 렌더링을 결합한 공동 시뮬레이션(co-simulation) 파이프라인 DeformX를 개발했으며, 물리적으로 그럴듯하고 접촉이 일관된 전선 형상, CAD 기반 전선 에셋, 다양한 시각적으로 근거한 장면을 만들 수 있습니다. 간단한 기준선으로 WireSeg-32k만으로 SAM3에 LoRA 미세조정을 적용하면 상용 모델 대비 실세계 mAP@75가 10.2% 향상되어, 물리적으로 근거한 합성 데이터가 실제 전선 인식으로 전이될 수 있음을 보여줍니다.
A core obstacle to alignment evaluation is evaluation awareness: capable models can tell when they are being tested rather than deployed, weakening the conclusions a safety evaluation can support. We present two techniques that make simulated alignment evaluations harder to distinguish from real deployments. Our first technique, critique refinement, spends additional inference-time compute on each simulator action: the simulator generates multiple candidate actions, refines them using feedback from an instance of the target model on how to make them more realistic, and continues the evaluation with the most deployment-like candidate. Our second technique, DISH (Deployment-Imitating SWE-Agent Harness), wraps the target in an agent harness, reducing the gap between simulated and real deployment environments in coding settings. We test the techniques on multiple target models and find that applying both yields larger realism gains than either alone. We show that automated approaches can improve the realism of alignment evaluations, and that they compare favorably to spending the extra compute on more audit attempts.
Molecular line lists disagree at a level that measurably shifts retrieved C/O, so retrievals conditioned on a single list omit a dominant systematic from their error bars. Classical marginalization re-runs a full retrieval under K opacity realizations per object. We move that cost into simulation: while generating the training set we randomize opacities (smooth multiplicative Fourier fields in log(nu) plus a discrete line-list mixture, under a deliberately broad prior informed by measured inter-list differences), so a neural posterior estimator learns an already-marginalized posterior. On perturbed held-out spectra (n=2048) a standard NPE covers nominal 90% C/O intervals 43.3% of the time against 82.4% for a compute-matched marginalized network; three-seed control ensembles reach 93.2% on clean data but 61.4% under perturbation, against 89.4% marginalized, isolating the deficit as opacity-induced (TARP joint expected-coverage deviation 0.005 against 0.035 for a single control network). An axis ablation traces the gain to the continuous field (45.4% for the discrete swaps alone, 79.3% for the field alone). In simulation, real JWST bad-pixel masks with median filling cut the marginalized ensemble's C/O coverage from 0.894 to 0.683; an ensemble retrained with mask augmentation holds 0.918 under the full deployment preprocessing (n=1024). We apply the networks to one pipeline-combined JWST/NIRSpec G395H spectrum each of 14 late-T and Y dwarfs. The recommended mask-augmented ensemble returns C/O medians between 0.24 and 0.62 and contains the Hood et al. (2024) value of 0.53 for the T8 benchmark 2MASS J0415-0935, which every fixed-opacity control network excludes. With 90% intervals spanning about 0.43 of the prior width, this is a consistency check, not a validation.
Robots that act on everyday language requests need maps that record what each object is and where it is. Existing open-vocabulary 3D mapping systems usually require a depth sensor and commit to each object the first time they see it, when the evidence is weakest. We present VOIM, a training-free system that builds open-vocabulary 3D instance maps from RGB-D or from monocular RGB video alone. VOIM accumulates soft label evidence from frozen, off-the-shelf perception models in every voxel across views, and decides labels and object instances only after that evidence has built up. On ScanNet++, VOIM achieves the highest present-class mIoU among the compared methods, 10 points above the strongest baseline. From monocular video alone, VOIM also exceeds HOV-SG and ConceptGraphs run with depth and ground-truth poses. We also conduct real-robot navigation experiments on a quadruped to show that VOIM maps are usable in the real world.
Local learning methods offer an alternative to end-to-end backpropagation, but their unstructured local objectives can produce weak feature learning in deep networks. We study whether the order of hidden-state optimization can address this limitation. We propose a boundary-first inference schedule that partitions a model into chunks, first coordinates hidden states at chunk boundaries, and then refines representations within each chunk. We instantiate this schedule in predictive coding networks (PCNs), a local-learning framework in which hidden activities and prediction errors are explicitly exposed during inference. On CIFAR-10, the resulting boundary-first predictive-coding instantiation improves accuracy over standard predictive coding by $9.77\%$ under a standard parametrization and by $5.51\%$ under a $\mu$-parametrization. Diagnostic analyses further show more non-trivial early-layer updates, lower initial-to-final CKA, and more diverse layerwise gradients, consistent with stronger feature learning. These results support boundary-first, chunk-based inference as a practical design principle for predictive-coding training and motivate its study in broader local-learning systems.
End-to-end weather forecasting systems produce skillful global gridded and station forecasts directly from raw Earth observations, replacing the numerical weather prediction pipeline, including data assimilation, at a fraction of its cost. These systems are deterministic and issue no uncertainty. Here we render the Aardvark Weather model probabilistic by attaching one stochastic mechanism to each component: learned, input-dependent noise at the observation encoder, aimed at the uncertainty inherited from the observing system, and Monte Carlo dropout in the processor, aimed at the uncertainty of the learned dynamics. The resulting nested ensemble attributes forecast spread to the two sources through a law-of-total-variance decomposition, cross-checked by withholding observation streams. Probabilistic finetuning significantly improves the mean forecast, by 11% on average across variables and lead times. The ensemble is close to calibrated against ERA5 through the medium range, is at least as accurate as the deterministic model at weather stations while beating it in relative economic value at every lead time, and remains considerably behind the operational ECMWF IFS ensemble. The encoder branch behaves as observation-driven uncertainty. Component-attributed uncertainty makes end-to-end forecasts more transparent, a step toward observation-driven digital twins of the atmosphere.
지식 그래프 상의 경로 탐색(path-finding)은 다중 홉 질문에 대한 LLM의 추론을 그라운딩하는 효과적인 방법이 되었다. 그러나 생의학 QA에는 일반 도메인 방법이 대응하도록 설계되지 않은 두 가지 뚜렷한 어려움이 있다. (i) 질의가 중간 추론 과정을 드러내지 않으며 여러 개의 유효한 경로를 통해 답할 수 있다는 점, (ii) 생의학 지식 그래프는 연결이 매우 조밀하여 경로 탐색 방법이 쉽게 잘못된 경로로 들어선다는 점이다. 이러한 문제를 해결하기 위해 본 논문은 Path-Bank에서 질의 적응형 메타 경로(meta-path)를 검색하는 경로 탐색 프레임워크인 AdaPath를 제안한다. Path-Bank는 질의의 의미와 생의학 지식 그래프의 구조를 모두 포착한다. AdaPath는 생의학 질의에서 누락된 단서를 제공하는 동시에, 다중 홉 추론 과정에서 조밀한 지식 그래프 이웃을 효과적으로 가지치기한다. 또한 다중 홉 질의를 중간 추론이 얼마나 드러나는지에 따라 계층화한 생의학 KGQA 벤치마크인 BioStrat-QA를 공개한다. 여러 생의학 KGQA 벤치마크 전반에서 AdaPath는 기준 모델(baseline)들을 일관되게 능가하며, 다중 홉 질의가 표면적 정보를 덜 드러내는 경우에도 의미 있는 경로 탐색을 유지한다. 소스 코드는 https://github.com/Jun-Hyeong-Kim/AdaPath 에서 확인할 수 있다.
대규모 비전-언어 모델(LVLM)은 여러 멀티모달 태스크에서 뛰어난 성능을 달성하지만, 종종 물체와 배경 간의 허위 상관관계(spurious correlation)를 활용하여 대상 물체와 관련된 시각적 증거보다는 문맥적 지름길(contextual shortcut)에 기반한 예측을 생성합니다. 환각(hallucination) 및 강건성 평가에 대한 관심이 높아지고 있음에도 불구하고, 기존 벤치마크는 모델의 예측이 대상 물체에 그라운딩(grounding)된 것인지 아니면 상관된 배경 단서에 의해 유도된 것인지 제어하는 데 한계가 있습니다. 본 연구에서는 LVLM에서 허위 상관관계로 인해 발생하는 오류를 구축, 벤치마킹 및 완화하기 위한 프레임워크인 PURGE(Partition-aware Unlearning for Removing spurious-correlation Generated Errors)를 소개합니다. 이 프레임워크는 다음으로 구성됩니다. (1) 예시들을 물체 관련 증거와 허위 배경 단서로 분할(partition)하여 지름길 의존성에 대한 제어된 진단을 가능하게 하는 세 가지 상호보완적 구조화 데이터 구축 전략을 개발하는 구조화된 데이터셋 구축, 그리고 (2) 이러한 파티션을 활용하여 물체 기반 추론 능력은 보존하면서 허위 물체-배경 연관성을 선택적으로 제거하는 파티션 인지 언러닝(partition-aware unlearning). 저자들은 CHAIR, POPE, Causal-HalBench, MM-SpuBench, AMBER, MMHal, Waterbirds를 포함한 다양한 벤치마크 스위트에서 비전-언어 인코더인 CLIP과 함께 LLaVA-1.6-7B, Qwen3-VL-8B-Instruct, Qwen3.5-9B를 포함한 여러 LVLM에 걸쳐 PURGE 프레임워크를 평가합니다. 실험 결과, PURGE는 대부분의 평가 환경에서 전반적인 성능을 유지하거나 향상시키는 동시에 환각 및 허위 상관관계로 인한 오류를 일관되게 줄여줌을 보여주며, 더 신뢰할 수 있는 LVLM을 위한 재사용 가능한 평가 프로토콜과 효과적인 완화 프레임워크를 모두 제공합니다.
주파수 특징(frequency feature)과 압축 불변 표현 학습은 비디오 압축 환경에서도 유지되는 딥페이크 탐지의 핵심으로 널리 간주됩니다. 우리는 블록 DCT 및 FFT 위상 스트림, 압축 수준 조건부 대역 어텐션(attention), 적대적(그래디언트 반전) 압축 불변성을 갖춘 CAFRL을 통해 이를 검증하고 통제된 부정적 결과(negative result)를 보고합니다. 모델 용량과 증강(augmentation)을 일치시킨 대조군과 사전 등록된 프로토콜에 따라 평가했을 때, 다중 품질 데이터로 학습된 기본 EfficientNet-B0는 FaceForensics++ 테스트 분할의 모든 압축 수준에서 기존 명세의 CAFRL을 능가했으며, CRF 40에서는 3.66 AUC 포인트 더 높았습니다(쌍체 비교, 단일 시드). 자체 부정적 결과에 대한 자체 감사 결과 주파수 가설에 불리하게 작용했던 네 가지 결함이 발견되었으며, 이 네 가지를 모두 수정한 사전 명시적 재테스트를 통해 이러한 성능 격차가 아키텍처 실패가 아닌 학습 레시피의 인위적 결과(recipe artifact)임을 확인했습니다. 즉, 기준선(baseline) 레시피는 일치하는 기존 배포 레시피 변형 대비 3.96 포인트를 회복했습니다. 주파수 경로는 감지할 수 있는 차이를 만들지 못했습니다. 단독으로는 판별력이 있었으나(학습 후반 독립 검증(validation) AUC 0.91-0.98), 4.0M 트렁크의 두 가지 특징 너비와 0을 포함한 데이터셋 내 압축 대비에 대한 모든 시드 통합 구간에서 이러한 융합 방식 하에서는 한계 가치(marginal value)가 없었습니다. 테스트된 단일 미학습 조작 기법에 대해서도 공정하게 맞춰진 변형들이 기본 백본보다 낮은 성능을 보였습니다. 기존 명세대로의 적대적 브랜치는 아무런 이득을 더하지 못했고 자체 조건부 추정기를 저하시켰으며, 공정한 레시피 조건에서는 테스트되지 않았습니다. 단일 패스 H.264 재인코딩 하에서의 강건성은 주파수가 아닌 데이터 다양성에서 비롯되었습니다. 실제 고정 왜곡률 인자(CRF) 변형 데이터가 합성 JPEG 증강보다 7.3 포인트 앞섰습니다(단일 실행, 겹치지 않는 신뢰 구간). 본 증거는 FaceForensics++ 계열, GAN(생성적 적대 신경망) 시대 및 단일 코덱 기반입니다. 모델 용량뿐만 아니라 학습 레시피에 대한 대조군도 일치시켜야 하며, 아키텍처 이전에 코덱 다양성을 통해 압축 강건성을 확보해야 합니다.
Randomized sketch-and-solve algorithms accelerate overconstrained $\ell_2$ regression by replacing the input with a much smaller, randomly projected problem. While standard subspace embeddings nearly preserve the regression cost, coordinate-wise accuracy is more delicate: rather than the objective value, we want the solution vector itself to be close to the optimum in $\ell_\infty$ norm. Specifically, we seek $x'\in\mathbb{R}^d$ with $\|x'-x^\star\|_\infty\leq\frac{\epsilon}{\sqrt d}\|Ax^\star-b\|_2\|A^\dagger\|_{\mathrm{op}}$, where $A\in\mathbb{R}^{n\times d}$ is the design matrix, $b\in\mathbb{R}^n$ the label vector, $x^\star$ the optimal solution and $A^\dagger$ the pseudo-inverse of $A$. Price, Song and Woodruff introduced this problem and claimed that the subsampled randomized Hadamard transform (SRHT) with $O(\epsilon^{-2}d^{1+\Theta(\sqrt{\log\log n/\log d})})$ rows achieves this guarantee; we point out a gap in the proof. Song, Ye, Yin and Zhang later claimed $O(\epsilon^{-2}d\log^3 n)$ rows, but their proof relies on an independence assumption that does not hold in general, and we exhibit an explicit instance on which it fails. To achieve a truly nearly-linear-in-$d$ row count, we introduce a new fast, dense randomized transform combining 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 whose noise is independent of the entire sketched design, exactly what Song, Ye, Yin and Zhang's argument was missing. Our sketch yields the $\ell_\infty$ guarantee with $m=O(\epsilon^{-2}d\log d)$ rows, using one Hadamard pass with padded internal dimension $N=\widetilde{O}(n+\epsilon^{-2}d^3)$. It is also efficient to apply: $(SA,Sb)$ can be computed in $O(Nd\log N)=\widetilde{O}(nd+\epsilon^{-2}d^4)$ time.