Top-conf paper digest - week of July 20-24, 2026
Eight newly posted arXiv papers with explicit NeurIPS, ICML, ICLR, or CVPR 2026 status, grouped across multimodal reasoning, agent evaluation, discrete diffusion, visual RL, 3D rendering, and diffusion inference.
Selection basis
- Window: first arXiv submissions dated July 20–24, 2026. The eight selected records are v1 submissions dated July 20 or July 23; same-day Monday records from July 27 are outside this issue.
- Status rule: acceptance, spotlight, or conference-publication language is copied from each arXiv abstract record. A record that only names a conference without explicit acceptance wording is labeled as a preprint.
- Coverage rule: the set spans multimodal reasoning, agent evaluation, discrete diffusion, visual reinforcement learning, 3D rendering, and diffusion-model inference. Each entry includes a method mechanism, a comparison point, and the strongest quantitative evidence available in the paper.
- Affiliations: the arXiv HTML record exposes author-to-institution mappings for some papers but not all. Missing mappings are stated rather than inferred.
| Area | Paper | Status | Reading signal |
|---|---|---|---|
| Multimodal reasoning | ST-Veto | Preprint; arXiv comment says “ICML 2026 - main” | Training-free decoding uses temporal confidence and image grounding together. |
| Agent evaluation | Copyright-Bench | ICML 2026 Spotlight | Measures copyright violations separately from task completion. |
| Discrete diffusion | Context-weighted Discrete Flow Matching | NeurIPS 2026 preprint; under review | Reweights token updates by local context rather than treating all positions alike. |
| Visual RL | PAMD | Accepted to ICML 2026 | A structured pairwise metric replaces both fixed norms and unconstrained distance heads. |
| RL theory and practice | Relative Value Learning | Published as an ICLR 2026 conference paper | Learns value differences directly and reconstructs an advantage estimator. |
| 3D systems | CaT-GS | Preprint; arXiv comment says “CVPR 2026” | Caches inter-frame work and splits overloaded rasterization tiles. |
| Diffusion inference | DiFA | Accepted to ICML 2026 | A solver-compatible temporal consensus improves few-step sampling without extra NFEs. |
| Diffusion inversion | SlerpFlow | Accepted to ICML 2026 | Spherical trajectory correction reduces inversion error at the same evaluation budget. |
Multimodal reasoning and agent evaluation
ST-Veto: vetoing unstable and weakly grounded tokens in diffusion MLLMs
Area tag: Multimodal reasoning / diffusion MLLMs
arXiv: 2607.17884
Authors / institutions: Keuntae Kim, Beomseok Lee, Hyunwoo Kim, and Yong Suk Choi. The paper record does not expose clear author-to-institution mappings.
Peer-review status: Preprint; the arXiv comment says “ICML 2026 - main.” 1
Problem: Diffusion multimodal language models can revise many token positions in parallel, but a token that looks confident at the current denoising step may become an unreliable anchor later. Autoregressive chain-of-thought methods also assume a sequential context that diffusion decoding does not provide.
Method: ST-Veto is a training-free decoding policy. It tracks each token’s confidence across diffusion steps, uses a second-order Taylor prediction to identify unstable trajectories, and measures image-attention mass as a visual-grounding signal. Tokens that fail either test enter a veto set. The policy swaps them only with near-boundary candidates that meet the safety criteria, so it can perform a partial swap instead of forcing a dubious replacement. 1
Comparison with prior work: The experiments compare confidence, entropy, and margin decoding with DDCoT, CCoT, SCAFFOLD, ICoT, EAH, OPERA, and FarSight. The change is to use both a temporal stability test and a visual-grounding test before unmasking, rather than adapting a chain-of-thought policy designed for autoregressive VLMs.
Results / takeaway: On LaViDa, ST-Veto reports 30.21 / 31.87 on MMBench, 38.05 / 38.25 on SQA-IMG, and 47.80 / 48.54 on V* Bench, compared with the base-confidence values of 27.96 / 29.55, 33.96 / 31.55, and 45.81 / 40.06. On MMaDA, the corresponding ST-Veto figures are 27.52 / 27.74, 36.55 / 36.74, and 47.25 / 45.31. The paper reports accuracy gains of up to 9% and negligible overhead relative to base diffusion decoding. 1
Resources and limitations: No code or project URL is listed. The method depends on confidence and attention as imperfect proxies, and Taylor prediction is less reliable when confidence trajectories are non-smooth. The experiments focus on single-image reasoning; multi-image, video, and very long generations remain open cases.
Copyright-Bench: separating task completion from copyright compliance
Area tag: Agents / AI governance
arXiv: 2607.21799
Authors / institutions: Zheng Hui, Doni Bloomfield, and Noam Kolt. The paper identifies the Hebrew University Governance of AI Lab and lists supporting organizations, but does not provide a complete author-to-institution mapping.
Peer-review status: ICML 2026 Spotlight. 2
Problem: Agent benchmarks usually score whether an agent finishes a task, not whether it uses lawful inputs while doing so. Copyright-Bench tests that gap in three workflows: commercial web development, merchandise and derivative generation, and pitch-deck creation.
Method: The benchmark presents a partially observable asset-selection task under U.S. copyright law. Each task has at least one valid compliant solution. Its 200 high-resolution assets balance public-domain images from the National Park Service against restricted images from DepositPhotos. CLIP-ViT-L/14 embeddings and an LAION-Aesthetics V2 predictor pair visually similar assets, while JSON sidecars expose the decisive copyright notice through a
read_metadata tool. Agents must inspect files, view images, read metadata, and submit a selection. The main metric is Violation Rate, the share of tasks in which at least one restricted asset is selected; Task Success Rate tracks whether the underlying task was completed.Comparison with prior work: The paper adds legal compliance to the evaluation target instead of treating asset choice as a hidden implementation detail. It also varies the prompt with neutral, IP-aware, hurried, and IP-dismissive conditions, then compares proprietary and open-weight models with a 176-person human baseline. 2
Results / takeaway: In the standard environment, mean violation rates are 27.6% for Claude 4.5 Opus, 28.7% for Claude 4.5 Sonnet, 30.6% for GPT-5.2, and 36.2% for GPT-4o. Open-weight models are higher in the reported table: 41.0% for Llama-4-Maverick, 43.9% for Qwen-3VL, and 48.6% for Llama-4-Scout. An IP-aware prompt sharply lowers violations; Gemini 3 Pro’s WebDev rate falls from 39.6% to 10.0% in the cited comparison. The useful result is diagnostic: many agents can complete the commercial task while still selecting a restricted asset. 2
Resources and limitations: Code and data repository. The benchmark covers U.S. copyright law, a binary public-domain versus restricted distinction, and third-party asset retrieval. It does not test fair-use scenarios, model memorization, trademarks, privacy, or generated text, audio, and video.
Loading content card…
Discrete diffusion and generative modeling
Context-weighted Discrete Flow Matching: give predictable tokens more training weight
Area tag: Discrete diffusion / language modeling
arXiv: 2607.21427
Authors / institutions: Daniil Cherniavskii, University of Amsterdam; Daniel Severo and Karen Ullrich, Meta FAIR. The paper notes that the first author’s work was done during an internship at Meta FAIR.
Peer-review status: Preprint; the arXiv comment says “Under review at NeurIPS 2026.” 3
Problem: Standard discrete flow-matching objectives mix easy, predictable token targets with ambiguous, high-entropy targets. The paper argues that local context provides a useful conditioning signal: positions surrounded by more revealed context are often easier to predict, but the standard objective does not weight them differently.
Method: Context-weighted Discrete Flow Matching modifies the underlying continuous-time Markov chain so token-update probabilities depend on local context while preserving the boundary conditions and marginal unmasking probabilities. The inference variants are neighbor-weighted and entropy-weighted sampling. The training variant, Scaled Cross-Entropy, reweights coordinate losses toward better-conditioned positions. The authors also give an exact train-time path sampler that samples the number of unmasked tokens and then samples positions one at a time.
Comparison with prior work: The experiments compare DFM with autoregressive modeling, BD3LM, SEDD, and MDLM, and compare CE, NELBO, Bregman, and SCE objectives. The contribution is a context-dependent weighting rule that keeps any-order generation rather than turning the model into a conventional autoregressive decoder. 3
Results / takeaway: On OpenWebText, neighbor-weighted sampling improves MAUVE by up to 24% and reduces generative perplexity by up to 63% relative to Euler in the reported experiments. With Scaled Cross-Entropy, the paper reports MAUVE 0.777 and generative perplexity 30.20, compared with 0.628 and 82.17 for standard uniform DFM. On QM9, neighbor-weighted sampling nearly triples valid molecules and doubles novel molecules over Euler. The tradeoff is visible in standard test perplexity: on Wikitext103, DFM uniform is 63.38 while DFM plus SCE is 91.53, even though the generation metrics improve. 3
Resources and limitations: No code repository is listed. The method adds little computation and keeps the same basic sampling complexity, but neighbor-weighted sampling is weaker at very low NFE. The experiments center on small models and one-dimensional discrete sequences, so behavior on larger models, longer contexts, and images is not established.
DiFA: align diffusion predictions before the solver step
Area tag: Diffusion-model inference
arXiv: 2607.17972
Authors / institutions: Shigui Li and Delu Zeng. The paper record does not expose author affiliations.
Peer-review status: Accepted to ICML 2026. 4
Problem: Few-step diffusion sampling treats each clean-signal prediction as an exact estimator. In high-curvature regions, local linearization can create bias, and errors can accumulate into prediction drift. Distillation addresses speed through retraining, while higher-order solvers improve numerical integration without using temporal redundancy among successive predictions.
Method: DiFA is a training-free wrapper around an existing solver. It treats several clean-signal predictions along the reverse trajectory as correlated observations, then forms a forward-aligned temporal consensus using structural consistency and noise-level compatibility. A deviation-guidance term preserves useful residual information so the consensus does not simply over-smooth the prediction. DiFA refines the clean prediction supplied to each solver step; it does not replace the solver and adds no network function evaluations.
Comparison with prior work: The paper tests DDPM, DDIM, DPM-Solver and its variants, UniPC, Heun, EDM, EVODiff, and other diffusion samplers on CIFAR-10, ImageNet-64, LSUN Bedroom, and SiT-XL/2 flow-matching settings. The comparison is therefore against both stronger numerical solvers and methods that trade inference cost for distillation.
Results / takeaway: On CIFAR-10, DPM-Solver++ with DiFA reaches FID 1.96 at 20 NFE, compared with 2.33 for the solver alone. On ImageNet-64, DPM-Solver++ with DiFA reports FID 1.64 at 25 NFE versus 1.83 for the baseline, while UniPC with DiFA reports 1.63 versus 1.73. In the SiT-XL/2 flow-matching experiment at 10 NFE, DiFA improves IS from 80.2694 to 91.7977 and lowers FID from 27.8142 to 21.5990. 4
Resources and limitations: The paper says code is available but does not expose a concrete URL. DiFA assumes a clean-prediction parameterization, fixed hyperparameters, and a static anchor approximation. Those assumptions may need revision for text-to-image, video, inverse-problem, and distillation pipelines.
SlerpFlow: correct rectified-flow inversion on a spherical path
Area tag: Diffusion inversion / image editing
arXiv: 2607.21326
Authors / institutions: Wenbin Duan, Yan Shu, Zhuoyuan Fu, Fangmin Zhao, Yan Li, Yaru Zhao, and Binyang Li. The paper record does not expose author affiliations.
Peer-review status: Accepted at ICML 2026. 5
Problem: Rectified-flow diffusion transformers such as FLUX need inversion to map an image back to latent noise before faithful reconstruction or editing. Linear solvers introduce discretization error, which can distort both the reconstruction and the edited image.
Method: SlerpFlow separates radial and angular latent dynamics and uses spherical linear interpolation to predict the next direction on the hypersphere. It then estimates the target radial shell and synthesizes a chordal velocity that lands on that shell. The method follows a predictor-corrector structure similar to Heun, while caching the current velocity so only the future velocity needs a new model evaluation.
Comparison with prior work: The reconstruction comparison includes Euler, Heun, RF-Solver, and FireFlow. Editing comparisons include P2P, PnP, PnP-Inv., MasaCtrl, InfEdit, RF-Inv., RF-Solver, and FireFlow. SlerpFlow keeps the same step and NFE budget as FireFlow, so the result isolates the trajectory correction rather than an added evaluation budget.
Results / takeaway: On PIE-Bench unconditional reconstruction, SlerpFlow reports PSNR 20.73, SSIM 75.91, and LPIPS 30.56, compared with FireFlow at 20.19, 74.47, and 31.09. In the conditional setting, it reports PSNR 21.19, SSIM 79.35, and LPIPS 26.01 versus FireFlow at 20.39, 63.85, and 27.96. For text-driven editing, the method uses 15 steps and 32 NFE, and the paper reports the highest CLIP-Whole and CLIP-Edited scores among the compared methods. 5
Resources and limitations: The paper says code is available but does not show a concrete URL in the record. The spherical latent-space assumption can over-correct when local geometry is anisotropic or flat, and a global angular correction can perturb unedited background regions under a strong prompt shift.
Visual reinforcement learning
PAMD: a structured adaptive distance for bisimulation representations
Area tag: Visual reinforcement learning
arXiv: 2607.18004
Authors / institutions: Daegyeong Roh, Juho Bae, and Han-Lim Choi. The paper record does not expose author affiliations.
Peer-review status: Accepted to the 43rd International Conference on Machine Learning (ICML 2026). 6
Problem: Bisimulation representation learning needs a latent distance that tracks behavioral distance. A fixed norm or cosine comparator is too rigid, while an unconstrained pairwise MLP can absorb the supervision in its distance head without teaching the encoder a useful structure.
Method: PAMD, or Pairwise Adaptive Mahalanobis Distance, predicts a pair-conditioned positive-definite quadratic form. MetricNet emits lower-triangular parameters, a Cholesky construction makes the matrix positive definite, symmetry handles input order, and trace normalization removes scale freedom. The module is a plug-in comparator for bisimulation-based visual RL methods.
Comparison with prior work: The experiments compare DBC, DBC-Det, MICo, and SimSR, with reference comparisons to TACO, DrQ-v2, A-LIX, DrQ, and CURL. Structural ablations compare global diagonal and adaptive diagonal Mahalanobis comparators with an unconstrained pairwise MLP. The central comparison is between flexibility and a constraint strong enough to prevent the distance head from taking over.
Results / takeaway: Across the evaluated DeepMind Control Suite tasks, DBC-Det plus PAMD yields substantially higher returns than DBC-Det, and PAMD also improves MICo- and SimSR-style operators. Removing trace normalization consistently degrades performance. In a frozen-encoder residual diagnostic, the unconstrained MLP quickly fits the residual, while the structured positive-definite distance can reduce it only modestly; once the encoder is trainable, PAMD continues reducing the residual. This supports the paper’s claim that the constraint shifts learning pressure back toward the representation, although the paper’s text extraction does not expose a complete numeric return table. 6
Resources and limitations: Code repository. The dense pair-conditioned quadratic form scales poorly with latent dimension. Experiments are centered on pixel-based DMControl and a fixed set of baseline pipelines, while comparisons with stronger recent methods such as TACO, MLR, and DrM are not fully controlled.
Loading content card…
Relative Value Learning: estimate differences instead of absolute critics
Area tag: Reinforcement learning / value estimation
arXiv: 2607.21120
Authors / institutions: Marc Höftmann, Jan Robine, and Stefan Harmeling. The paper record does not expose author affiliations.
Peer-review status: Published as a conference paper at ICLR 2026. 7
Problem: Standard critics estimate an absolute state value, V(s), even though control uses differences between situations. The paper asks whether learning those differences directly can produce a sound critic and an unbiased policy-gradient estimator.
Method: Relative Value Learning defines an antisymmetric pairwise function, Δ(sᵢ, sⱼ) = V(sᵢ) − V(sⱼ), and introduces a pairwise Bellman operator. The authors prove that the operator is a γ-contraction with a unique fixed point equal to the true value differences. They derive 1-step, n-step, and λ-return targets, then reconstruct generalized advantage estimation as R-GAE from pairwise differences.
Comparison with prior work: The practical comparison integrates relative value estimation with PPO and evaluates it against standard PPO on the 49-game Atari benchmark. The theoretical comparison is against the usual absolute-critic formulation: the proposed estimator changes the learned object while retaining a policy-gradient route through advantage differences.
Results / takeaway: On the Atari benchmark covering 49 ALE games, the paper reports competitive performance for PPO with Relative Value Learning compared with standard PPO. The abstract does not provide an aggregate score or per-game table, so the evidence supports parity across the benchmark rather than a numeric claim of improvement. The paper is most relevant to readers interested in the critic’s target and estimator bias, not only in another Atari leaderboard result. 7
Resources and limitations: No project or code URL is listed on the arXiv record. The central evidence is theoretical plus the Atari integration; the abstract does not establish how the method behaves in continuous-control or large-scale actor-critic settings.
3D rendering
CaT-GS: cache the work that 3D Gaussian rendering repeats
Area tag: 3D vision systems / Gaussian splatting
arXiv: 2607.17842
Authors / institutions: Tingjia Zhang, Bo Chen, Shengzhong Liu, Fan Wu, and Guihai Chen. The paper lists Shanghai Jiao Tong University for Zhang, Liu, Wu, and Chen, and the University of Illinois Urbana-Champaign for Bo Chen.
Peer-review status: Preprint; the arXiv comment says “CVPR 2026.” 8
Problem: 3D Gaussian Splatting slows down on large scenes because tile-based rasterization creates more work and uneven GPU utilization. Existing accelerators either require retraining or optimize only rasterization while leaving repeated inter-frame processing in the pipeline.
Method: CaT-GS combines speculative multi-frame preprocessing, inter-frame caching, and load-aware task splitting. Motion prediction and fine-grained intersection checks determine when a Gaussian trail can be reused. Frustum caching skips repeated culling, sort caching reuses key-frame ordering, and a CUDA kernel splits overloaded tiles into smaller tasks. When motion becomes too large, the method shrinks or skips speculation rather than forcing a stale cache.
Comparison with prior work: The baselines are 3DGS, ADR-GS, and Flash-GS. CaT-GS differs by treating frame-to-frame preprocessing and tile imbalance as first-class sources of redundancy, not only by optimizing the final rasterization kernel.
Results / takeaway: On the reported trajectories, CaT-GS reaches 736.5 FPS on Truck versus 478.3 for Flash-GS, 892.2 versus 528.3 on Train, and 922.3 versus 634.2 on Playroom. On five UAV scenes, it reports 202.5–378.5 FPS, while the paper’s summary claims up to 10× speedup over original 3DGS and up to 70% over prior state of the art. Quality remains close to FlashGS in the reported examples: on Truck, Ours-Sub is 28.16 / 0.852 versus FlashGS at 28.19 / 0.856. 8
Resources and limitations: No code repository is listed, and the UAV city dataset is described as future open-source data. Large inter-frame motion can cause Gaussian mispositioning or force speculation off. Task splitting also adds work when early termination makes later subtasks unnecessary.
Reading order
Open DiFA first if your diffusion pipeline is limited by few-step quality and cannot afford retraining. Read SlerpFlow for FLUX-style inversion and editing under a fixed evaluation budget. For multimodal reasoning, ST-Veto is the direct methods paper; Copyright-Bench is the direct evaluation paper for agents that choose external assets. CaT-GS is the systems read for large-scene rendering, while PAMD and Relative Value Learning target different layers of visual RL, a learned distance versus the critic’s value representation. Context-weighted Discrete Flow Matching is the one to open if your discrete diffusion model shows a gap between generation quality and standard perplexity.
References
- 1
- 2
- 3
- 4
- 5
- 6
- 7Relative Value Learning
arxiv.org
- 8
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.