Top-conf paper digest - week of July 27-31, 2026
Six first-submission arXiv papers from July 27–31, spanning multimodal systems, vision, flow models, graph transfer, topological learning, and tensor programming, with methods, results, status, and resources.
Selection basis
- Window: first arXiv submissions dated July 27–31, 2026. The Monday, August 3 batch is outside this issue.
- Status rule: explicit acceptance or publication language is preserved. A record that only lists a conference in its journal-reference or comment field is labeled as a preprint.
- Set: six records with explicit ICML, ICLR, or CVPR 2026 affiliation in the arXiv record. No qualifying NeurIPS 2026 record surfaced in the verified window. Workshop-only papers were excluded.
- Coverage: multimodal systems and vision, generative modeling, graph learning, and tensor programming. Numeric results are included when the paper exposes them; missing affiliation or resource links are stated rather than inferred.
| Area | Paper | Status | Reading signal |
|---|---|---|---|
| Multimodal systems | MODUS | Accepted to ICML 2026 | One decoder-only model handles text, images, geometry, segmentation, and representations. |
| Vision | SCDT | Accepted by CVPR 2026 | A single denoising tracker adapts to missing and complete RGBT inputs. |
| Generative modeling | SGFlow | Preprint; arXiv comment says ICLR 2026 | Stop-gradient dynamics learn flow maps without nested differentiation or an inverse model. |
| Graph learning | Same Graph Cross-Task Transfer | Preprint; arXiv lists journal reference ICML 2026 | Homophily predicts when node classification transfers to link prediction. |
| Graph learning | TopoFormer | Preprint; arXiv lists journal reference ICLR 2026 | Overlapping topological slices become ordered tokens for a Transformer. |
| ML systems | einx | Published at ICLR 2026 (oral) | A declarative notation makes vectorization and axis semantics explicit. |
Multimodal systems and vision
MODUS: a decoder-only model for any-to-any modalities
Area tag: Multimodal systems
arXiv: 2607.25948
Authors / institutions: Mingqiao Ye, Zhaochong An, Zhitong Gao, Xian Liu, François Fleuret, Chuan Li, Amir Zadeh, Serge Belongie, Afshin Dehghan, Jesse Allardice, David Mizrahi, Oğuzhan Fatih Kar, Roman Bachmann, and Amir Zamir. The arXiv record does not expose a complete author-to-institution mapping.
Peer-review status: Accepted to ICML 2026. The v1 record was submitted on July 28, 2026. 1
Problem: Any-to-any models should predict one modality from any combination of others, but most systems use encoder-decoder or diffusion designs trained from scratch. That makes it harder to reuse a strong pretrained decoder-only language model and often creates modality-specific heads, losses, or task pipelines.
Method: MODUS starts from the pretrained BAGEL-7B Mixture-of-Transformers architecture and uses two experts inside one autoregressive decoder: a 1D expert for sequential or discrete modalities and a 2D expert for spatial or continuous modalities. Text, RGB images, depth, surface normals, segmentation, edges, grounding boxes, and learned visual representations share one token sequence. The training objective combines next-token prediction for 1D modalities with flow matching for 2D modalities. A three-stage curriculum adds 1D data, then 2D data, then more conditioning modalities per example. 1
Comparison with prior work: The paper compares one MODUS checkpoint with specialist models such as DeepSeek-VL2, FLUX.1-dev, DepthAnything2, Marigold, GroundingDINO, and DINOv2; encoder-decoder systems such as 4M-21 and Unified-IO 2; diffusion models such as OneDiffusion; and decoder-only models including Bagel, Janus-Pro, and GPT-4o. The relevant comparison is architectural: MODUS keeps a single decoder-only model and removes separate task heads instead of assembling a collection of specialist pipelines.
Results / takeaway: On the paper's zero-shot table, MODUS reports 51.1 on MMMU, 0.81 on GenEval, 0.285 on DIODE, 19.92 on NYUv2, 54.5 on RefCOCOval, and 77.9 / 92.5 on ImageNet retrieval. It is close to specialist systems on several tasks rather than winning every column: FLUX.1-dev scores 0.82 on GenEval, GroundingDINO scores 50.4 on RefCOCOval, and DINOv2 scores 82.1 / 93.9 on ImageNet retrieval. The model also supports chained generation, such as RGB → depth → surface normal, and cross-modal self-verification; the latter rises from 0.81 for the baseline to 0.84 with VQA plus grounding verification. 1
Resources and limitations: The paper says it releases code, two checkpoints, and the 29-million-sample Modus-Dataset. Project page and materials. The current modality set is representative rather than exhaustive; adding audio or 3D is left to future dataset and tokenization work. Chained generation adds decoding steps and therefore latency.
SCDT: one RGBT tracker for missing and complete inputs
Area tag: Vision / multimodal tracking
arXiv: 2607.24701
Authors / institutions: Andong Lu, Ziyi Zha, Jiandong Jin, Shihao Li, Chenglong Li, Jin Tang, and Bin Luo. Lu, Zha, Jin, and Tang are from the School of Computer Science and Technology, Anhui University; Li and Luo are from the School of Artificial Intelligence, Anhui University.
Peer-review status: Accepted by CVPR 2026. The v1 record was submitted on July 27, 2026. 2
Problem: RGBT trackers lose performance when the visible or thermal stream disappears. Existing approaches often reconstruct the missing modality from the current frame, or switch between separate branches for missing and complete inputs. Those choices can produce spatially biased or temporally inconsistent features.
Method: The Spatio-Temporal Conditional Denoising Transformer (SCDT) treats feature reconstruction as a conditional denoising process. Current-frame spatial cues, short-term temporal cues from recent frames, and long-term cues that encode modality evolution condition a denoiser. Cross-attention handles short-term context; FiLM-style scale and shift use the long-term token. A noise-modulated adaptation mechanism changes the denoising strength according to modality availability, so one architecture handles both missing and complete cases. The objective combines reconstruction, alignment, and tracking losses, with reconstruction active for missing inputs and alignment active for complete inputs. 2
Comparison with prior work: The evaluation includes FlexTrack, IPL, STTrack, SUTrack, AINet, CAFormer, OneTracker, SDSTrack, UnTrack, ViPT, TBSI, and other recent trackers. SCDT's distinction is not a new sensor or a separate missing-modality branch; it is the use of temporal conditioning and a single noise-controlled denoiser across both operating conditions.
Results / takeaway: On complete-modality benchmarks, SCDT reports 77.4 / 61.0 on LasHeR, 93.1 / 69.6 on RGBT234, and 93.6 / 78.9 on VTUAV for the reported precision and success metrics. On missing-modality versions, it reports 69.3 / 54.4 on LasHeR-Miss, 88.1 / 64.3 on RGBT234-Miss, and 84.1 / 69.6 on VTUAV-Miss. The paper reports a 4.2-point precision and 2.1-point success gain over FlexTrack on LasHeR-Miss, and calls 88.1 / 64.3 on RGBT234-Miss a new state of the art. The ablation also gives a practical signal: four denoising layers beat both two and six layers on LasHeR-Miss, suggesting that deeper refinement can start to oversmooth. 2
Resources and limitations: Code repository. The paper's results are on three RGBT tracking datasets and their missing-modality variants; the central evidence does not establish performance on other sensor combinations.
Generative modeling
SGFlow: learn flow maps with stop-gradient dynamics
Area tag: Diffusion and flow models
arXiv: 2607.26398
Authors / institutions: Mark Goldstein, Flatiron Institute; Anshuk Uppal, Technical University of Denmark; Raghav Singhal, Aahlad Puli, and Rajesh Ranganath, New York University. 3
Peer-review status: Preprint; the arXiv comment says “ICLR 2026.” The v1 record was submitted on July 29, 2026. 3
Problem: Diffusion and flow models are easy to train with regression losses, but sampling requires numerical integration. Consistency-style flow-map methods can reduce the number of sampling steps, yet some require an inverse model, backpropagation through nested model calls, or a guarantee that the desired ODE flow map is actually a solution of the training objective.
Method: SGFlow, short for StopGrad Flow, learns the flow map and the implied velocity together. It uses stop-gradient estimates inside a non-conservative vector-flow objective and computes the needed derivatives with Jacobian-vector products rather than explicit Jacobian matrices. The paper proves that the true flow map is a stationary point of the stop-gradient dynamics under its stated conditions. The practical target is a multistep, simulation-free learner that follows the ODE without requiring the inverse during training. 3
Comparison with prior work: The CIFAR-10 comparison covers Flow Matching, Meanflow, and Lagrangian flow-map matching; the related-work table also contrasts consistency distillation, consistency training, consistency trajectory models, ESD, and PSD. SGFlow is the only row in the reported comparison marked simultaneously as multistep, ODE-following, simulation-free, regression-based, inverse-free, theoretically guaranteed at the optimum, and free of nested differentiation.
Results / takeaway: On unconditional CIFAR-10 generation, the reported FID scores are 37.29 / 4.00 / 3.87 / 3.85 for SGFlow at 1 / 10 / 50 / 100 sampling steps. Meanflow is strongest at one step with 8.67, while Lagrangian matching is strongest at 50 and 100 steps with 3.37 and 3.20. SGFlow's best result is 4.00 at 10 steps, ahead of Flow Matching at 84.39 and Lagrangian matching at 4.08 at the same step count. The method's signal is therefore not a universal FID win; it is the combination of competitive few-step quality and a stationary-point guarantee. 3
Resources and limitations: No external code or project URL is exposed in the verified record. The experiment is limited to CIFAR-10 and a modified diffusion U-Net; the authors say compute constraints prevented a broad architecture study. The paper also notes that diffusion-style architectures may not be optimal for flow-map learning.
Graph learning
Same Graph Cross-Task Transfer: when node labels help link prediction
Area tag: Graph transfer learning
arXiv: 2607.28525
Authors / institutions: Neelam Akula, Surbhi Kumar, Murat Kantarcioglu, and Baris Coskunuzer. The arXiv HTML record does not expose author-to-institution mappings.
Peer-review status: Preprint; the arXiv record lists journal reference ICML 2026 without explicit acceptance wording. The v1 record was submitted on July 30, 2026. 4
Problem: A graph often supports several tasks, such as node classification (NC) and link prediction (LP), but transfer studies can leak evaluated edges or change the negative-sampling rules between methods. Without fixed splits and a shared message-passing graph, a reported transfer gain may reflect the protocol rather than the representation.
Method: The paper fixes node splits, edge splits, and LP negatives, while removing evaluated edges from the message-passing graph. It tests warm starts, embedding transfer, a two-view encoder with InfoNCE-style alignment, and joint training under a weighted NC-plus-LP loss. The CoTask Score (CTS) summarizes relative utility when one shared encoder must serve both tasks. Experiments use GCN, GraphSAGE, and GPS across 11 node-level benchmarks covering homophilic, heterophilic, and structure-dominant graphs. 4
Comparison with prior work: The main comparison is between transfer regimes and single-task baselines under the same leakage-free protocol. It is less a new GNN backbone than a correction to how cross-task transfer is measured. The paper asks which direction transfers, and which graph statistics predict the answer, instead of treating NC → LP and LP → NC as symmetric.
Results / takeaway: NC → LP is consistently beneficial on homophilic graphs, while LP → NC can degrade accuracy under naive reuse. The paper's selected examples show a 12.3-point best gain over the GCN reference on Cora and 13.2 points on Citeseer for NC → LP; on PubMed, one GraphSAGE embedding-transfer setting is 4.3 points worse than its reference. For the reverse direction, embedding replacement is 21.4 points worse on PubMed, but joint transfer reaches 11.9 points on Texas and 13.7 points on Brazil in the cited examples. The reported Pearson correlations make the protocol lesson concrete: NC → LP gains correlate strongly with homophily, with coefficients up to 0.904, whereas LP → NC is more dependent on the structure-dominant regime and the transfer method. 4
Resources and limitations: Code repository. The study is limited to same-graph NC and LP transfer with three backbones. Cross-graph transfer, newer sequence-inspired graph architectures, and other tasks remain untested.
TopoFormer: turn graph topology into Transformer tokens
Area tag: Topological graph representation learning
arXiv: 2607.28259
Authors / institutions: Md Joshem Uddin, University of Texas at Dallas; Astrit Tola, Florida State University; Cuneyt Gurcan Akcora, University of Central Florida; and Baris Coskunuzer, University of Texas at Dallas.
Peer-review status: Preprint; the arXiv record lists journal reference ICLR 2026 without explicit acceptance wording. The v1 record was submitted on July 30, 2026. 5
Problem: Persistent-homology pipelines summarize graph structure with persistence diagrams and then vectorize them. On graphs, cumulative filtrations can saturate early, and the vectorization step can discard the order in which local and global structure appears. The authors want topological signals that a standard attention model can consume without the full persistence-diagram pipeline.
Method: Topo-Scan creates overlapping graph slices from a scalar filtration rather than one cumulative sequence. For each slice it records connected components, cycles, node count, and edge count, producing an ordered token sequence across resolutions. A Transformer with positional encoding processes those tokens. The sliding-window construction is intended to preserve late-emerging structure and remain parallelizable, while the low-dimensional token keeps the representation lightweight. 5
Comparison with prior work: Graph-classification comparisons include GCN, DGCNN, DiffPool, ECC, GIN, GraphSAGE, several topological methods such as PersLay and DMP, augmentation and contrastive methods, and the graph-kernel method DASP. Direct ablations compare TopoFormer with PH-MLP and PH-TR under degree, Ollivier–Ricci, and HKS filtrations. The paper's claim is specifically about preserving ordered topological evolution, not merely adding another graph feature vector.
Results / takeaway: On eight graph-classification benchmarks, TopoFormer is reported as best or second-best on 7/8, with average deviation 0.5 and average rank 1.5. Selected accuracies include 94.68 ± 4.30 on MUTAG, 77.64 ± 3.64 on PROTEINS, 78.90 ± 3.31 on IMDB-B, and 91.50 ± 1.89 on REDDIT-B. On OGBG-MOLHIV, TopoFormer* reports 78.19 ± 0.19 ROC AUC, within about two points of Graphormer according to the paper. On seven MoleculeNet tasks, TopoFormer* reports 96.5 ± 0.6 ROC AUC on ClinTox and 95.9 ± 0.3 on BACE, with the best reported scores on ToxCast, ClinTox, and BACE. 5
Resources and limitations: No code or project URL is listed in the verified record. The method targets graph-level tasks, uses only H0/H1 homology and a fixed clique complex, and evaluates a small set of standard filtrations. Node- and edge-level tasks, temporal graphs, and citation graphs are left for future work.
ML systems
einx: make tensor vectorization explicit
Area tag: Tensor programming / ML systems
arXiv: 2607.27987
Authors / institutions: Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, and Michael Arens, all from Fraunhofer IOSB.
Peer-review status: Published at ICLR 2026 (oral). The v1 record was submitted on July 30, 2026. 6
Problem: Numpy-like tensor APIs accumulate special cases for broadcasting, reshaping, indexing, and reductions. That makes code hard to read and makes shape errors easy to introduce. einsum and einops improve notation for important subsets of operations, but neither is presented as a universal language for arbitrary tensor operations.
Method: einx separates an elementary operation, such as
dot, sum, add, or get_at, from its vectorization over axes. Its declarative expressions use brackets for the axes consumed by the elementary operation, parentheses for flattened axes, plus signs for concatenation, and ellipses for generalized expansion. The implementation parses the expression, solves axis constraints, builds a computation graph, generates backend code, and caches the compiled function. A multi-head attention implementation can be written in three einx lines, with the operation axes and softmax axis visible in the expression. 6Comparison with prior work: einsum handles a narrow set of contractions and reductions; einops extends the notation to reshaping, broadcasting, and reductions but remains operation-limited. Named-tensor systems such as xarray and torchdim track symbolic axes but still require additional axis management. einx's proposed difference is a single pointful notation that maps both the elementary operation and the vectorization explicitly, while integrating with existing tensor frameworks.
Results / takeaway: On the Numpy backend, first-call compilation takes 6.8 ± 1.3 ms for a transpose-like identity, 9.3 ± 2.4 ms for a dot product, and 23.5 ± 3.1 ms for a larger broadcast-and-add expression. Cached calls take 0.058 ± 0.001 ms, 0.070 ± 0.007 ms, and 0.077 ± 0.003 ms respectively. The engineering takeaway is concrete: einx trades a small first-call compilation cost for readable shape semantics, then relies on caching; the paper says JAX JIT can remove the remaining footprint in compiled workloads. 6
Resources and limitations: GitHub repository. The implementation generates Python code with
exec, so initialization and cache lookup are part of the runtime design. Some concatenation cases are supported only for einx.id, and several operations can have more than one valid decomposition into elementary operations.Reading order
Open MODUS first if your work needs one model to move between language, image, geometry, and representation spaces. Read SCDT if missing sensor streams are the operational failure mode. SGFlow is the theory-and-sampler paper for readers reducing diffusion steps without adding inverse or nested-gradient machinery. For graph practitioners, Same Graph Cross-Task Transfer gives the sharper protocol lesson, while TopoFormer gives the representation-design lesson. Open einx when tensor shape manipulation, rather than model capacity, is the part of your stack that keeps breaking.
References
- 1
- 2
- 3
- 4
- 5
- 6
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.