Top-conf paper digest - week of July 13-17, 2026

Selection basis

Window: first arXiv submissions dated July 13-17, 2026. The nine selected records are v1 submissions dated July 13-16; older cross-list/revision records were excluded.
Status rule: the arXiv abstract record must explicitly identify acceptance at NeurIPS, ICML, ICLR, or CVPR. Workshop-only records and papers accepted to other conferences are excluded.
Coverage rule: prioritize papers with a concrete method mechanism and a result that helps a reader decide whether to open the full paper, while keeping coverage across LLMs, agents, vision, video, ML systems, and scientific ML.
Affiliations: where the paper-level record did not expose a reliable author-to-institution mapping, that gap is stated instead of inferred.

Area	Paper	Status	Why to triage it
Vision-language auditing	Symbal	ICML 2026	Finds recurring caption errors across 1.7 million image-text pairs without access to the captioning model. 1
Video generation systems	FlashDecoder	CVPR 2026	Uses a rolling KV cache to make latent-to-pixel decoding streamable, reaching 41.55 dB PSNR at 1080p with much lower latency and memory. 2
3D generation	HIVE-3D	ICML 2026	Refines a coarse scene hierarchically instead of asking a single model to generate every detail at once. 3
Continual vision	SIKD	ICML 2026	Distills spatial co-occurrence and class-level semantic structure to reduce forgetting in incremental detection. 4
Agents / multimodal RAG	EvoGraph-R1	CVPR 2026	Lets an agent retrieve, search, edit, and answer while the multimodal hypergraph changes during reasoning. 5
Dynamic 3D vision	SPIN-4DGS	ICLR 2026	Predicts Gaussian attributes from spatiotemporal positions, improving reconstruction when motion creates large frame-to-frame displacements. 6
Embodied agents	PSC-AVDN	CVPR 2026	Adds parsing, search, confirmation, and explicit spatial memory to training-free aerial vision-and-dialog navigation. 7
Formal reasoning	Theory-Level Autoformalization	ICML 2026 Spotlight	Reframes autoformalization as building a dependency-complete theory library, not translating isolated statements. 8
Scientific ML	MetaPerch	ICML 2026	Uses recording metadata such as location and time as auxiliary supervision for bioacoustic foundation models across 17 datasets. 9

LLMs, agents, and formal reasoning

EvoGraph-R1: a knowledge graph that changes while the agent reasons

Area tag: Agents / multimodal retrieval arXiv: 2607.12764 Authors / institutions: Jiashi Lin, Changhong Jiang, Xiangru Lin, Ruifei Zhang, Xinyi Zhu, Jiyao Liu, Cheng Tang, Ye Du, Shujian Gao, Junzhi Ning, Lihao Liu, Ziyan Huang, Tianbin Li, Jin Ye, and Junjun He. The paper lists Northwestern Polytechnical University, Shanghai Artificial Intelligence Laboratory, The University of Hong Kong, Monash University, and The Chinese University of Hong Kong (Shenzhen). Peer-review status: CVPR 2026 accepted paper. 5

Problem: GraphRAG systems usually build a static graph offline and query it once. That makes it difficult to add newly found evidence, correct a bad edge, or refine retrieval after the first answer attempt.

Method: EvoGraph-R1 builds a multimodal hypergraph from textual and visual subgraphs, then treats retrieval as a Markov decision process. The agent can take four actions: GraphRetrieve, WebSearch, GraphEdit with insert/update/delete operations, or Answer. Group Relative Policy Optimization trains the policy with structural, answer, and overall outcome rewards, so the graph itself becomes part of the reasoning state.

Comparison with prior work: The comparison is against standard RAG, static GraphRAG, and RL-augmented retrieval systems. The important change is not another graph encoder; it is allowing the graph to evolve in response to the question and the evidence gathered during the run. 5

Result / takeaway: On text QA, EvoGraph-R1-7B reports 68.5 F1 on 2WikiMultiHopQA, 65.4 on HotpotQA, and 56.8 on Natural Questions, averaging 63.57. The paper reports gains over Graph-R1 of 3.5, 2.7, and 6.9 points on those three datasets. On multimodal QA it reports 43.6 on E-VQA, 42.3 on InfoSeek, and 68.6 on OK-VQA; removing web search drops E-VQA from 43.6 to 32.4 and 2WikiMultiHopQA from 68.5 to 58.9 in the ablation. The practical question is whether the extra graph-editing loop pays for itself on tasks where evidence is incomplete or changes during retrieval.

Resources: Project page. The paper's limitations are the cost of maintaining multimodal structure and the risk that extraction or editing errors propagate through later reasoning. 5

Area tag: Embodied agents / vision-language navigation arXiv: 2607.11529 Authors / institutions: Yu Qi, Hongyu Li, Shaofei Huang, Tianrui Hui, Yaxiong Wang, Lechao Cheng, Zhun Zhong, Si Liu, and Meng Wang. The paper associates authors with Hefei University of Technology, University of Macau, Jianghuai Advance Technology Center, Anhui Provincial Key Laboratory of Humanoid Robots, and Beihang University; the record does not expose a complete one-to-one mapping for every author. Peer-review status: Accepted to CVPR 2026. 7

Problem: A high-altitude UAV must resolve small targets, weak landmarks, changing scale, and ambiguous direction phrases. A single MLLM call tends to lose directional grounding and long-range spatial consistency.

Method: PSC-AVDN is a training-free three-stage pipeline. Parsing converts phrases such as clock directions and compass directions into a common angular representation. Search uses a Search Chain-of-Thought to analyze the destination, inspect multi-scale observations, build a reference grid, and localize candidates. Confirmation then checks candidate regions at finer scale. Structured Spatial Memory combines multi-scale visual observation, spatial visual memory, and structured geometric memory.

Comparison with prior work: The paper compares with supervised navigation systems and training-free GPT-4o and Qwen-VL-Max baselines. Its distinction is a decomposed search process with persistent spatial state rather than a single free-form response. 7

Result / takeaway: On the unseen-test split of ANDH, PSC-AVDN reaches 13.5 SPL, 16.4 success rate, and 28.2 GP, compared with 5.7, 6.2, and 6.7 for Qwen-VL-Max. On the unseen-test split of ANDH-Full it reaches 12.1, 14.4, and 54.5, compared with 6.4, 6.9, and 1.6. The full spatial-memory configuration is also stronger than the Qwen-VL-Max baseline on ANDH unseen validation, rising to 17.8 SPL, 22.6 success rate, and 39.2 GP. The remaining limitation is visual scale: the setting still contains tiny, weakly distinctive landmarks and complex spatial relations.

Resources: Code repository. 7

Theory-Level Autoformalization: from statements to theory libraries

Area tag: Formal reasoning / evaluation arXiv: 2607.13292 Authors / institutions: Marcus J. Min, Mike He, Zhaoyu Li, Zixuan Yi, Sharad Malik, Aarti Gupta, Xujie Si, and Osbert Bastani. The paper-level record exposes funding affiliations but not a complete author-to-institution mapping. Peer-review status: ICML 2026 Spotlight. 8

Problem: Most autoformalization work translates one natural-language statement at a time. Real formal mathematics and verification projects depend on a web of primitives, definitions, lemmas, notation, and tactics that must exist before a target theorem can be stated or checked.

Method: This position paper proposes a four-layer view: axiomatic primitives, derivative definitions, tooling infrastructure, and target theorems. It argues for theory-level benchmarks with sound equivalence checking, general-purpose models rather than only fine-tuned specialists, and a common intermediate representation that can support theoretical discovery across formal systems.

Comparison with prior work: The paper shifts the unit of evaluation from statement-level exact match or theorem completion to dependency-complete libraries. It also separates the problem of producing a formal expression from the harder problem of proving that it is semantically equivalent under the relevant background theory.

Result / takeaway: This is not a new benchmark run, so its numbers are a survey of the current landscape rather than new experimental evidence. It reports 71.4% for the best current method on layer-3 statements when layers 0-2 are already human-formalized. It also cites 118 corrected errors among 371 ProofNet problems and at least 58 corrected errors among 672 PutnamBench formalizations. The decision point for readers is whether their current pipeline can construct and verify dependencies, rather than only generate a plausible isolated statement. 8

Resources: Awesome-Autoformalization survey. The paper identifies the absence of theory-level benchmarks and the scarcity of paired data for DSLs such as SMT-LIB, TPTP, Ivy, SVA, Cypher, CodeQL, and Cedar as central open problems. 8

Vision, multimodal reasoning, and generation

Symbal: auditing systematic caption errors without the captioning model

Area tag: Vision-language auditing arXiv: 2607.15216 Authors / institutions: Maya Varma, Jean-Benoit Delbrouck, Sophie Ostmeier, Akshay Chaudhari, and Curtis Langlotz. The full paper contains Stanford and clinical-AI support information, but the arXiv record does not expose a complete author-to-institution mapping. Peer-review status: ICML 2026. 1

Problem: A captioning model may repeatedly hallucinate a particular phrase when a recurring visual feature is present. Dataset users need to find those systematic misalignments, but may not have access to the model that generated the captions.

Method: Symbal uses a structured two-stage pipeline. It splits captions into textual facts, clusters them, scores their image alignment, and summarizes the lowest-alignment cluster. It then clusters the images associated with that suspected error and summarizes the visual feature linked to it. The stages use off-the-shelf embedding, text-only, and vision-language scorers rather than internal model states.

Comparison with prior work: The benchmark compares Symbal with direct single-stage prompting of large models, including Llama 3.3 70B, Qwen2.5-VL 72B, and GPT-OSS 120B. The paper's change is to search for recurring error-feature associations over a dataset, rather than ask a model to inspect the whole dataset and name an error in one pass. 1

Result / takeaway: SymbalBench contains about 1.7 million image-text pairs across 420 dataset settings covering natural and medical images. Symbal correctly identifies the systematic misalignment in 63.8% of settings, compared with 17.1% for the strongest reported direct-prompting baseline. In real-world checks, the method found caption associations whose error rates were 3.1x, 4.6x, and 17.2x higher in specific visual contexts. Stage 2, which identifies the associated visual feature, is substantially harder than finding the text error, especially in reference-free medical settings.

Resources: Code repository. 1

HIVE-3D: hierarchical refinement for single-image scene generation

Area tag: 3D scene generation arXiv: 2607.13468 Authors / institutions: Bin Zang, Wenting Zheng, Xiaoliang Luo, Zhiyuan Fang, Shi Li, Lvchun Wang, Wei Yu, Yi Zhao, Tian Xie, Yuchi Huo, and Rengan Xie. The paper-level record does not expose a reliable author-to-institution mapping. Peer-review status: Accepted at ICML 2026. 3

Problem: Single-image 3D generation methods can produce convincing objects but struggle to represent a complete scene at useful resolution. Increasing resolution everywhere is expensive and can break global layout consistency.

Method: HIVE-3D first uses TRELLIS to produce a coarse scene. Florence-2 and SAM 2 provide 2D component segmentation, which is matched to 3D voxels to build a hierarchical scene tree. A voxel super-resolution model then refines each component from coarse to fine, with image conditioning and registration used to place the refined components back into the global scene. The pipeline uses Amodal3R for occluded regions and RANSAC for point-cloud registration.

Comparison with prior work: The baselines include TRELLIS, MIDI, SceneGen, and Gen3DSR. HIVE-3D keeps a coarse global scaffold and spends refinement capacity on scene components, so the comparison is about structured allocation of detail rather than a single larger unconditional generator. 3

Result / takeaway: On 3D-FRONT, HIVE-3D reports Chamfer distance 0.0035 versus 0.0038 for TRELLIS, F-score 84.34 versus 81.29, and runtime 36.5 seconds versus 6.3 seconds. Its gains are not uniform: IoU is 0.7449 versus TRELLIS at 0.8603. On the visual metrics table, it reports PSNR 13.39 versus 13.31 and CLIP 0.97 versus 0.95, while LPIPS is 0.33 versus 0.31. The method is therefore most compelling when component-level detail matters more than minimum runtime or every geometry metric.

Resources and limitations: A project page is listed. The pipeline depends on TRELLIS and upstream detection/segmentation; if the coarse model omits an object or creates incomplete geometry, later registration can fail. 3

SPIN-4DGS: implicit attributes for fast motion in 4D Gaussian splatting

Area tag: Dynamic 3D vision arXiv: 2607.12362 Authors / institutions: Seung-gyeom Kim, Areum Kim, Yongjae Yoo, and Sukmin Yun. The arXiv record does not expose author affiliations. Peer-review status: Accepted at ICLR 2026. 6

Problem: In 4D Gaussian Splatting, large motion between frames can make Gaussian attributes poorly learned, causing fast-moving objects to disappear or become unstable in the reconstruction.

Method: SPIN-4DGS predicts Gaussian attributes from explicitly collected spatiotemporal positions instead of directly modeling temporal displacements. A lightweight feed-forward network predicts the attributes under a rasterization reconstruction loss, sharing representations across Gaussians so that appearance and geometry remain coherent over time.

Comparison with prior work: The paper targets the failure mode of conventional 4DGS methods that optimize attributes independently across time or rely on displacement modeling. It evaluates against D3DGS and other 4D reconstruction baselines on challenging sports scenes from the CMU Panoptic dataset. 6

Result / takeaway: On the Basketball scene, SPIN-4DGS improves PSNR by 1.83 over the strongest baseline reported in the abstract, while also improving SSIM across the large-displacement sports scenes. The useful triage signal is narrow but clear: this is a motion-robustness paper for readers whose current 4DGS pipeline loses fast objects, not a general claim that implicit attribute prediction improves every dynamic-scene regime.

Resources: Project page. The arXiv record does not list a code repository. 6

SIKD: distilling object symbiosis in incremental detection

Area tag: Continual vision / object detection arXiv: 2607.13452 Authors / institutions: Mingyue Zeng, De Cheng, Zhipeng Xu, Huaijie Wang, Nannan Wang, and Xinbo Gao. The paper-level record does not expose a reliable author-to-institution mapping. Peer-review status: Accepted by ICML 2026. 4

Problem: Incremental object detectors must learn new classes without forgetting old ones. Standard class-incremental distillation can separate feature spaces too aggressively and discard useful co-occurrence and occlusion relationships between old and new objects.

Method: Symbiosis-Inspired Knowledge Distillation uses Spatial Symbiosis Distillation to preserve spatial evidence in high-overlap regions and Semantic Symbiosis Distillation to preserve class-level topology. The latter builds confidence-weighted prototypes and aligns soft ranks in the old-class logit space. A Consistent Feature Enhancement module is used only during training, so the inference graph does not carry that extra module.

Comparison with prior work: The paper compares with LwF, RILOD, SID, ERD, TLR, CL-DETR, DyQ-DETR, ACF, and DCA across COCO 2017 and DIOR incremental settings. Its distinction is to preserve both spatial co-occurrence and semantic rank structure rather than applying a single generic feature or logit distillation loss. 4

Result / takeaway: In COCO 70+10, SIKD reaches 44.3 AP, 62.9 AP50, and 47.8 AP75; AP is 1.9 points above DyQ-DETR, 3.0 above DCA, and 8.5 above CL-DETR in the reported comparison. In DIOR 10+10, it reports 69.2 overall performance versus 53.1 for CL-DETR. The ablation separates the contributions: the baseline is 41.2 AP, Spatial Symbiosis alone 42.3, Semantic Symbiosis alone 43.9, and the full method 44.3.

Resources and limitations: No code repository is listed in the arXiv record. Training cost rises to 694.169 GFLOPs versus 500 GFLOPs for the baseline, and the paper notes that its initial base-training phase can trail some prior methods even when later incremental phases are stronger. 4

FlashDecoder: streaming video decoding with a bounded cache

Area tag: Video generation systems arXiv: 2607.14898 Authors / institutions: Minguk Kang and Suha Kwak, with affiliations to Pika Labs and POSTECH. Peer-review status: CVPR 2026. 2

Problem: Latent video diffusion systems can denoise quickly but still spend substantial time and memory decoding latents into pixels. Conventional 3D convolutional decoders become expensive at high resolution and over long sequences.

Method: FlashDecoder is a pure Transformer decoder that processes one frame at a time. Each frame attends only to a fixed-size rolling KV cache of past frames, with a temporal window of two frames in the reported setup. Sequential processing supplies temporal causality without a full attention mask, and temporal upsampling is performed before spatial upsampling.

Comparison with prior work: The paper compares against convolutional decoders in Wan2.1 and Wan2.2 latent spaces, as well as OmniTokenizer, HunyuanVideo, MAGI-1, and AToken. The tradeoff is explicit: bounded temporal context and causal processing are used to keep latency and memory independent of total video length. 2

Result / takeaway: In the Wan2.2 latent space at 1080p, FlashDecoder-XL reports 41.55 dB PSNR versus 41.49 for the convolutional decoder, while decoding 3.6x to 4.7x faster with up to 11x less memory on one H100. Architecture-aware inference optimization raises the reported speedup to 12x. The caveat is that rFVD remains behind Wan2.2 and HunyuanVideo, so the strongest case is low-latency reconstruction rather than a universal quality win.

Resources: No code or project repository is listed on the arXiv record. 2

Scientific and domain-shifted ML

MetaPerch: metadata as supervision for bioacoustic foundation models

Area tag: Scientific ML / bioacoustics arXiv: 2607.14072 Authors / institutions: Mustafa Chasmai, Vincent Dumoulin, and Jenny Hamer. The arXiv record does not expose author affiliations. Peer-review status: Accepted to ICML 2026. 9

Problem: Bioacoustic foundation models trained on citizen-science recordings see strong geographic and ecological shifts. The recordings also carry metadata such as location and time, but standard supervision often ignores it.

Method: MetaPerch adds auxiliary metadata losses so the representation can use species-metadata correlations alongside the vocal signal. The paper studies nine metadata sources and tests whether those signals improve species identification and generalization to species-distribution and acoustic-domain shifts across 17 bioacoustic datasets.

Comparison with prior work: The baseline idea is vocalization-only supervision. MetaPerch asks whether metadata should be treated as a training signal rather than only as a filter or post hoc annotation, while still aiming to preserve useful acoustic representations for deployment in passive acoustic monitoring.

Result / takeaway: The abstract reports strong species-identification performance across multiple challenging domains and an empirical study covering nine metadata sources and 17 datasets, but it does not expose headline metric values. The HTML conversion was unavailable for this record, so this entry does not infer a numerical gain. Read it if your deployment domain has structured collection context that a foundation model currently discards; otherwise the missing metric detail is a reason to open the full paper before comparing it with other encoders.

Resources: No code or project link is listed on the arXiv record. 9

Reading order

Start with Symbal if you audit multimodal datasets or generated captions: its benchmark is large and its error-detection task is concrete. Read EvoGraph-R1 next for agentic retrieval, especially if graph structure must change as evidence arrives. For systems work, FlashDecoder is the clearest speed-memory tradeoff; for 3D vision, compare HIVE-3D when scene detail matters with SPIN-4DGS when fast motion is the failure mode. PSC-AVDN is the most direct read for training-free embodied reasoning, while Theory-Level Autoformalization is the conceptual paper for readers designing formal reasoning infrastructure rather than another statement-level benchmark.