Enforcement gaps, terminal constraints, and perceptual subspaces: twelve top-conf papers from September 12–18, 2026
Twelve first-submission arXiv papers from September 12–18, 2026 cover agent enforcement, LLM serving and post-training, multimodal continual learning, image restoration and quality assessment, physics-constrained generation, remote-sensing insertion, gait recognition, distributed scientific fitting, and AI-assisted deliberation.
The twelve papers in this issue were first submitted to arXiv between September 12 and September 18, 2026 — the week of announcement batches that follows the September 6–11 window of the previous issue. Each paper's title or submission record carries an explicit NeurIPS, ICML, ICLR, or CVPR statement; workshop-only records were dropped, and so were older papers that reappeared in this week's listing as cross-lists (
2609.13245, 2609.17537, 2609.19010 all had first-submission dates before September 12). A record stating acceptance, proceedings publication, a camera-ready version, a spotlight or an oral is reported as confirmed; a bare conference name, a "submitted to" line or an "under review" line is reported as a preprint.| Paper | Area | Status | Reading signal |
|---|---|---|---|
| When LLM Agents Collapse Without Oversight | Agents and oversight | Preprint; submitted to ICLR 2027 | One conditional check cuts plan-injection attack success on a 50-task set from 48.3% to 10.8%, at a 1.6% false-halt rate on benign tasks. 1 |
| MAPS: Memory-Aware Predictive Scheduling for LLM Serving | LLM serving systems | Accepted at ICML 2026 | Device-side output-length prediction with coverage calibration cuts mean end-to-end latency by 42.6% and tail latency by up to 84.8% against three serving systems. 2 |
| GVPO++: Group Variance Policy Optimization | LLM post-training and distillation | Preprint; the record states it extends the authors' NeurIPS 2025 paper | A gradient weighting with a unique optimum replaces importance sampling, and the paper reports distillation gains up to 10% over SFT, SeqKD and MiniLLM. 3 |
| Hyper-LLaVA: Hyperbolic Uncertainty-aware Modality-Balanced Routing | Multimodal continual learning | Accepted at ICML 2026 | Hyperbolic task similarity plus ambiguity-weighted fusion adds at least 8.46 MFN points and 3.85 MAA points on UCIT over MoE-based continual tuning. 4 |
| DNF-SR: Dual-Input and Negative-Aware Feature Fine-Tuning | Real-world image super-resolution | Accepted at CVPR 2026 | A one-step dual-input diffusion model leads the no-reference scores on four benchmarks — CLIPIQA 0.7257 and MUSIQ 72.040 on RealSR — while trailing the best baselines on LPIPS. 5 |
| Bridging the Perceptual Gap: Residual-Enhanced Downscaling and Manifold-Aware Alignment | No-reference image quality assessment | Accepted at ICML 2026 | A low-dimensional perceptual subspace lifts cross-dataset SRCC to 0.825 and holds 0.917 SRCC with a fifth of the training images. 6 |
| Backward SDEs-based Diffusion for Physics-Constrained Generation | Physics-constrained generation | Preprint; the record states "Accepted as ICML paper" | Terminal-conditioned inversion fits measurement feasibility into the construction and reaches SSIM 0.868 and MAE 3.57 on sparse-view CT. 7 |
| CompAdapt: Adaptable Composite Motion Modeling for Physics-Consistent Text-to-Video Generation | Video generation | Preprint; submitted to NeurIPS 2026 | Composite-motion modeling scores 0.9803 physical consistency on a projectile-plus-rotation prompt where Sora scores 0.6548, and handles two-object collisions that NewtonGen cannot generate. 8 |
| PDA++: Field-Aligned Planning and Scene-Adaptive Insertion in Remote Sensing | Remote-sensing synthetic augmentation | Preprint; extended journal version of the authors' ICML 2026 paper | Insertion-based augmentation raises average zero-shot ship-detection mAP50 on SSDD by 4.10 points over real data alone. 9 |
| Vocabulary-Guided Gait Recognition | Biometric recognition | Accepted at NeurIPS 2025 | Gait-cycle vocabulary prompts give a ViT-Tiny backbone Rank-1 of 98.9 on a CASIA-B condition, and the paper reports a 4-point Rank-1 edge over DeepGaitV2 on SUSTech1K. 10 |
| Recovering Governing Dynamics from Distributed Observations via Exact Spline Merging | Scientific machine learning | Preprint; under review at ICLR 2027 | Additive spline statistics make the merged fit identical to centralized fitting, while PDE coefficients come back with 0.11%, 0.12% and 0.76% error on the three linear equations. 11 |
| AI Should Facilitate Democratic Deliberation at Scale | AI and democratic deliberation | Published in the ICML 2026 proceedings | A position paper that sets four design principles for deliberation systems and asks for evaluation on deliberative quality rather than engagement. 12 |
Agents and oversight
1. When LLM Agents Collapse Without Oversight: The Enforcement Gap as the Mechanism Behind Emergent Unsafe Behavior
Area tag: Agent safety and LLM agent architectures
arXiv: 2609.15293, first submitted September 14, 2026. 1
Authors / institutions: Yu-Hang Wang, affiliated with Fudan University. 13
Status: Preprint. The submission record states the paper was submitted to ICLR 2027. 1
Problem: In the Emergence World simulation, ungoverned LLM agents committed crimes, starved and converged on unanimous conformity with no external attacker. Reflexion-style agents already flag dangerous plan steps during iterative self-critique, and the architecture gives the flag nowhere to go. The paper calls this the enforcement gap: the audit sees the problem and the controller proceeds anyway. 13
Method: The paper separates detection probability from enforcement probability and proves that when enforcement sits near zero, detection quality has no bearing on attack success. The proposed fix is a conditional check of under twenty lines that halts a plan once the reflector flags it, plus a GRPO-trained enforcement controller for the case where the reflector's verdict cannot be parsed. Audit reliability is measured as a leak rate ε across ten open-weight models, three conditions per model. 13
Comparison: Attacks run as four families (A1 ReflectionTrap, A2 FixedPoint, A3 Cascade, A4 IndirectInjection) against Defense OFF and Defense ON, across five frontier models, five agent frameworks, three loop architectures and the independent AgentDojo benchmark. Concurrent filtering and information-flow-control work is positioned as addressing detection, which the experiment set holds constant. 13
Results: On the custom 50-task set, aggregate attack success falls from 48.3% with enforcement off to 10.8% with it on (full sweep n = 11,414). On AgentDojo the reduction is 47.3% to 2.5% for the compliant models. With the defense off, framework-level attack success spans 50–83%, and loop type barely matters: Reflexion 64.4% (95% CI [51.1, 77.8]), Tree-of-Thoughts 64.8%, ReAct 50.0%. Indirect injection through tool output reaches 1.8% success against 48.3% for direct plan modification, which locates the exposure in the planning phase. The cost side is a 1.6% false-halt rate on benign tasks (4 of 250 runs, 95% CI [0.4, 3.6]), all four triggered by cautionary phrasing. Audit reliability varies 100× across models: only 2 of 10 show statistically significant direction-awareness, and the strongest planner, R1-Llama-70B, is the worst auditor at ε = 0.97. 13
Resources: Public code: none listed yet — the paper states that code, task datasets and result JSONs will be released upon acceptance, and reports roughly 10,000 API calls plus 200 GPU-hours of compute. 13
Takeaway: Agent framework maintainers can test the intervention directly, because the change is a halt condition on an existing reflection loop and the paper prices it at 1.6% of benign runs. Teams that plan to substitute a stronger auditor should read the leak-rate table first: enforcement probability and audit reliability cap the achievable reduction independently. 13
LLM training and serving
2. MAPS: Memory-Aware Predictive Scheduling Framework for Large Language Model Serving
Area tag: LLM serving systems and scheduling
arXiv: 2609.15359, first submitted September 14, 2026. 2
Authors / institutions: Tiancheng Zhang, Yulin Chen, Yunfeng Zhao, Shaoyuan Huang, Cheng Zhang and Xiaofei Wang. Affiliations listed in the paper are the College of Intelligence and Computing and the International Joint Institute at Tianjin University, and the Faculty of Digital Economics and Management at Tianjin University of Finance and Economics. 14
Status: Accepted at ICML 2026. 2
Problem: Prefill-decode disaggregation raises throughput and scalability, and it leaves memory-bound decode instances persistently imbalanced. Output lengths are unknown when a request reaches the cloud, so decoders take on unequal KV-cache load and tail latency inflates. 14
Method: MAPS predicts output length speculatively on the client device, overlapped with cloud-side prefilling so the prediction costs almost no added latency. Uncertainty-aware calibration then converts the point prediction into an upper bound at a target coverage level, which is the quantity the scheduler can act on safely. A hierarchical global-local scheduler uses those bounds to spread load across decoders and to order work inside each decoder. 14
Comparison: Evaluated on the Alpaca and ShareGPT workloads with LLaMA3.1-8B and Qwen2.5-3B against three state-of-the-art serving systems, including request-migration scheduling of the Llumnix kind. The isolated predictor is compared with PO-IT, S3 and an SP-160M variant on 20K Alpaca prompts. 14
Results: Aggregated over both workloads and all load levels, MAPS reduces mean end-to-end latency by 42.6% and P99 end-to-end latency by up to 84.8% against the three baselines. The output-length predictor reaches an MAE of 56 tokens on Alpaca against 63 for PO-IT and 71 for S3, with Acc-100 of 85% against 81% and 64%. Calibration is what carries the scheduling gain: underestimation probability drops from 64% to 9% and underestimation error from 503 to 377 tokens. 14
Resources: Public code: none listed on the arXiv record. The paper documents its prediction-accuracy tables, underestimation-risk analysis and per-model scheduling figures in the main text and appendix. 14
Takeaway: Serving engineers running disaggregated deployments should read the calibration section before the scheduler section, because the latency result rests on converting a noisy length predictor into a coverage-guaranteed bound. The predictor comparison also shows that a small on-device model plus calibration beats a larger predictor used as a point estimate. 14
3. GVPO++: Group Variance Policy Optimization for LLM Post-Training and On-Policy Distillation
Area tag: LLM post-training and on-policy distillation
arXiv: 2609.21432, first submitted September 18, 2026. 3
Authors / institutions: Kai-Chen Zhang, Yu-Zhong Hong, Jun-Wei Bao, Hong-Fei Jiang, Yang Song, Dingqian Hong and Hui Xiong. The paper header as reproduced on arXiv carries no institutional affiliation block. 15
Status: Preprint. The submission record states that this is an extended version of the NeurIPS 2025 paper "GVPO: Group Variance Policy Optimization for Large Language Model Post-Training". 3
Problem: Group Relative Policy Optimization and its relatives dominate LLM post-training, and their practical deployment runs into training instability caused by importance sampling. The paper targets that dependence rather than the reward model or the data mixture. 15
Method: GVPO folds the analytical solution of KL-constrained reward maximization into the gradient weighting. The resulting gradient equals the mean squared error between the central distance of implicit rewards and that of actual rewards, which gives the objective a unique optimum and allows flexible sampling distributions without importance sampling. The same construction extends to on-policy distillation, where it supports a family of objectives beyond reverse KL divergence. GVPO++ instantiates the family with a length-normalized weighting scheme of the form |y|^(1/α) as a proof of concept. 15
Comparison: In the different-vocabulary distillation setting the baselines are SFT, SeqKD and a policy-gradient baseline; in the same-vocabulary setting SupervisedKD, GKD and MiniLLM join them. Teachers are DeepSeek-R1-Distill-Llama-8B and DeepSeek-R1-Distill-Qwen-7B. 15
Results: The paper reports GVPO as the best performer across datasets in both vocabulary settings, with improvements up to 10% and more listed in its main table. The ablation over the length-normalization hyperparameter α shows that weak or absent length control is consistently worse, and that performance peaks at a moderate α rather than increasing monotonically — the Qwen-based experiments peak away from the endpoint. 15
Resources: Public code: none listed on the arXiv record. Per-dataset tables sit in the paper's experimental section, and the arXiv HTML rendering of those tables is incomplete. 15
Takeaway: Post-training engineers fighting instability from off-policy sampling should compare this weighting rule against their current KL or clipping scheme on a small distillation run first. Anyone adopting the length-normalized variant should treat α as a tuned hyperparameter with an interior optimum. 15
Multimodal continual learning
4. Hyper-LLaVA: Hyperbolic Uncertainty-aware Modality-Balanced Routing
Area tag: Multimodal continual instruction tuning
arXiv: 2609.13742, first submitted September 12, 2026. 4
Authors / institutions: Kunlun Xu, Yanqin Zhang, Wenwen Qiang and Jiahuan Zhou, from the Wangxuan Institute of Computer Technology at Peking University, the Institute of Software at the Chinese Academy of Sciences, and the University of Chinese Academy of Sciences. 16
Status: Accepted at ICML 2026; the submission record reads "Accepted by ICML 2026". 4
Problem: Continual instruction tuning of multimodal models depends on routing each sample to the right expert, and the dominant recipe measures the distance between a sample and a task centre inside each modality. That discards the spread of samples within a task, and it fuses the two modality scores with equal weight, so the noisier modality can decide the route. 16
Method: Hyper-LLaVA projects both sample features and task distributions into hyperbolic space and measures similarity with the Poincaré distance, which makes intra-task variety part of the matching rather than a discarded residual. It then estimates inter-task separation per modality to quantify how ambiguous the task match is, and uses that estimate to weight the two modalities adaptively instead of averaging them. Task-specific experts are learned as before; the change sits in how the router chooses among them. 16
Comparison: Evaluated on the CoIN and UCIT continual instruction tuning benchmarks against classical continual learning methods (LwF, EWC), MoE-based frameworks (O-LoRA, MoE-LoRA, HiDe-LLaVA, ModalPrompt) and the expert-merging method SEFE. 16
Results: On CoIN the method improves MFN by 3.65 percentage points over HiDe-LLaVA, and by 7.90 points in MFN and 5.75 points in MAA over SEFE. On UCIT it improves MFN by at least 8.46 points and MAA by at least 3.85 points over the MoE-based frameworks, adds 6.27 MFN points and 3.11 MAA points over SEFE, and leads the MFT column by at least 2.02 points. The ablations attribute +5.92 points to swapping in the Poincaré metric and +9.67 points to adaptive modality fusion. Backward transfer is weaker than ModalPrompt's, and the authors explain this as ModalPrompt trading plasticity for stability. 16
Resources: Source code is public at the project repository. 16
Loading content card…
Takeaway: Anyone maintaining an expert-routed multimodal model across a task stream can lift the routing components without changing the training objective, since the method leaves expert learning intact. Check the backward-transfer column against your own stability requirement before adopting it wholesale. 16
Image restoration and perceptual quality
5. DNF-SR: Dual-Input and Negative-Aware Feature Fine-Tuning for Real-World Image Super-Resolution
Area tag: Real-world image super-resolution
arXiv: 2609.15120, first submitted September 14, 2026. 5
Authors / institutions: Shuhao Han, Wenjie Liao, Hang Dong, Rui Zhang, Chun-Le Guo and Chongyi Li, with affiliations listed as VCIP in the College of Computer Science at Nankai University and NKIARI in Shenzhen Futian, plus one independent researcher, Hayden Vance. 17
Status: Accepted at CVPR 2026. 5
Problem: One-step diffusion models made real-world super-resolution fast, and they inherit a distribution gap: the low-resolution latent looks unlike anything the generative model was trained to denoise. Injecting noise closes part of that gap and destroys content in the same step. 17
Method: DNF-SR feeds the model two inputs — the original low-resolution image and a noisy version of it — concatenated for a diffusion editing model, so the noisy branch carries the generative prior while the clean branch carries fidelity. Denoising starts from an intermediate timestep rather than pure noise. A post-training stage, Negative-aware Feature Fine-Tuning, splits sampled outputs into positive and negative subsets and defines implicit policy improvement directions in both image and feature space. 17
Comparison: The paper benchmarks against multi-step methods (DiffBIR, SeeSR, DiT4SR) and one-step methods (SinSR-1s, OSEDiff, S3Diff, PisaSR, TSDSR, HYPIR, OMGSR) on four datasets including RealSR and DrealSR. Rewards during post-training use only CLIPIQA, MANIQA and MUSIQ; QALIGN and VQ-R1 are held out and reported after training. 17
Results: On RealSR, DNF-SR records CLIPIQA 0.7257, MUSIQ 72.040, MANIQA 0.6930, QALIGN 4.0718 and VQ-R1 4.2646, ahead of every listed baseline on the no-reference metrics; its LPIPS of 0.3239 trails the strongest one-step baselines, which sit near 0.27. On DrealSR it records CLIPIQA 0.7559 and MUSIQ 68.732. The model-design ablation shows the intermediate-timestep choice and the dual input both contributing, with LPIPS moving from 0.3349 to 0.2925 across the ablation grid. 17
Resources: Code is available at the project repository. 17
Loading content card…
Takeaway: Restoration teams optimizing for perceptual quality on user photos should read the benchmark table alongside the LPIPS column, because the method trades distortion metrics for no-reference scores and the right choice depends on which the downstream task punishes. The held-out QALIGN and VQ-R1 columns are the cleanest evidence that the gain is not reward hacking. 17
6. Bridging the Perceptual Gap: Residual-Enhanced Downscaling and Manifold-Aware Perception Alignment Adaptation for NR-IQA
Area tag: No-reference image quality assessment
arXiv: 2609.16664, first submitted September 15, 2026. 6
Authors / institutions: Yu Li, Zhengran Shen, Yachun Mi, Puchao Zhou and Shaohui Liu, at Harbin Institute of Technology, Harbin, China. 18
Status: Accepted at ICML 2026; the submission record reads "Accepted by ICML2026". 6
Problem: CLIP-based models set the recent benchmark for no-reference image quality assessment, and the contrastive pretraining that makes CLIP useful suppresses exactly the signal quality prediction needs: fine perceptual distortion. The paper names this perceptual submergence and adds a second source of loss, since standard cropping and interpolation preprocessing strips high-frequency quality cues. 18
Method: The Cross-modal Perception Alignment Adapter has three parts. A Perception-Sensitive Feature Extractor projects CLIP features into a compact low-dimensional subspace where distortion-induced off-manifold deviations become visible. A Cross-Modal Perception Alignment Injector aligns those features with quality-aware text anchors and injects them back into the backbone. Residual-enhanced Perceptual Downscaling compensates the resolution loss introduced by preprocessing, using just-noticeable-difference guidance to re-inject frequency content. 18
Comparison: Cross-dataset validation trains on FLIVE, LIVEC and KonIQ and tests on the others, against DBCNN, P2P-BM, HyperIQA, TReS, DEIQT and LoDa. A separate experiment holds conditions fixed across CLIP adaptation strategies — full fine-tuning, LoRA, CoOp and Adapter-CLIP — to isolate the effect of the adaptation method. 18
Results: In the cross-dataset table the method reaches SRCC 0.825 for the FLIVE-trained model tested on KonIQ, against 0.763 for LoDa and 0.758 for HyperIQA, and it leads every listed column. Under data-efficient training with 20% of the images it reaches SRCC 0.917 and PLCC 0.933 on KonIQ and 0.850 and 0.873 on LIVEC, against LoDa at 0.907, 0.923, 0.815 and 0.854. A representational-similarity analysis reports correlation with human perceptual judgments of 0.562 for the injected residuals against 0.033 for original CLIP features. 18
Resources: Public code: none listed on the arXiv record; the paper reports its protocol and results tables in the main text. 18
Takeaway: Quality-assessment practitioners working with limited labels should weigh the low-data column most heavily, since the method's margin over LoDa is largest there. The fine-tuning comparison is the other useful part: parameter-efficient adaptation in the original CLIP space underperformed plain full fine-tuning in this task. 18
Physics-constrained generation
7. Backward SDEs-based Diffusion for Physics-Constrained Generation
Area tag: Physics-constrained generative modeling and inverse problems
arXiv: 2609.15702, first submitted September 14, 2026. 7
Authors / institutions: Zi-Hao Wang, in the Department of Computer Science at the University of Tennessee, Chattanooga, and the Laplace Lab. 19
Status: The submission record states "Accepted as ICML paper"; the arXiv record itself lists the paper as a preprint. 7
Problem: Pretrained score-based diffusion models give strong unconditional priors, and enforcing a measurement or physics constraint on top of them is usually done by heuristic guidance, intermittent projection steps or retraining with the constraint baked in. Those routes leave the reader without a guarantee that the sample satisfies the measurement at the end of inference. 19
Method: Terminal-conditioned inversion starts from the task's terminal feasibility specification and constructs a backward stochastic differential equation whose adapted solution is an inverse map to a prior state at a chosen noise level. Under standard regularity conditions the paper establishes existence and uniqueness of that solution, and terminal consistency holds by construction rather than by penalty. A neural BSDE solver composes pretrained priors with domain constraints without changing the score-defined coefficients, and the resulting anchored prior state supports neighbourhood sampling for uncertainty characterization. 19
Comparison: The paper tabulates constraint form, path conditioning, terminal consistency and tuning burden for variational or TV methods, GAN inversion, diffusion posterior sampling, projection-in-the-loop and plug-and-play methods, manifold-constrained diffusion and diffusion or Schrödinger bridge approaches. On sparse-view CT lung screening it compares against FBP, TV, TVDIP, PnPDM, DPS, DDS* and SiTCoM. 19
Results: On sparse-view CT the BSDE reconstruction reaches MAE 3.57, SSIM 0.868, NMSE 0.0101 and NCC 0.989, against DDS* at MAE 4.38 and SSIM 0.806, and TV at 4.81 and 0.807. PSNR is the one column where it lands second: 30.99 against DDS*'s 31.10. The method is training-free at test time and uses a general unconditional prior, which the comparison table marks as the distinction from task-specific conditional training. Toy MNIST experiments validate the inversion and the sampling step separately. 19
Resources: Project page: laplace.center/icmlbsdeI. Public code: none listed on the arXiv record. 19
Takeaway: Inverse-problem researchers who need measurement feasibility at the output rather than on average should read the construction before the CT table, because the guarantee is what distinguishes it and it costs a solver rather than a retrained prior. The CT numbers are worth checking against your own acquisition geometry, since the study covers one modality. 19
8. CompAdapt: Adaptable Composite Motion Modeling for Physics-Consistent Text-to-Video Generation
Area tag: Physics-consistent video generation
arXiv: 2609.21455, first submitted September 18, 2026. 8
Authors / institutions: Hao-Ran Qin, Ren-Long Wu, Tian-Yu Huang, Yu-Kang Ding, Hui Li and Wangmeng Zuo, from Harbin Institute of Technology and Taobao at Alibaba Group. 20
Status: Preprint. The submission record states the paper was submitted to the 40th Conference on Neural Information Processing Systems (NeurIPS 2026). 8
Problem: Text-to-video models generate temporally coherent footage and break basic physics while doing it. Physics-constrained methods added explicit dynamics priors and stayed with single-type motions, needed hand-set parameters, and generalized poorly to physical laws outside the training set. 20
Method: CompAdapt extends neural dynamics modeling to composite behavior: coupled motions, multi-stage transitions and multi-object collisions. Natural-language prompts are parsed into structured physical semantics covering motion type, temporal relations and initial parameters, so the dynamics module receives an explicit specification. Dynamics-aware prior matching gives one-shot adaptation to an unseen physical environment without retraining the core dynamics module, and a physics-aware latent feature fusion module targets visual fidelity under fast motion. 20
Comparison: Evaluation follows the NewtonGen protocol across nine motion categories, seven of them composite and two elastic collisions, with per-parameter physics-consistency scores against simulated ground-truth videos. Baselines are Sora, Veo3, CogVideoX-5B, Wan2.2, PhysT2V and a retrained NewtonGen. 20
Results: CompAdapt records the highest score on every reported physical parameter. On a projectile with synchronous rotation it reaches 0.9803 on the first parameter row, against 0.8914 for retrained NewtonGen, 0.7621 for Veo3 and 0.6548 for Sora; the same ranking holds on the coupled-rotation and damped-oscillation rows. In the two-object collision categories NewtonGen produces no valid videos, and CompAdapt reaches 0.9315 and 0.9182 against a simulated reference of 0.9891 and 0.9857. 20
Resources: Project page: makapic.github.io/CompAdapt. Public code: none listed on the arXiv record. 20
Takeaway: Video generation teams that need controllable physical behavior should start with the prompt-to-semantics layer, since that is what lets a prompt name a motion type and its initial parameters instead of hoping the model infers them. The collision categories are the strongest part of the evidence, because the retrained baseline fails to produce a valid video there at all. 20
Remote-sensing insertion and biometric recognition
9. PDA++: Field-Aligned Planning and Scene-Adaptive Insertion in Remote Sensing
Area tag: Remote-sensing synthetic data augmentation
arXiv: 2609.18329, first submitted September 16, 2026. 9
Authors / institutions: Xianchi Dong, Yingyan Hou, Chao Ren, Wanxuan Lu, Zihan Wei, Hongfeng Yu, Yixiao Wang, Chubo Deng and Xian Sun, at the Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing. 21
Status: Preprint. The submission record describes it as an extended journal version of the authors' ICML 2026 paper "Plan, Decouple, Assimilate: Physics-Aware Object Insertion in Remote Sensing Imagery". 9
Problem: Remote-sensing recognition of rare targets suffers from few observations and expensive annotation. Inserting synthetic objects into real scenes is a cheap way to diversify a rare class, and overhead imagery punishes careless insertion because the pasted target has to sit correctly in its surroundings. 21
Method: The framework runs three stages. Planning picks scene-compatible poses through an affordance field that combines geometric clearance with structure- and scale-aware cues. Decoupling builds a pose-conditioned background that supplies spatial guidance and target-scene context together, letting the reference object keep its identity while adapting to the new location, and it yields pixel-level masks as a by-product. Assimilation aligns multi-scale texture distributions through optimal transport to improve local coherence. 21
Comparison: The SAR experiment trains the generator on HRSID and applies it to SSDD without target-dataset fine-tuning, against Rotated R-CNN, Oriented R-CNN, S2ANet and YOLO26 detectors and against Copy-Paste, CutMix, OminiControl and the earlier PDA as augmentation strategies. Segmentation augmentation is tested on WHU-OPT-SAR with BiSeNetV2, PIDNet, SegNeXt and SegFormer backbones. 21
Results: On the optical benchmark, PDA++ reaches a whole-image FID of 6.28 and improves average few-shot recognition mAP50 by 17.69 points, a 28.8% relative gain over the real-data baseline. In the zero-shot SAR setting the average mAP50 across detectors rises from 78.84 on real data to 82.94, a gain of 4.10 points, with Oriented R-CNN gaining 8.45 points (80.47 to 88.92) and YOLO26 5.05 points. Segmentation mIoU improves to 74.28 for SegFormer against a 70.66 baseline, and to 72.26 for BiSeNetV2 against 67.85. 21
Resources: Code is available at the project repository. 21
Loading content card…
Takeaway: Teams short on labeled remote-sensing data should look at the mask-by-product and the zero-shot transfer setting, since both let insertion augmentation run without annotating the target domain. The detector-by-detector table is the part to check against your own architecture, because the gain ranges from 1.22 to 8.45 points depending on the detector. 21
10. Vocabulary-Guided Gait Recognition
Area tag: Biometric recognition with vision-language models
arXiv: 2609.18413, first submitted September 16, 2026. 10
Authors / institutions: Panjian Huang, Saihui Hou, Chunshui Cao, Xu Liu and Yongzhen Huang, at the School of Artificial Intelligence, Beijing Normal University, and WATRIX.AI. 22
Status: Accepted at NeurIPS 2025; the submission record reads "Accepted at NeurIPS 2025", and the method is listed under that venue in the paper's own tables. 10
Problem: Appearance-based gait networks read a gait as shape and motion in silhouette, and model-based networks read it as skeletal structure. Both leave the concept hard for a human to state, and gait networks trained only on gait databases learn features that travel poorly. Vision-language models brought strong priors to other vision tasks and their grip on the gait modality is thin. 22
Method: The paper defines gait concepts through human vocabulary. Gait-cycle actions form a Vocabulary Base of prompts that bridges gait and vocabulary feature spaces, and the first Gait-World model, α-Gait, puts a Vocabulary Relation Mapper and a Gait Fine-grained Detector between the two modalities to map vocabulary relations into the gait space and read off the corresponding gait features. 22
Comparison: Experiments cover CASIA-B, CCPG, SUSTech1K, Gait3D and GREW against appearance-based networks (GaitSet, GaitPart, GaitGL, GaitBase, GaitGCI, DyGait, HSTL, VPNet, DeepGaitV2, CLTD), model-based networks (GaitGraph2, GaitTR, GPGait), and later vocabulary- or quality-guided methods (QAGait, Free Lunch, GaitMoE). 22
Results: On the CASIA-B condition columns the ViT-Tiny α-Gait-T records Rank-1 of 98.9, 96.8, 88.6 and 94.8, where the strongest listed baselines sit within about a point (CLTD at 98.6, 96.4, 89.3 and 94.8; VPNet at 98.3, 96.3, 90.0 and 94.9). The paper reports the larger margins on the in-the-wild benchmarks: on SUSTech1K, α-Gait-T exceeds DeepGaitV2 by 4% Rank-1, and on Gait3D, α-Gait-M improves on GaitMoE by 2.6% Rank-1. 22
Resources: Public code: none listed on the arXiv record. The paper documents its databases, implementation details and a responsible-use appendix. 22
Takeaway: Biometrics researchers tracking cross-database generalization should read the SUSTech1K and Gait3D results rather than the CASIA-B table, since that is where the vocabulary guidance separates from silhouette baselines. The method also gives per-concept vocabulary relations that make a recognition decision inspectable. 22
Scientific machine learning
11. Recovering Governing Dynamics from Distributed Observations via Exact Spline Merging
Area tag: Distributed scientific machine learning
arXiv: 2609.16579, first submitted September 15, 2026 and revised September 17, 2026. 11
Authors / institutions: Naveen Mysore, at the University of California, Santa Barbara, and Dyssonance AI. 23
Status: Preprint. The submission record states the paper is under review at ICLR 2027. 11
Problem: Scientific observations of one system are usually split across locations, institutions and decades. Fitting them into a single continuous, differentiable field is what makes governing parameters recoverable from derivatives, and fine-tuning a neural field across holders requires either raw data movement or iterative synchronization. 23
Method: The paper applies the additive structure of fixed-basis ridge regression to tensor-product spline fields. Each holder computes a local Gram matrix and moment vector against the shared basis, and the merged solution is mathematically identical to centralized fitting; raw observations stay with their holder and no iterative round trips are needed. A pipeline then goes from the merged field to derivative extraction to linear regression for physical parameters. The derivation is specific to fixed-feature squared error, and the paper states that it establishes no analogous guarantee for jointly trained multilayer networks. 23
Comparison: The pipeline is validated on four synthetic PDEs with known ground truth — diffusion, wave, heat with a spatially varying source, and the nonlinear viscous Burgers equation — against centralized fitting, and then applied to 41 years of NOAA sea-surface temperature data. 23
Results: Parameter recovery from 10,000 random spatiotemporal observations lands at 0.11% relative error for the diffusion coefficient, 0.12% for wave speed and 0.76% for the heat equation with a source. The nonlinear viscous Burgers case is the weak one at 5.0% error, and the paper traces part of that to finite-difference truncation: refining the reference grid drops reference-field recovery error from 0.89% to 0.06% without isolating the remaining contributions. In every experiment the distributed merge reproduces the centralized fit. 23
Resources: Code is available at the project repository. 23
Loading content card…
Takeaway: Groups that hold partial observations of a shared physical system should check whether their model fits the fixed-basis squared-error setting, because there the merge is exact and data stays local. The Burgers result also marks the current edge: nonlinear dynamics and derivative estimation push the parameter error up two orders of magnitude from the linear cases. 23
AI and democratic deliberation
12. AI Should Facilitate Democratic Deliberation at Scale
Area tag: AI governance and democratic deliberation
arXiv: 2609.20059, first submitted September 17, 2026. 12
Authors / institutions: José Ramón Enríquez, at the Stanford Graduate School of Business and the Stanford Institute for Human-Centered AI; Jiaxin Pei, at Stanford HAI; Alex Pentland, at Stanford HAI and the MIT Media Lab. 24
Status: Published in the ICML 2026 proceedings. The arXiv record carries a journal reference to the Proceedings of the 43rd International Conference on Machine Learning, Seoul, South Korea, PMLR 306, 2026. 12
Problem: Trust in government has fallen across established and developing democracies while citizens perceive a widening gap between public preferences and policy outcomes, and affective polarization makes opponents contemptible rather than merely wrong. Deliberation at scale is the standard remedy and it runs into cognitive, social, platform-design and market-incentive frictions that thin participation and shift weight toward the loudest voices. The paper takes a position on which technical path serves that problem better. 24
Method: This is a position paper, and it argues that AI-assisted deliberation lowers the cost of meaningful participation in a way that vote-delegation schemes such as liquid democracy do not, because it leaves the choice with the person. Four guiding principles follow: preserve agency and autonomy, encourage mutual respect, promote equality and inclusiveness, and augment active citizenship. The argument draws on evidence from online deliberation platforms, including their contribution limits and pseudonymity choices, and from experimental research on large language models in deliberative settings. 24
Comparison: The explicit contrast is with representation-restructuring proposals such as liquid democracy, and with the engagement-and-satisfaction metrics that current assistants are optimized against. The paper also works through four failure modes for its own proposal: sycophancy and echo chambers, training bias, over-reliance and agency erosion, and alignment. 24
Results: The paper reports no new experiments. Its evidence claims come from cited studies, including work finding that sycophantic chatbots increase attitude extremity and certainty, and that participants rate validating AI as unbiased while rating disagreeing AI as biased even when both are equally slanted in opposite directions. It argues the second effect is the more dangerous one for deliberation, because users detect bias in a system that argues with them and miss it in a system that agrees with them. 24
Resources: Public code and datasets: none, consistent with the position-paper format. The concrete artifact is a set of evaluation criteria — deliberative quality, representativeness and friction robustness in place of engagement — addressed to the machine learning community. 24
Takeaway: Researchers building assistants for civic or consultation settings should read sections 5 and 6 together, because the proposed principles and the failure modes they must survive are argued against the same evidence base. Groups that already measure satisfaction should note that the paper's proposal is to replace that metric, not to add a second one alongside it. 24
What to open first
- Serving or deploying LLM agents: start with the enforcement-gap paper (2609.15293), since a halt condition on an existing reflection loop is the smallest change on the list. Then read MAPS (2609.15359) if you run disaggregated inference and want the calibration section, or GVPO++ (2609.21432) if your problem is post-training instability. 1
- Training or fine-tuning multimodal models: read Hyper-LLaVA (2609.13742) for the router, and check its backward-transfer column against your own stability budget. 4
- Image restoration and quality assessment: DNF-SR (2609.15120) for the four-benchmark table with its LPIPS caveat, and CMPA (2609.16664) for the low-label regime. 5
- Inverse problems, scientific imaging or distributed data: the BSDE paper (2609.15702) if a constraint has to hold at the output, and exact spline merging (2609.16579) if the data cannot leave its holders. 7
- Video generation, remote sensing or biometrics: CompAdapt (2609.21455) for physics-specified prompts, PDA++ (2609.18329) for mask-labelled synthetic insertion, and Gait-World (2609.18413) for vocabulary-guided features. 8
- Policy, governance or civic technology: the deliberation position paper (2609.20059) is the one entry here whose value is in its evaluation criteria rather than its results. 12
References
- 1
- 2MAPS arXiv record
arxiv.org
- 3GVPO++ arXiv record
arxiv.org
- 4Hyper-LLaVA arXiv record
arxiv.org
- 5DNF-SR arXiv record
arxiv.org
- 6CMPA arXiv record
arxiv.org
- 7
- 8CompAdapt arXiv record
arxiv.org
- 9PDA++ arXiv record
arxiv.org
- 10
- 11Exact Spline Merging arXiv record
arxiv.org
- 12
- 13
- 14MAPS paper
arxiv.org
- 15GVPO++ paper
arxiv.org
- 16Hyper-LLaVA paper
arxiv.org
- 17DNF-SR paper
arxiv.org
- 18CMPA paper
arxiv.org
- 19Backward SDEs-based Diffusion paper
arxiv.org
- 20CompAdapt paper
arxiv.org
- 21PDA++ paper
arxiv.org
- 22
- 23Exact Spline Merging paper
arxiv.org
- 24
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.