qwenGPU is an experimental model that compiles a SIMT GPU directly into the weights of Qwen3.5-2B. Developer Vixhal (@TheVixhal) released it on GitHub under the MIT license. The model runs real CUDA, compiled to NVIDIA PTX, inside its own forward passes. For developers, it tests one idea: exact computation inside a pretrained LLM, not beside it.
TL;DR
Size: Qwen3.5-2B (24 layers) along with 45 compiled GPU layers, 69 in total. Context: 262,144 tokens (base model).
Runs on: Apple silicon, an NVIDIA GPU or a CPU, with about 16 GB of free memory. No quantized build.
Performance: Every launch in the developer’s runs was exact; text drift is a KL of about 7e-4.
Best: The reference interpreter matched a real NVIDIA Tesla T4 on 49 of 49 kernels.
Worst: About 10 to 15 tokens per second on Apple MPS; Qwen-written CUDA passed 4 of 5.
What is qwenGPU?
qwenGPU is Qwen3.5-2B with 45 compiled transformer layers placed in front of Qwen’s own 24. None of those layers are trained. A compiler computes every weight.
How does a GPU fit inside model weights?
The GPU is a 4-lane SIMT machine. It has 15 registers per lane, a 4-stage pipeline, a 32-block ROM and 126 memory words. It uses only Qwen3.5’s own layer types: Gated DeltaNet linear attention and SwiGLU MLPs.
The trick is the Gated DeltaNet recurrent state. In the circuit’s heads, the gate is held open and the write strength is 1. Every key is one-hot. Each update overwrites exactly 1 row, and a one-hot query reads it back. So each row works as a memory cell.
That state holds the program ROM, registers and global memory. The README lists 16 register rows per lane, with 1 that always reads zero. That matches the 15 usable registers in the X post.
The GPU supports integer arithmetic, predicates, divergent branches and loops. It also supports shared memory, atomicAdd, __syncthreads() and warp shuffles.
&&
What happens during a CUDA launch?
The host plays the role of NVRTC and the CUDA driver. It compiles CUDA to PTX with clang, then lowers PTX to machine blocks. It feeds the program and input arrays to the model as tokens.
From there, the model writes the launch itself. Each EXEC(b, w) token runs block b for warp w in 1 forward pass. The VAL tokens after it carry register results. Memory writes return as MEM records. The host reads results from the recurrent state and computes none of them.
In the demo video, a dot product of [1, 2, 3] and [4, 5, 6] returns 32. The model wrote that launch in 47 tokens.
Meet qwenGPU: Qwen3.5-2B with a SIMT GPU inside its weights, 45 compiled layers in front of Qwen’s own, using its Gated DeltaNet layers as registers and memory.
It runs real CUDA in its forward passes. The host compiles the kernel to PTX and loads it, then the model writes every… pic.twitter.com/iM5IETEbIv
Not by the developer’s measure. The circuit is muted on text tokens, and Qwen’s own weights are unchanged. The merged model’s next-token distribution stays within a KL of about 7e-4 of the original.
How accurate is qwenGPU?
All numbers are self-reported, measured on Apple silicon with MPS in bf16:
Every launch was exact under --check replay.
The reference interpreter matched an NVIDIA Tesla T4 on 49 of 49 test kernels.
Normal mode answered 20 of 24 sampled chat questions with GPU-computed numbers.
Advanced mode wrote a correct kernel first try on 4 of 5 questions.
In normal mode, Qwen writes an expression like [1, 2, 3] @ [4, 5, 6] for ready-made kernels. In advanced mode, Qwen writes the CUDA kernel itself. The README warns that a 2B model’s CUDA is still unreliable.
How does qwenGPU compare with Qwenputer and Percepta?
qwenGPU uses the transformer compiler from Qwenputer by Divyansh Agrawal. Qwenputer compiles a 32-bit RISC-V CPU into Qwen3.5-4B. Percepta built a WebAssembly interpreter into a transformer with 2D attention heads in March 2026.
Zyphra has released Zyphra Hierarchical Attention (ZHA), a sparse attention method built around a tree-structured indexer. ZHA targets a known gap in sparse attention: the indexer that picks tokens still scans the whole context. Zyphra’s technical paper shows ZHA can be distilled into frozen pretrained models with about 4B tokens. On Qwen3.6-35B-A3B it stays close to dense attention on reasoning, agentic coding and 256K retrieval benchmarks.
TL;DR
Size: It is a method that adds 81,920 trainable parameters to Qwen3.6-35B-A3B (about 0.00024%). Tested up to 256K context.
Runs on: Retrofits frozen pretrained models after a 4B-token KL warmup at 32K context. Speedups come from a modeled H100 BF16 roofline. No optimized kernels released. An FP4 router variant was also tested.
Performance: Within a few points of dense attention on 7 reasoning, coding and retrieval benchmarks.
Best: LongBench v2 hits 63.4% (deep tree) vs 58.0% dense, using about 1.1% of keys.
ZHA is a sparse attention method from Zyphra that finds relevant keys by searching a tree. It keeps the full KV cache. The tree only decides which cached keys each query reads.
Fixed-state methods like SSMs and sliding windows saturate at long contexts. Fine-grained sparse methods like DeepSeek Sparse Attention score every past token with an indexer. The final attention is sparse, but the indexer still grows with the full context.
Near 1M tokens, cited work shows indexing takes 60 to 90% of layer time. ZHA makes the search itself sparse.
How does ZHA’s tree indexer work?
Building the tree
ZHA splits keys into contiguous chunks and compresses each chunk into a summary key. Summaries are grouped and compressed again, level by level. With chunk size 4, 6 levels cover 4,096 tokens per top node.
Summaries come from a small set of learned reference queries. Each reference query gives a tangent-plane estimate of a chunk’s log attention mass. The max over these planes approximates how much attention any future query would give the chunk.
How does top-p routing pick tokens?
At decode time, a query scores the top-level nodes. It keeps the smallest set holding a fraction p of estimated attention mass. It then expands only those children, repeating until it reaches raw KV entries.
Top-p makes the budget adaptive. A single-fact lookup touches few branches. A query that needs many scattered facts expands more. The first token is always kept as an attention sink.
How does ZHA perform against dense attention and MSA?
Zyphra team converted Qwen3.6-35B-A3B with 2 trees: deep (chunk 4, 6 levels) and wide (chunk 256, 1 level). Both use 16 reference queries and p = 0.95. It also trained a MiniMax Sparse Attention (MSA) indexer on 40B tokens, 10x ZHA’s budget.
Benchmark (Qwen3.6-35B-A3B)
Dense
ZHA deep
ZHA wide
MSA
AIME 2026
93.1
91.2
90.8
91.7
SWE-bench Verified
67.4
67.4
68.4
63.6
Terminal-Bench 2.1
49.9
46.7
53.3
46.1
MMLU-Pro
84.9
85.3
85.2
85.3
GPQA Diamond
84.8
85.0
85.0
83.1
RULER-13 (256K)
96.0
93.3
96.2
77.1
LongBench v2 (thinking)
58.0
63.4
62.4
49.3
Source:arXiv 2610.42112, Figure 5. All numbers self-reported by Zyphra.
The biggest gap is on retrieval. MSA drops to 77.1% on RULER-13, while both ZHA trees stay above 93%. On LongBench v2, the deep tree and MSA read similar key shares (1.1% vs 1%). The deep tree scores 14.1 points higher.
Across 28 tree geometries, RULER accuracy held steady when attending to 10 to 30% of keys. At fixed compression, deeper trees kept accuracy at higher sparsity. A 4-parameter model predicts RULER accuracy from geometry with 1.44-point RMSE.
arXiv cs.LGarXiv:2610.39316v2 Announce Type: replace
Abstract: Existing gut microbiome foundation models typically use flat taxon representations, limiting their ability to exploit known taxonomic hierarchies. We introduce TaxoMFM, a microbiome foundation model that integrates taxonomic structure into self-supervised learning. TaxoMFM jointly learns abundance patterns across taxonomic ranks and relative proportions within groups through taxonomy-aware encoding and hierarchical compositional reconstruction. We further construct TaxoGut-HPP, comprising 100,806 gut microbiome profiles from 231 source cohorts and covering 40 classification and 8 regression tasks. TaxoMFM achieves competitive performance across disease, gut health, host trait, and liver ultrasound phenotype prediction. In cross-cohort zero-shot transfer, where target cohorts are excluded from both pretraining and downstream adaptation, TaxoMFM achieves AUROCs of 0.838, 0.736, and 0.722 for colorectal cancer, type 2 diabetes, and inflammatory bowel disease, respectively. Further analyses show that TaxoMFM recovers community changes across taxonomic scales, while preserving community information and producing consistent representations under database updates and species-vocabulary shifts. Our code is available at https://anonymous.4open.science/r/TaxoMFM-F16D/.
arXiv cs.LGarXiv:2607.02834v2 Announce Type: replace
Abstract: Molecular optimization often starts from a pretrained generative model that captures a broad prior over valid molecular structures. At test time, however, the goal is not to sample from this prior, but to use a limited oracle budget to shift generation toward task-specific high-reward molecules. We study this adaptation problem for discrete diffusion models. Each online round couples several choices. The loop must decide which candidates to evaluate, how rewards become model updates, which feedback to reuse, and how far to move beyond the pretrained prior. These choices have mostly been studied in isolation, leaving open whether they complement one another, become redundant, or interfere inside a full online adaptation loop. We conduct controlled studies across six small-molecule binding-affinity tasks and three protein-fitness tasks. We find that acquisition, reward shaping, and model debiasing provide complementary routes to higher reward, especially for small molecules. Replay further stabilizes learning, while validity penalties keep small-molecule exploration on the valid molecular manifold. Together, these findings point to a practical recipe for feedback-efficient molecular optimization: online fine-tuning with acquisition, reward shaping, debiasing, replay, and validity control. This recipe outperforms offline fine-tuning and inference-time search baselines under matched oracle-call budgets and GPU-hour accounting. The gains are largest when high-reward candidates require larger shifts from the pretrained prior.
arXiv cs.LGarXiv:2610.41818v1 Announce Type: new
Abstract: Neural Architecture Search (NAS) has advanced the automated design of deep networks, but remains computationally expensive due to reliance on model training or performance predictors. We introduce DIFFNAS, a predictor-free NAS framework that leverages a discrete graph diffusion model guided by zero-cost performance proxies. Unlike prior diffusion-based NAS methods that require learned performance predictors, DIFFNAS conditions the generative process on training-free metrics such as MeCo and SWAP, enabling efficient architecture search with minimal computational overhead. Experiments on NAS-Bench-201 show that DIFFNAS finds the optimal architectures for CIFAR-10 and CIFAR-100 (94.37% and 73.51%) without a pretrained predictor, while reducing search cost by up to 17.5x compared to DiffusionNAG. On the MobileNetV3 search space, its accuracy is lower than that of DiffusionNAG. These results highlight the promise of diffusion models combined with zero-cost proxies for scalable and effective architecture search.
arXiv cs.LGarXiv:2610.41711v1 Announce Type: new
Abstract: Residual connections underpin Transformers, but they implicitly tie the residual width carried through depth to the compute width used inside each layer. We introduce Hierarchical Residuals (HiRes), a minimalist architectural change that decouples these dimensions, widening the residual state while keeping per-layer compute width fixed. HiRes groups layers into blocks: standard d-dimensional PreNorm layers operate inside each block, while a wider D-dimensional PostNorm outer residual connects blocks. Learned read/write projections bridge the two spaces at block boundaries, leaving individual Transformer layers unchanged. Across different model sizes and depths, HiRes consistently outperforms vanilla residual addition, with fitted scaling laws indicating a 1.30-fold improvement in FLOPs efficiency. On OLMoE-1B-7B pretraining for 500B tokens, HiRes achieves 0.0208 lower training loss and 1.14-point higher downstream accuracy. HiRes is also competitive with other residual expansion methods such as VWN, mHC and AttnRes with less activation memory and I/O costs. Built from only dense projections and normalizations, it requires neither specialized kernels nor infrastructure-level optimizations. HiRes thus exposes residual width as a practical scaling variable using only standard operators.
arXiv cs.LGarXiv:2610.40857v1 Announce Type: new
Abstract: Scientific foundation models (SciFMs) have achieved remarkable performance on a wide range of PDEs, including the ability to efficiently adapt to unseen data distributions with additional finetuning. As with all neural architectures, finetuning such models comes with the risk of catastrophic forgetting, whereby adapting to new tasks degrades prior knowledge of the model. We study catastrophic forgetting in Walrus-1.3B, a SciFM for continuum dynamics, over the course of a two-stage sim-to-real adaptation: first finetuning on simulated flow around a cylinder and then adapting to experimental data of the same system. We find that finetuning on the simulations rapidly degrades rollout performance on the original pretraining datasets, with one-step errors increasing up to 35x after just 20 epochs. Subsequent adaptation to real data further erodes performance on the simulated task. We characterize the potential drivers of forgetting, such as dataset dependence and task representation similarity. We then show that replaying as little as 1% simulated data during real-data finetuning recovers most of the lost simulation performance while retaining effective adaptation to the experimental dynamics. Our results identify catastrophic forgetting as a practical obstacle to sim-to-real adaptation of SciFMs and show that lightweight replay strategies significantly improve retention.
arXiv cs.CVarXiv:2610.39232v2 Announce Type: replace
Abstract: Embodied world models should not only imagine plausible futures, but also perceive the geometric, semantic, and motion structure of the evolving world. We present GenSight, a unified embodied world model that brings future video generation and dense perception into a single pretrained video diffusion transformer. GenSight uses a unified spatiotemporal representation for video, depth, embodiment segmentation, and first-frame-relative 2D flow. Built on a shared DiT backbone, GenSight uses lightweight task modulation to predict all modalities within the same architecture. All tasks are jointly optimized within a unified velocity prediction space. At inference, GenSight supports two complementary modes: multi-step future video generation with aligned dense predictions and single-step dense perception from an observed video. By unifying modality representation, model architecture, and prediction space, GenSight avoids separate modality-specific pipelines and couples appearance, geometry, semantics, and motion in a shared representation. It promotes cross-modal consistency and allows dense perception to benefit from the spatiotemporal priors learned through video generation. Trained on embodied manipulation data spanning multiple robot embodiments and human-hand interactions, GenSight demonstrates strong performance in both video generation and dense 4D perception, highlighting the effectiveness of unifying seeing and imagining for embodied world modeling.
arXiv cs.CVarXiv:2610.38346v2 Announce Type: replace
Abstract: Visual foundation models learn rich pathological morphology but lack explicit concept-level supervision. Vision--language pretraining can provide this supervision but requires costly image--text pairs. Here, we introduce GloSAGE, a semantic self-distillation framework that uses lesion profiles predicted by a pretrained visual teacher to construct structured pathology descriptions. Reliability-guided selection identifies image--text anchors, while Knowledge-Structured Pathology Text Synthesis and Evidence-Guided Negative-pair Construction account for partially observed lesions. Image--text alignment trains projection layers while retaining the frozen visual encoder; an intermediate Semantic Transfer Layer supplies features for downstream tasks. Developed on GloMillion and evaluated across seven glomerulus- and case-level cohorts, GloSAGE supported zero-shot lesion recognition, image--text retrieval, hard-negative discrimination and concept-intervention analyses. It outperformed its source visual model in most lesion-recognition tasks and improved lesion grading, while supporting competitive case-level disease classification, injury grading and baseline renal function estimation. Ablations supported the combined contributions of knowledge-structured text and evidence-guided negatives, and the intermediate transfer layer improved mean lesion-recognition ROC-AUC relative to the complete projection output. A frozen language model conditioned on GloSAGE features generated glomerular descriptions. These findings suggest that teacher-derived lesion profiles can provide annotation-efficient semantic supervision for extending a glomerular visual foundation model.
arXiv cs.CVarXiv:2610.18857v2 Announce Type: replace
Abstract: Pixel-space diffusion models provide an appealing alternative to latent-space diffusion models by directly modeling raw images without relying on a learned tokenizer. However, scaling pixel-space models has predominantly followed dense strategies, tightly coupling model capacity with active computation. In this work, we show that standard LLM-style Mixture-of-Experts (MoE) architectures can be effectively transferred to pixel-space diffusion. Building on this finding, we introduce PixelMoE, a pixel-space generative framework that sparsely scales model capacity with standard LLM-style MoE, without diffusion-specific routing mechanisms or auxiliary routing objectives. Specifically, PixelMoE combines standard Token-Choice (TC) routing with fine-grained routed experts for input-adaptive sparse computation, a shared expert for capturing common representations, and a bias-based load-balancing strategy. To complement sparse capacity scaling, PixelMoE further introduces Pixel-Space Hierarchical Flow Matching, which varies spatial resolution along the denoising trajectory from global structure formation to fine-grained detail refinement. At comparable active parameter counts, PixelMoE consistently outperforms its dense counterparts across model scales, demonstrating effective sparse capacity scaling in pixel space. On class-conditional ImageNet-1K, PixelMoE achieves FID scores of 1.47 at 256x256 and 1.53 at 512x512 without representation alignment or perceptual losses, comparing favorably with leading pixel-space and latent-space generative models. Code and pretrained models are available at https://github.com/KlingAIResearch/PixelMoE and https://huggingface.co/KlingTeam/PixelMoE, respectively.
arXiv cs.CVarXiv:2609.31670v2 Announce Type: replace
Abstract: We study rectified flows for denoising without clean training targets. For an unconditional rectified flow trained on Noise2Noise pairs, the optimal velocity field yields the minimum mean squared error (MMSE) denoiser in one step. In a jointly Gaussian model, we show that integrating the flow before applying its denoising readout increases error with integration time. The flow transports between noisy distributions, so its endpoint is not a clean-image estimate. Experiments on low-dose CT and Gaussian-corrupted natural images follow this pattern: one-step flow performs similarly to direct regression on the same pairs, whereas additional Euler steps reduce PSNR. A direct regressor using spatially separated, similar patches gives our best label-free CT result, a $1.99$\,dB gain over the noisy input. We report these results in CTDenoiser, a benchmark covering five architectures, supervised and self-supervised training, and two per-image methods. Supervised CNNs gain about $4$\,dB, while label-free performance depends strongly on how training pairs account for correlated CT noise. A channelized Hotelling observer also shows that fidelity gains need not improve lesion detectability: detectability falls at $160$\,HU despite higher PSNR and SSIM, while changes at $40$--$80$\,HU are small. Code, model definitions, and data-loading utilities will be released publicly; the repository link is withheld for anonymous review.
arXiv cs.CVarXiv:2511.11051v4 Announce Type: replace
Abstract: Low-Rank Adaptation (LoRA) fusion enables the composition of subject and style representations for controllable generation without retraining. However, existing approaches primarily operate through weight-level merging, without explicitly modeling how independently trained LoRAs interact in the shared parameter space. We adopt a geometric perspective on LoRA fusion, interpreting content and style LoRAs as occupying overlapping, non-orthogonal low-rank subspaces, where such overlap can lead to conflicting parameter updates that affect generation quality. This observation motivates us to reformulate LoRA fusion not merely as parameter combination, but as a problem of controlling how updates from overlapping subspaces are combined. Based on this insight, we propose Null Space Projection LoRA (NP-LoRA), a training-free framework that employs projection as a fusion operator to explicitly modulate cross-LoRA interactions. Specifically, NP-LoRA uses principal directions of the style LoRA to define a projection subspace and projects the content LoRA onto the complementary subspace (i.e., the null space of the style LoRA), suppressing interference along dominant style directions while preserving complementary information. To avoid the overly aggressive suppression of hard projection, we further formulate soft projection as a regularized optimization problem that balances content preservation against style-subspace suppression. This objective admits a closed-form solution, yielding a projection operator controlled by a single parameter that continuously interpolates between linear merging and hard projection. Extensive experiments across multiple pretrained LoRA pairs show that NP-LoRA achieves more balanced content-style composition compared to strong baselines, without requiring retraining. Code is available at https://github.com/blackzebra7/NP-LoRA.
arXiv cs.CVarXiv:2303.16570v3 Announce Type: replace
Abstract: We present point2vec, a self-supervised method for learning point cloud representations through masked prediction of a teacher network's latent features. Building on data2vec, we investigate a limitation of applying this approach to point clouds: positional information associated with masked patches reveals the overall object shape to the student, weakening the learning task even under heavy masking. point2vec addresses this issue by passing only visible patches to the student encoder and introducing a shallow decoder to predict representations for masked patches. The teacher processes the complete point cloud to provide contextualized training targets. These targets capture global context through the teacher's self-attention, providing supervision beyond the local geometry of individual patches. This design conceals masked-patch positions from the student and reduces its computational cost during pre-training. Our ablations show that the decoder alone provides only part of the improvement; concealing positional information is key. After pre-training on ShapeNet, point2vec outperforms the evaluated self-supervised baselines on shape classification on ModelNet40 and ScanObjectNN and few-shot classification on ModelNet40, while achieving competitive part segmentation results on ShapeNetPart.
arXiv cs.CVarXiv:2610.40662v1 Announce Type: cross
Abstract: HER2 status decides who receives HER2-targeted therapy. Pathologists score HER2 protein by immunohistochemistry (IHC); equivocal IHC 2+ tumours, their hardest calls, wait for in situ hybridisation (ISH). We built a HER2 reader for routine H&E slides, owned end to end, as a baseline for pathologists. A 22M-parameter ViT-S/16 tile encoder was trained on public TCGA slides by self-supervision, distillation from H-optimus-0 and two rounds of JEPA refinement; five cross-validated attention-based slide classifiers average their probabilities; and an operating point is fixed on training data. Compared under one paired ten-seed protocol, readers on our encoder rank first of 7 on external validation from TCGA-BRCA to Yale-HER2 (AUC 0.812), level with readers on encoders 4 and 50 times larger. We also rebuilt PathBench's encoder benchmark on public cohorts; our replica reproduces its published order of reference encoders, and under it our encoder ranks first of 7 on TCGA-BRCA (0.689) and sits level with H-optimus-0 and Phikon on Yale-HER2 (0.897). On three external hospitals our deployed reader reaches AUC 0.82 on Yale-HER2 (about 0.81 with training slides balanced for tissue size), 0.75 on HEROHE and 0.63 on BCNB biopsies, and among HEROHE's 85 IHC 2+ cases it ranks ISH-amplified above non-amplified tumours at 0.76, best among our TCGA-trained readers. Per-laboratory score normalisation and a laboratory's own slides improve it further on a new site. Without a GPU our reader needs 5-6 s per slide. It serves as a research-use pre-read in a pathology platform, where it proposes a probability and a call beside IHC and helps prioritise ISH for equivocal cases. Weights, per-slide predictions and the data behind every figure are public at https://huggingface.co/cloudkites/her2-reader.
arXiv cs.CVarXiv:2610.42136v1 Announce Type: new
Abstract: Dense 3D tracking from monocular video requires establishing point correspondences over time, estimating their depth, and inferring their trajectories during occlusion. Existing methods typically learn these tasks jointly, a demanding objective given the scarcity of 4D supervision. We show that strong pretrained 2D trackers and 3D geometry models provide an effective foundation: lifted trajectories, with occluded segments simply interpolated, already rival dedicated 3D trackers using the same geometry. However, lifted trajectories inherit correspondence and depth errors, and under occlusion, their depth comes from the occluder rather than the tracked point. We introduce ReLift, an efficient feed-forward framework that recasts dense 3D tracking as geometric trajectory repair and completion. At its core, a trajectory corrector alternates attention along and across trajectories to preserve reliable observations, repair corrupted positions, and complete missing motion. Using only geometric inputs allows us to train directly on 3D trajectories with simulated lifting errors, decoupling training from images and from any specific tracker or geometry model. Concentrating 4D supervision on trajectory refinement yields a reusable model that consistently improves lifted trajectories across trackers and geometry providers without retraining, so it benefits directly from future advances in either component. On WorldTrack, ReLift achieves state-of-the-art mean accuracy, improving on the strongest baselines by 2.1 APD on all points and 3.0 on dynamic points, with faster configurations remaining competitive with leading 3D trackers while running faster than 4RC. Visit our supplementry page for more examples and code.
arXiv cs.CVarXiv:2610.42092v1 Announce Type: new
Abstract: Self-supervised learning relies on asymmetry between what a model observes and what it must explain. Recent methods learn 3D geometry from unposed images by predicting cameras and scene structure, then rendering held-out views. To avoid trivial solutions, they impose this asymmetry in the input space, restricting how views interact through separate branches or masked attention. We revisit this design with geometric accuracy, rather than rendering quality, as the main objective. When a held-out view lies between context views, the context already constrains the geometry it observes. We therefore impose the asymmetry in the 3D representation space, excluding held-out views only from the reconstructed scene while a shared backbone uses all views for correspondence. This improves depth and pose estimation over input-space designs. We further exploit view density: sparse and dense observations depict the same scene but provide different amounts of correspondence. Aligning sparse-context features with dense-context features supplies a geometric training signal alongside rendering, without a pretrained geometric teacher. We additionally align features with a frozen vision foundation model, which brings further gains. None of these signals requires geometric annotation, and the resulting model improves geometry estimation over prior self-supervised methods. Controlled experiments isolate the contribution of each component, and self-supervised pretraining also improves subsequent supervised training.
arXiv cs.CVarXiv:2610.41814v1 Announce Type: new
Abstract: Vision Transformers represent each image patch through a canonical token produced under one observation condition. That token is a powerful interface, but it is still one realized state of the patch. We introduce PMTT, which establishes two foundations: a patch representation object can extend from one canonical state to a canonical-centered Representation State Family, and identity-preserving variation opens Patch-Level Local Feature Space. We develop four parts of this representation program. Patch Information Acquisition and Dense Orbital Views expose lawful same-patch states. Visible-to-Hidden prediction tests whether observed states reveal structure beyond themselves. Intrinsic analysis identifies low-dimensional, energy-concentrated, patch-specific, predictive geometry. Representation Organization separates acquiring more information from organizing that information, with an analytic Manifold as the central structured representation. We realize PMTT on frozen DINOv3 S/B/L/H+ backbones without neural post-training of the representation. Across two independent 4,096-anchor cohorts, Manifold explains 70.8-79.6% of Hidden variation, compared with 55.0-72.3% for directly Observed states, and wins on more than 95% of anchors at every scale. Effective rank falls from about 2.5 to 2.0 while dominant energy concentration rises with backbone scale. Across six visual capabilities and four scales, Observed exceeds the canonical Center in 21 of 24 conditions, Manifold exceeds Center in 23, and Manifold exceeds Observed in 16. The strongest current consequence appears in physical correspondence, where MegaDepth satisfies Center < Observed < Manifold at every scale. These results show that pretrained visual backbones contain richer patch information than the standard single-token interface directly exposes. Code and experimental assets are available at https://github.com/zixiongwang-ai/PMTT.
arXiv cs.CVarXiv:2610.41763v1 Announce Type: new
Abstract: Reliable histopathology segmentation requires models that rely on pathology-relevant structures rather than spurious background correlations. In this work, we argue that improving architectures alone is insufficient if models continue to rely on spurious correlations rather than pathology-relevant features. We introduce a causality-inspired evaluation framework based on controlled interventions that explicitly probes model sensitivity to foreground tissue structure and background/domain variation. Building on this, we propose the Causal Reliance Score (CRS), a metric that quantifies the relative sensitivity of model predictions to foreground perturbation versus background/domain variation under controlled interventions. To support this analysis, we design augmentation strategies that act as implicit counterfactual interventions during training, encouraging invariance to nuisance factors while preserving morphology. We demonstrate how these interventions reveal behavioural differences between models that are not captured by standard metrics such as Dice or IoU. While we instantiate our framework using a self-supervised transformer (CAUSE-ViT), the primary contribution of this work is not a new architecture, but a principled intervention-based methodology and evaluation metric for auditing model behaviour in medical image segmentation. Our results suggest that evaluating model behaviour through counterfactual interventions is essential for developing reliable and clinically robust medical image segmentation systems.
arXiv cs.CVarXiv:2610.41728v1 Announce Type: new
Abstract: Multitemporal synthetic aperture radar (SAR) imagery provides continuous, cloud-free observations of crop phenology for agricultural monitoring in cloudy and rainy regions, yet high-accuracy classification typically requires extensive labeled samples. Although self-supervised pre-training mitigates label dependency in remote sensing, existing frameworks primarily target optical imagery or generic spatial SAR scenes, underutilizing temporal backscatter dynamics and polarimetric scattering mechanisms. We propose SAR-TS-BERT, a physics-guided BERT-style self-supervised pre-training framework for SAR time-series crop classification. SAR-TS-BERT integrates backscatter intensities and polarimetric ratio features (VV, VH, and VV/VH) with dual-polarimetric Cloude-Pottier decomposition parameters ($H/\alpha/A$) into a six-band input representation through two core designs: 1) a physics-guided span masking strategy (SpanPhysics) that adaptively selects contiguous masked spans based on temporal backscatter change rates to prioritize key phenological transitions; and 2) a dual-branch cross-modal fusion encoder that models backscatter and polarimetric decomposition features through separate Transformer branches followed by bidirectional cross-attention fusion. On independent test sets in Estonia (EuroCropsML) and the USA (California CDL), SAR-TS-BERT achieves overall accuracies (OAs) of 96.78% and 87.29%, respectively, and attains a mean monthly OA of 99.37% on Chongming Island, Shanghai. On the Estonia benchmark, pre-training improves OA and Macro F1 by 1.17 and 1.59 percentage points over training from scratch under full supervision, and by 2.58 and 5.19 percentage points using only 1% labeled data. Incorporating SAR scattering physics into temporal self-supervised pre-training effectively enhances crop classification accuracy, parcel-level mapping consistency, and few-shot transferability.
arXiv cs.CVarXiv:2610.41673v1 Announce Type: new
Abstract: Pose estimation is a computer vision task used to estimate a skeleton of dynamic systems to predict future movements. Most of the research in this direction is based on a supervised learning approach which requires a massive amount of labeled datasets. In this paper, a self-supervised three-stage model based on a contrastive learning approach is introduced for estimating a skeleton of dynamic construction machines; such as excavators without using any labeled images for the first stage. The whole model structure is divided into three stages: the pre-train stage using the SimCLR contrastive approach, and two fine-tuning stages for the transfer learning and the downstream task. The model can leverage the features and learn from a huge unlabeled dataset called ACID to two small datasets generated from NVIDIA Isaac and MATLAB Simscape simulators as well as transfer the knowledge to a smaller dataset with a ratio of 3.5% from the original ACID dataset. The results show that the proposed approach can improve the accuracy of pose estimation for heavy construction machines in real images by 11% and 13% in comparison to the normal self-supervised approach with two backbones ResNet-50 and HRNet-W32, respectively.
