Top-conf paper digest — week of August 3–7, 2026

Eight first-submission arXiv records dated August 3–7 carry explicit ICLR, ICML, or CVPR 2026 acceptance or publication language. They range from human-centered evaluation and probabilistic LLM programming to stochastic symmetry discovery, event-based restoration, robustness, attention dynamics, image restoration, and visual identity.

A useful way to read the batch is to ask where information is lost. Aggregate-then-Calibrate keeps human order and model scale; PPDL keeps a distribution over execution traces; LieStoNet keeps the algebraic relations that define a symmetry. In each case, the reported gain comes from adding an explicit structural constraint before the final prediction, not from interpreting an opaque output after the fact.

Area tag: Human-centered assessment arXiv: 2608.02455 Authors / institutions: Zejun Xie and Desheng Zhang, Rutgers University; Xintong Li, Renmin University of China; Guang Wang, Florida State University. 1 Peer-review status: Accepted by ICLR 2026. The v1 record was submitted on August 3, 2026. 1

Problem: Human-centered assessments often have comparative judgments but no verifiable ground-truth score. Human-only methods can mix annotators with different expertise and rating scales; model-only scores are consistent in scale but may learn from noisy proxies. The question is whether the two sources can be combined without treating either raw ratings or model scores as fully trustworthy.

Method: Aggregate-then-Calibrate (AtC) uses two stages. First, a heterogeneous rank-aggregation model estimates a consensus ordering while modeling annotator reliability. Second, isotonic projection adjusts a predictive model's scores so they respect that ordering while retaining metric information. The split is important: humans contribute ordinal structure, while the model contributes a usable score scale. The paper also proves an efficiency advantage for heterogeneous aggregation, risk bounds for calibration under a misspecified consensus ranking, and asymptotic improvement over model-only assessment. 1

Comparison with prior work: Judgment aggregation methods use human comparisons but do not supply a calibrated metric scale. Model-only assessment supplies a scale but inherits proxy-label or feature bias. AtC is therefore not another preference-optimization objective; it is a protocol for joining ordinal constraints with quantitative prediction. The experiments compare HRA variants, CrowdBT, CrowdTCV, BTL, and TCV under the same assessment setup.

Results / takeaway: On the semi-synthetic Reading Level dataset, the HRA-G row reports Kendall's τ of 0.375 / 0.399 / 0.410 for human-only, model-only, and AtC scores. The corresponding Wasserstein distances are 2.250 / 2.831 / 0.839, and the KS statistics are 0.500 / 0.300 / 0.163. The dataset contains 490 documents, 624 annotators, and 12,728 pairwise judgments. On the real-world Dots-activity task, the paper uses 300 participants judging 30 images, producing 8,700 pairwise comparisons. 1

Takeaway: Read this if your evaluation process has reliable pairwise judgments but an unstable absolute scale. The evidence supports the two-stage separation; it does not remove the need to audit the consensus ranking or obtain ground truth where the decision cost is high.

Resources: The paper releases code and datasets at CAMELLIAxt/12500_AtC_supp. 1

Area tag: LLM systems and probabilistic programming arXiv: 2608.05234 Authors / institutions: Louis Mandel, Mandana Vaziri, and Martin Hirzel, IBM; Guillaume Baudart, Université Paris Cité, Inria, CNRS, IRIF. 4 Peer-review status: Published at ICML 2026. The v1 record was submitted on August 5, 2026. 4

Problem: A multi-step LLM flow compounds uncertainty across model calls, tool calls, and repair steps. Inference scaling can improve accuracy by exploring multiple traces, but implementations often entangle majority voting, importance sampling, or sequential Monte Carlo with the flow's application logic. That makes it difficult to inspect the distribution of possible outputs or switch inference policies.

Method: PPDL extends a prompt-programming language with probabilistic factor statements. The flow remains the place where developers specify model and tool calls; factors score a trace with soft or hard constraints, and the interpreter maintains a distribution over execution states. The same flow can then run with single execution, majority voting, importance sampling, or sequential Monte Carlo. The implementation parallelizes independent calls and uses lazy data structures so it waits only for states needed by later expressions. 4

Comparison with prior work: Prompt programming makes LLM/tool composition easier but does not expose uncertainty. Hand-written inference-scaling loops expose the search procedure but duplicate flow logic and make alternative algorithms costly to try. PPDL's claim is narrower and more useful: it puts sampling and trace weighting into a language-level interface rather than asking every developer to reimplement them.

Results / takeaway: Table 2 caps the benchmarks at 500 samples and reports three-run means. For granite4-small on GSM8k, single execution reaches 83.8%, majority voting 90.4%, importance sampling 93.7%, and SMC 92.3%; on Math500 the corresponding numbers are 60.7%, 62.1%, 64.0%, and 65.5%. The best inference engine depends on the model and task rather than following one fixed ranking. In the Rocq theorem-proving case, 40 particles solve 87.7 ± 2.4 of 244 MiniF2F-Rocq problems with importance sampling and 95.0 ± 2.4 with SMC, a difference of 7.3 problems on average. 4

Takeaway: PPDL is most relevant when a workflow already has multiple uncertain steps and the engineering bottleneck is experimentation with inference policy. Its benchmark does not show that scaling always helps: factors must carry useful information, and pass@k remains an oracle upper bound rather than an executable strategy.

Resources: The paper's implementation is IBM/prompt-declaration-language, ICML 2026 branch. 4

Area tag: Transformer optimization and interpretability arXiv: 2608.06776 Authors / institutions: Rahul Vashisht, Department of CSE, IIT Madras; Harish G. Ramaswamy, Department of CSE and Department of DSAI, IIT Madras. 7 Peer-review status: Accepted to the 43rd International Conference on Machine Learning (ICML 2026). The v1 record was submitted on August 7, 2026. 7

Problem: Query-key parameters determine where attention goes, while output-value parameters determine how attended representations affect predictions. Two self-attention models can reach similar predictive loss yet produce very different attention patterns. The paper asks whether parameterization and relative learning speed explain that gap.

Method: Gradient-flow analysis separates the QK and OV circuits. Factorized parameterizations implicitly rescale the learning rates of the corresponding collapsed circuits, and the paper derives closed-form trajectories in which QK and OV parameters move along a line at different speeds. When QK learns faster than OV, the model increases attention mass on task-relevant tokens to compensate for slower output learning. This mechanism is tested by multiplying the QK learning rate while leaving the OV rate fixed. 7

Comparison with prior work: Earlier attention-interpretability work largely asks whether attention weights align with explanations. This paper changes the intervention point: it treats attention concentration as a consequence of optimization dynamics and parameterization, not only as a property to judge after training. The practical comparison is therefore equal predictive performance with different circuit learning rates.

Results / takeaway: On SQuAD, Faster QK raises test F1 only from 56.11 to 56.86, but Attention Confidence rises from 12.65 to 42.95, Attention Confidence & Model Confidence from 12.60 to 41.48, and comprehensiveness from 0.38 to 0.53. On SVA, Attention Confidence changes from 41.67 to 76.85 while test accuracy changes from 92.434 to 93.412. HateXplain shows the same separation: test accuracy rises from 55.74 to 56.90, while Attention Confidence goes from 0.00 to 43.50. The numbers support sharper attention proxies without treating them as proof that attention is a faithful explanation. 7

Takeaway: For interpretability experiments, learning-rate groups are not a harmless implementation detail. The theory is developed for a single-layer attention-only Transformer, so the multi-layer results are empirical evidence of extension, not a theorem about every modern Transformer.

Resources: Reproduction code is available at vashishtrahul/Faster-Query-Key-Learning-Sharpens-Attention-in-Self-Attention-Models. 7

Area tag: Scientific machine learning and stochastic dynamics arXiv: 2608.01582 Authors / institutions: Shida Liu, Sumit Sinha, and L. Mahadevan, Harvard University; Abhishek Gupta, Tracelink. The Harvard affiliations span SEAS, Physics, and OEB. 2 Peer-review status: Accepted to the International Conference on Machine Learning (ICML 2026). The v1 record was submitted on August 3, 2026. 2

Problem: Symmetry discovery for deterministic systems has a mature toolbox, but the continuous symmetries of stochastic differential equations are rarely known. A method that only fits trajectories can recover a predictor without recovering the transformations that leave the stochastic dynamics unchanged.

Method: LieStoNet first learns neural surrogates for drift and diffusion from trajectory increments. It then learns projectable Lie-point generators while enforcing the SDE determining equations, closure under Lie brackets, the bilinearity/antisymmetry/Jacobi axioms, and an independent basis. The surrogate also defines an associated Fokker–Planck equation, whose symmetries can be discovered in parallel. This separates learning the local dynamics from learning the algebraic structure that should organize them. 2

Comparison with prior work: Existing symmetry-discovery approaches in the paper's comparison mainly target deterministic equations or require a prescribed group, coordinate system, or generator template. LieStoNet removes those templates but still has a real boundary: it searches within projectable Lie-point SDE symmetries. That boundary is more informative than calling the method assumption-free.

Results / takeaway: On four canonical SDE examples, the selected-dimension principal angles between learned and analytic generator spans include 6.261°, 2.487°, and 2.313° for Example 1; smaller angles mean closer span alignment. A ten-dimensional Brownian SDE with a 12-dimensional symmetry algebra is recovered with principal angles ranging from 0.73° to 16.55°. The finite-flow check pushes 500 analytic trajectories along learned generators and reports small drift/diffusion residuals at ε values 1, 3, and 5. On high-frequency BTC/USDT data—about 1.8 million ticks aggregated to 500 ms and split into 575 sub-trajectories—the paper uses distributional invariance diagnostics rather than claiming a known ground-truth algebra. 2

Takeaway: This is a paper about recovering a reusable structural object, not merely fitting a stochastic predictor. The canonical-SDE results establish recovery under controlled ground truth; the BTC/USDT experiment asks whether the diagnostics remain useful when the true symmetry algebra is unavailable.

Resources: Code is available at sumit-sinha-seas/LieStoNet_Final. 2

The four vision papers intervene at different points in the pipeline. EvReflection adds missing physical evidence at the input, Suppress and Diversify changes internal route selection, FIT changes how pixels become tokens and return to pixels, and VisME changes the training target and sampling policy. The comparison is useful for implementation decisions: the same word—robustness—can mean better evidence, better routing, better token boundaries, or a better-defined capability.

Area tag: Universal multimodal embeddings arXiv: 2608.01794 Authors / institutions: Jiawei Cao, Junyi Feng, Jiashen Hua, Ziheng Huang, Bing Deng, Kaijie Wu, Chaochen Gu, and Jieping Ye; Shanghai Jiao Tong University and Alibaba Group. 3 Peer-review status: Accepted to CVPR 2026. The v1 record was submitted on August 3, 2026. 3

Problem: Universal multimodal embeddings are usually evaluated on classification, VQA, retrieval, and grounding, but those tasks do not necessarily test whether two images depict the same identity. That gap matters for instance retrieval, re-identification, identity grounding, and identity preservation during image generation or editing.

Method: The paper defines visual identity discrimination and introduces MVEB, a benchmark with four identity-centric meta-tasks—identity recognition, re-identification, visual identity grounding, and identity editing—across 28 datasets. It trains VisME with identity-aware sampling, hard negatives, and a contrastive objective that preserves intra-identity consistency while mixing general multimodal data. MVEB contains 522K samples; 20 datasets are used for training and 8 for out-of-domain evaluation. 3

Comparison with prior work: CLIP-family dual encoders and recent universal embedding models provide a strong general multimodal baseline, but their benchmark mix can hide identity failures. VisME changes both the evaluation surface and the sampling regime. The comparison is not simply a larger model: the 2B and 7B VisME variants are tested against dual encoders and VLM-based embedding models on the combined MMEB and MVEB suite.

Results / takeaway: On the combined benchmark, VisME with Qwen2.5-VL-7B reaches 78.8 average on MVEB and 72.2 on MMEB; the 2B VisME variant reaches 69.1 on MVEB. In the sampling ablation, identity-aware sampling improves MVEB in-distribution accuracy by 11.9 points and out-of-distribution accuracy by 12.8 points relative to the unconstrained sampling variant. The result to check is the capability separation: identity performance improves substantially while general multimodal retrieval remains competitive. 3

Takeaway: Read this if a multimodal embedding system will be used for identity-sensitive retrieval or editing. The benchmark is authored with the method, so the next question for deployment is whether its identity definitions and OOD splits match the identities that matter in the target application.

Resources: MVEB project page provides the paper's code and data entry point. 3

Area tag: Event-based image restoration arXiv: 2608.06184 Authors / institutions: Jiaxiao Wang, Dachun Kai, Huyue Zhu, Quanquan Hu, Zhenyang Xu, and Xiaoyan Sun; University of Science and Technology of China and the Institute of Artificial Intelligence, Hefei Comprehensive National Science Center. 5 Peer-review status: The paper's published PDF identifies it as part of the Proceedings of the 43rd International Conference on Machine Learning (PMLR 306, 2026). The arXiv v1 record was submitted on August 6, 2026. 5

Problem: A single RGB frame cannot reliably distinguish a reflection layer from the transmitted background when their textures look alike. Multi-frame methods can use parallax, but usually need enough camera movement to produce measurable optical flow.

Method: EvReflection uses an event camera to record brightness changes at microsecond resolution during slight motion. A Micro-Dynamics Decoupler estimates layer-specific motion cues from the event stream, and a Parallax-Attention Rectifier uses those cues to remove residual artifacts from the RGB image. To address data scarcity, the paper builds a parallax-aware simulation pipeline and the real-world EVR2 benchmark with synchronized event-RGB capture.

Comparison with prior work: Single-image methods rely on static priors and remain vulnerable to texture confusion; multi-image methods use motion but can require a large baseline and suffer registration artifacts. EvReflection's claim is a different sensing regime: small physical motion plus event timing supplies differential evidence without the large camera movement assumed by many multi-frame methods.

Results / takeaway: On SIR2, EvReflection reaches 29.09 dB average PSNR, compared with 26.71 dB for RDNet; its average SSIM is 0.940 and average LPIPS is 0.036. On the real EVR2 benchmark, it reaches 27.25 dB average PSNR, with 28.88 / 27.04 / 25.81 dB on the 3 mm, 5 mm, and 8 mm glass subsets. EVR2 contains 140 scenes recorded at three glass thicknesses, yielding 420 paired sequences. The gain over RDNet grows on the thick-glass subset, where ghosting is harder to remove. 5

Takeaway: The method's improvement comes from acquiring a cue that RGB-only restoration lacks. The practical prerequisite is equally clear: an event-RGB camera and a motion setup, not merely a new image-restoration checkpoint.

Resources: The paper releases code, EVR2 data, and pretrained models. 5

Area tag: Corruption robustness arXiv: 2608.06712 Authors / institutions: Jiangang Yang, Wenhui Shi, Xiaoran Xu, Wenyue Chong, Luqing Luo, Jing Xing, and Jian Liu; Institute of Microelectronics, Chinese Academy of Sciences, and University of Chinese Academy of Sciences, Beijing. 6 Peer-review status: Accepted at ICML 2026. The v1 record was submitted on August 7, 2026. 6

Problem: Natural corruptions expose a progressive decay of robust features through network layers, but most defenses improve robustness indirectly through augmentation, regularization, or ensembles. The paper asks whether internal computational pathways can be identified and changed directly.

Method: Suppress and Diversify (S&D) dynamically selects pathways associated with robust features and diversifies them through symmetry-preserving transformations. The selected group is merged into the model during training. S&D is architecture-agnostic, parameter-free, and adds no test-time computation; the paper also studies representation stability with CKA and tests classification, detection, segmentation, and test-time adaptation.

Comparison with prior work: Stochastic depth and sub-network selection alter routes, while dynamic weight evolution, consistency learning, and bio-inspired models pursue robustness through other inductive biases. S&D's difference is to use an explicit robust-pathway signal and symmetry-preserving diversification rather than relying only on input perturbations or a fixed ensemble.

Results / takeaway: In the ImageNet-C robustness-paradigm comparison, S&D scores 42.7% top-1 accuracy versus 39.2% for the baseline, and a stronger S&D variant reaches 46.3%. On corrupted semantic segmentation, Mask2Former improves from 40.8 to 47.1 mIoU on Cityscapes-C, a 6.3-point gain. On COCO-C, YOLOv5s improves from 19.9 to 22.1 mAP. On ACDC adverse-weather segmentation, GCNet improves by 3.8 mIoU points. These gains occur across tasks, but clean accuracy and training-recipe interactions still need to be checked per backbone. 6

Takeaway: S&D is a plug-in candidate when test-time latency cannot increase and the model can afford a training-time pathway intervention. The zero test-time overhead claim does not mean zero training cost or uniformly higher clean accuracy.

Resources: JGyoung-UCAS/suppress_and_diversify. 6

Area tag: All-in-one image restoration arXiv: 2608.06832 Authors / institutions: Zihao He, Yunfeng Wu, and Songhua Liu, School of Artificial Intelligence, Shanghai Jiao Tong University; Yunfeng Wu, Alibaba Group and School of Advanced Technology, Xi’an Jiaotong-Liverpool University; Xinchao Wang, Department of Electrical and Computer Engineering, National University of Singapore. 8 Peer-review status: Accepted to ICML 2026. The v1 record was submitted on August 7, 2026. 8

Problem: Unified restoration Transformers usually cut an image into fixed patches and inject degradation or task information only inside later backbone blocks. That leaves patch embedding and pixel reconstruction blind to local, spatially mixed corruption such as rain in one region and haze or reflection in another.

Method: Flexible Image Transformer (FIT) uses a dual-granularity Degradation Encoder: a global degradation vector and a spatial degradation map. Together they condition deformable patch embedding and inverse-warping unembedding, so patch sampling and reconstruction can bend toward local degradation structure. Task-token dropout regularizes the conditioning signal so the model does not rely on a perfect task label. The learned offsets also expose where the model adapts spatially. 8

Comparison with prior work: Recent all-in-one methods condition Transformer blocks after tokenization, or use global degradation representations and task labels. FIT moves the conditioning boundary outward to the operations that create and reconstruct tokens. In the ablation, deformable tokenization adds 0.54 dB over the fixed-token baseline, degradation conditioning adds another 0.57 dB, and seam-aware training reduces grid artifacts.

Results / takeaway: Across BSD68, Rain100L, SOTS, GoPro, and LOLv1, FIT reports 32.83 dB average PSNR on the three-degradation setting and 30.72 dB on the five-degradation setting, outperforming recent unified restoration methods by 0.5–1.1 dB. The five-degradation table reports 30.72 dB average with 0.919 average SSIM. The result to inspect is not only the average: the learned offset maps provide a visual check of whether the token grid follows the corruption rather than merely producing a higher aggregate score. 8

Takeaway: FIT is a direct answer to spatially non-uniform degradation. Its evidence is benchmark-based, so deployment on a target camera or mixed-corruption distribution still needs failure-case validation.

Resources: No public code URL is listed in the verified paper record. 8

Open Aggregate-then-Calibrate first if your evaluation data mixes pairwise human judgments with model scores. Choose PPDL if your LLM workflow needs to compare inference-scaling policies without duplicating flow logic. LieStoNet is the deepest structural-method paper here: its value depends on whether symmetry recovery, rather than prediction alone, is the object you need.

For vision systems, start with EvReflection when the bottleneck is missing physical evidence, Suppress and Diversify when inference latency rules out test-time adaptation, FIT when spatially mixed degradation breaks fixed patching, and VisME when general multimodal retrieval is not enough to guarantee identity discrimination. Faster Query-Key Learning belongs near the top for readers studying why attention patterns change even when task scores barely do.