Top-conf paper digest — week of June 30-July 3, 2026
Nine arXiv papers from the June 30-July 3 recent batches, grouped by vision, LLM systems, agents, robotics, and multimodal video. This issue highlights ICML/CVPR-tagged work on pixel-space 3D diffusion, distribution-wise rewards, large-kernel operators, open-world agent robustness, sub-1-bit KV caches, and task-agnostic robot pretraining.
This issue is dominated by papers that turn a practical bottleneck into a benchmarkable mechanism: faster large-kernel vision operators, less brittle agent training, lower-bit KV caches, task-agnostic robot pretraining, and visual-generation rewards that try to preserve distributional diversity instead of only optimizing single samples.
Selection basis: papers were taken from the June 30-July 3 arXiv recent batches in cs.LG, cs.CV, cs.CL, cs.AI, and cs.RO, then filtered for explicit ICML 2026 or CVPR 2026 status in the arXiv record. Workshop-only items were excluded from the main list.
| Area | Paper | Status | Why it is in this issue |
|---|---|---|---|
| Vision / 3D | PointDiT | ICML 2026 | A minimalist pixel-space diffusion transformer for monocular 3D point maps. 1 |
| Multimodal / video | DramaSR-LRM | ICML 2026 | A 532K-line long-form drama speaker-recognition benchmark plus a reasoning-model pipeline. 2 |
| Robotics | Task-agnostic VLA pretraining | ICML 2026 | A two-stage route that learns motor priors before language grounding. 3 |
| Visual generation | Distribution-wise rewards | ICML 2026 Main | RL fine-tuning for image generators using distribution-level rather than sample-level rewards. 4 |
| Efficient vision models | WBMM | ICML 2026 Spotlight | A batched-matrix-multiplication operator for large receptive fields without custom kernels. 5 |
| Computational design | Mirror Illusion Art | CVPR 2026 Highlight | Joint geometry-and-color optimization for printable mirror illusion objects. 6 |
| Language modeling | Set Diffusion | ICML 2026 | A diffusion LM design that supports flexible token sets and KV-cache updates. 7 |
| Agents | OpenAgent | ICML 2026 | A controlled setup for measuring tool-use agents under open-world shifts. 8 |
| LLM systems | GSRQ | ICML 2026 | Gain-shape residual quantization for sub-1-bit KV-cache compression. 9 |
Vision and generation
PointDiT: pixel-space diffusion for monocular geometry estimation
Area tag: Vision / 3D reconstruction
arXiv: 2607.02515
Authors / institutions: Haofei Xu, Rundi Wu, Philipp Henzler, Nikolai Kalischek, Michael Oechsle, Fabian Manhardt, Marc Pollefeys, Andreas Geiger, Federico Tombari, and Michael Niemeyer. The arXiv abstract page and HTML conversion list the authors but do not expose institutional affiliations. 1
Peer-review status: ICML 2026, confirmed in the arXiv comments field. 1
The problem is single-image geometry: recovering a dense 3D point map from one RGB image, where depth and scale are ambiguous. Prior approaches either use complex hybrid regressors with carefully designed losses or move geometry into latent spaces so that pretrained latent diffusion models can be reused. 1
PointDiT strips the pipeline down. It uses a plain ViT-style diffusion transformer directly on raw point-map patches, conditioned on image tokens from DINOv3, and trains the diffusion backbone from scratch rather than building a point-map tokenizer. The main claim is architectural: for this task, the authors argue that latent compression and hybrid losses are not necessary to beat more elaborate baselines. 1
Result / takeaway: The abstract reports sharper geometry, better behavior in ambiguous regions such as transparent objects, and performance above complex latent-diffusion alternatives, but the abstract page does not expose a single headline numeric score. Code is not listed on the arXiv page, but the comments point to a project page. 1
Optimizing visual generative models via distribution-wise rewards
Area tag: Visual generation / RL alignment
arXiv: 2607.02291
Authors / institutions: Ruihang Li, Mengde Xu, Shuyang Gu, Leigang Qu, Fuli Feng, Han Hu, and Wenjie Wang. Institutional affiliations were not visible in the arXiv abstract page or the extracted HTML text. 4
Peer-review status: ICML 2026 Main, confirmed in the arXiv comments field. 4
Most RL-style fine-tuning for image generators uses sample-wise reward models. The paper’s concern is that this can optimize individual images while damaging the generated distribution: diversity drops, modes collapse, and visual artifacts appear. 4
The proposed fix is to reward sets of samples rather than one image at a time. Because estimating distribution-wise rewards is expensive, the authors introduce a subset-replace strategy that updates only a small subset of a generated reference set. They also use RL to tune post-hoc model-merging coefficients, which is meant to reduce train-inference mismatch from SDE-based RL practice. 4
Result / takeaway: The abstract reports FID-50K improvements from 8.30 to 5.77 for SiT and 3.74 to 3.52 for EDM2, while preserving sample diversity in qualitative evaluation. No public code link is visible in the arXiv abstract page. 4
Mirror Illusion Art
Area tag: Vision / computational design
arXiv: 2607.02015
Authors / institutions: Xiaopei Zhu, Zeyuan Li, Jun Zhu, and Xiaolin Hu; the HTML conversion lists Tsinghua University, Huazhong University of Science and Technology, the IDG/McGovern Institute for Brain Research at Tsinghua, and the Chinese Institute for Brain Research. 6
Peer-review status: CVPR 2026 Highlight; the arXiv comments also mention an Efficient CVPR award. 6
The paper defines mirror illusion art as an inverse-design task: given two 2D targets, produce a printable 3D object whose direct view and mirror reflection look like different objects. Prior topology-driven and shadow-art methods require heavy manual work, optimize shape but not color, and can produce incomplete or rough geometry. 6
AutoMIA jointly optimizes shape and color. The stabilization pieces are projection-alignment component selection, position-weighted adaptive suppression, internal voxel preservation, and shape-color decoupled optimization. In plain terms, the method tries to keep the object printable and smooth while separately controlling which surface regions explain the front view, the reflected view, and the colored appearance. 6
Result / takeaway: The abstract reports successful digital and physical mirror-illusion artworks with about 76 seconds design time and 2.6 GB memory on a single RTX 3090. Code is available through the GitHub link in the abstract. 6
WBMM: windowed batch matrix multiplication for large receptive fields
Area tag: Efficient vision models / operators
arXiv: 2607.02097
Authors / institutions: Wan Song, Wei Zhou, Rui Wang, Jun Yu, Toru Kurihara, Jiajia Xu, and Shu Zhan. Institutional affiliations were not visible in the arXiv abstract page or extracted HTML text. 5
Peer-review status: ICML 2026 Spotlight, confirmed in the arXiv comments field. 5
Large-kernel depthwise convolutions improve receptive field size, but the gather-style memory access gets worse as kernel size grows. The authors argue that this makes some large-kernel accelerations counterproductive on large feature maps. 5
WBMM partitions inputs into contiguous windows and indexes a relative-position bias table to build weight matrices, then executes the operation as batched matrix multiplication. The key contrast with depthwise convolution is memory regularity: larger windows become more favorable for WBMM, while larger kernels usually degrade depthwise-convolution throughput. 5
Result / takeaway: The paper reports that WBMM with 14 x 14 windows beats 5 x 5 depthwise-convolution baselines in speed while giving a 7.8x larger per-layer receptive field. In model experiments, it reports comparable or higher accuracy on ImageNet-1K, COCO, and ADE20K with 1.31-1.88x training speedup, without specialized acceleration kernels. Code is linked from the arXiv comments. 5
Language models, agents, and systems
Set Diffusion: flexible token sets between autoregression and diffusion
Area tag: Language modeling / decoding
arXiv: 2607.01775
Authors / institutions: Marianne Arriola and Volodymyr Kuleshov. The arXiv abstract page lists the authors but does not expose affiliations, and arXiv HTML was unavailable for this paper. 7
Peer-review status: ICML 2026, confirmed in the arXiv comments field. 7
Discrete diffusion language models have improved, but they usually assume fixed-length generation and lack KV-cache support. Block diffusion moves partway toward autoregression by generating fixed-size blocks left to right, but that still limits ordering flexibility and parallelism. 7
Set Diffusion changes the unit of generation from a fixed block to flexible-position, flexible-length token sets. Its set-causal architecture supports KV-cache updates after every inference step, and the paper explicitly calls out arbitrary-order and sliding-window decoding as supported cases. 7
Result / takeaway: The abstract reports better speed-quality tradeoffs than prior diffusion language models on mathematical reasoning, summarization, and unconditional generation, plus stronger infilling than block diffusion. Code, model weights, and a blog post are linked through the project page and GitHub noted in the arXiv record. 7
Can Agents Generalize to the Open World?
Area tag: Agents / tool use
arXiv: 2607.01084
Authors / institutions: Song-Lin Lv, Weiming Wu, Rui Zhu, Zi-Jian Cheng, and Lan-Zhe Guo. Institutional affiliations were not visible in the arXiv abstract page or extracted HTML text. 8
Peer-review status: ICML 2026, confirmed in the arXiv comments field. 8
The paper attacks a familiar weakness in agent benchmarks: a model can look competent when queries, tools, schemas, and observation dynamics stay close to training conditions, then break when the environment changes. The authors formalize this as OpenAgent, with shifts in user queries, action/tool sets, observations, and task domains. 8
Their diagnostic setup uses a controlled sandbox with a four-tier shift hierarchy: perception, interaction, reasoning, and internalization. The paper reports that both supervised fine-tuning and reinforcement learning agents degrade under open-world shifts, but with different failure modes. The proposed mitigation is Perturbation-Augmented Fine-Tuning, a disturbance-based SFT intervention. 8
Result / takeaway: The abstract does not expose a single headline score, but the contribution is a benchmark framing for agent robustness rather than another closed-set tool-use leaderboard. The code is announced for release at the GitHub URL in the abstract. 8
GSRQ: gain-shape residual quantization for sub-1-bit KV cache
Area tag: LLM systems / compression
arXiv: 2607.01065
Authors / institutions: Soosung Kim, Minjae Park, Eui-Young Chung, and Jaeyong Chung. Institutional affiliations were not visible in the arXiv abstract page or extracted HTML text. 9
Peer-review status: ICML 2026, confirmed in the arXiv comments field. 9
The deployment problem is KV-cache memory: for long-context LLM inference, attention cache size grows with sequence length and batch size. The paper focuses on vector quantization, especially residual quantization, as a route to sub-1-bit storage. 9
The authors identify centroid shrinkage in standard Euclidean K-means as a high-dimensional failure mode: averaged centroids can hurt directional preservation. Their Gain-Shape K-means is presented as a drop-in replacement that separates magnitude and direction, then feeds a weighted version into a residual-quantization pipeline called GSRQ. 9
Result / takeaway: On LLaMA-3-8B, the abstract reports that at 1-bit GSRQ improves average LongBench accuracy from 11.34 to 33.54, a 22.20 percentage-point gain over VQLLM. No public code link is visible in the arXiv abstract page. 9
Reasoning LLM improves speaker recognition in long-form TV dramas
Area tag: Multimodal video / speaker recognition
arXiv: 2607.02504
Authors / institutions: Yuxuan Li, Lingxi Xie, Xinyue Huo, Jihao Qiu, Jiacheng Shao, Pengfei Chen, Jiannan Ge, Kaiwen Duan, and Qi Tian. Institutional affiliations were not visible in the arXiv abstract page or extracted HTML text. 2
Peer-review status: ICML 2026, confirmed in the arXiv comments field. 2
The paper targets speaker recognition in long-form dramas: assigning each utterance to the right character when audio, visual identity, dialogue context, and plot relationships all matter. This differs from short diarization settings because a drama can include many recurring characters, secondary roles, scene cuts, and short utterances where voice alone is weak evidence. 2
The dataset contribution is DramaSR-532K, with 532K annotated dialogue lines and more than 900 unique characters. The method contribution, DramaSR-LRM, uses a large reasoning model with multimodal tool use to combine acoustic, linguistic, and visual evidence. 2
Result / takeaway: The abstract reports that DramaSR-LRM outperforms existing baselines, especially on short utterances where acoustic biometrics are unreliable, but does not expose a headline numeric score. The authors say all data and code will be made public on the project page. 2
Robotics and embodied AI
Learning to move before learning to do: task-agnostic pretraining for VLAs
Area tag: Robotics / VLA models
arXiv: 2607.02466
Authors / institutions: Junhao Shi, Siyin Wang, Xiaopeng Yu, Li Ji, Jingjing Gong, and Xipeng Qiu; the HTML conversion lists Fudan University and Shanghai Innovation Institute. 3
Peer-review status: ICML 2026, confirmed in the arXiv comments field. 3
The paper separates two objectives that are often bundled together in VLA training: physical competence, meaning how to move, and semantic alignment, meaning what instruction to execute. The authors argue that only the second part requires language-supervised expert demonstrations. 3
Task-Agnostic Pretraining first learns motor priors from cheaper unlabeled interaction data, including off-task trajectories and autonomous robot play, using a self-supervised inverse-dynamics objective. A smaller second stage then grounds those priors in language using expert data. 3
Result / takeaway: On SIMPLER, TAP matches models trained on more than 1M expert trajectories while using far less labeled data, and it gives a 10% absolute gain over standard behavior cloning. On a real WidowX platform, TAP keeps 25% success under camera perturbations while internet-scale baselines collapse to 0%. The HTML conversion exposes a homepage, GitHub repository, and Hugging Face model collection. 3
Reading order
Open the systems papers first if you are serving long-context models: GSRQ is the most directly actionable for KV-cache compression, while WBMM is the clearest operator-level engineering paper in the set. If you work on agents, OpenAgent gives a useful failure taxonomy and Set Diffusion is the more speculative decoding direction. For embodied AI, TAP is the cleanest paper to read end to end because the hypothesis, data regime, and robot results line up tightly.
References
- 1arXiv:2607.02515
arxiv.org
- 2arXiv:2607.02504
arxiv.org
- 3arXiv:2607.02466
arxiv.org
- 4arXiv:2607.02291
arxiv.org
- 5arXiv:2607.02097
arxiv.org
- 6arXiv:2607.02015
arxiv.org
- 7arXiv:2607.01775
arxiv.org
- 8arXiv:2607.01084
arxiv.org
- 9arXiv:2607.01065
arxiv.org
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.