# 990 Published articles for 990. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Off-Policy RL with Online Inference Tuning: My Solution to the CoRL 2026 RoboEval Challenge DevFeed: [Off-Policy RL with Online Inference Tuning: My Solution to the CoRL 2026 RoboEval Challenge]() Original publisher: [Read original article]() Author: Ilia Larchenko Published: 2026-11-10T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [grpo](), [thompson sampling]() Tags: [2026](), [990](), [challenge](), [competition](), [convergence](), [cs-cv](), [cs-lg](), [cs-ro](), [data](), [exploration](), [failures](), [full](), [inference](), [online](), [policy](), [rds](), [recovery](), [reward](), [rl](), [sampling](), [score](), [tasks](), [thompson-sampling]() ### AI overview The solution combines off-policy reinforcement learning with targeted demonstrations and recovery data to improve performance on eight bimanual RoboEval tasks. Its generic method trains a π₀.₅-based policy that also serves as its value function, using AWR and RECAP on advantages derived from the official score. Each rollout also tests about 20 inference parameters through Thompson sampling. This reached 0.85 RDS on held-out seeds in two days, though training was far from converged. The competition-specific additions--generated demonstrations reaching full reward and recovery episodes beginning from the policy's own failures--addressed situations local exploration did not find. The submitted policy reached 0.990 RDS on held-out seeds and 0.987 on official seeds. ### Source excerpt arXiv:2611.11464v1 Announce Type: cross Abstract: The CoRL 2026 RoboEval challenge has eight bimanual tasks, and it ranks a policy not only on whether it succeeds but on how: every stage of the task, done safely, efficiently and with both arms in coordination, scored against human demonstrations. My solution has two parts. The generic part, carried over from my LeHome challenge solution, is off-policy RL on a $\pi_{0.5}$-based policy that is also its own value function. It is trained with AWR and RECAP on advantages of the exact official score, in a loop where every rollout is at the same time a Thompson-sampling trial of about 20 inference parameters. Out of the box, it took the policy to 0.85 RDS on held-out seeds in two days (and it was still very far from convergence). The competition-specific part adds data that local exploration cannot find: generated demonstrations that reach the full reward, and recovery episodes that start from the policy's own failures. The submitted policy reaches 0.990 RDS on held-out seeds and 0.987 on the official seeds. ## Hybrid acoustic and pretrained embedding feature fusion for unmanned aerial vehicle detection using stacking ensemble learning under a leakage-aware evaluation protocol DevFeed: [Hybrid acoustic and pretrained embedding feature fusion for unmanned aerial vehicle detection using stacking ensemble learning under a leakage-aware evaluation protocol]() Original publisher: [Read original article]() Author: Zahid Iqbal Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Feature Engineering](), [Embeddings](), [benchmark overfitting machine learning](), [Hyperparameter optimization]() Tags: [000](), [2](), [22](), [3](), [990](), [abstract](), [all](), [area](), [cs-lg](), [cs-sd](), [detection](), [embedding](), [ensemble-learning](), [evaluation](), [feature](), [fusion](), [hybrid](), [learning](), [protocol](), [stacking](), [under](), [using]() ### AI overview On DroneAudioDataset, a source-disjoint evaluation produced a mean accuracy of 95.29% for a stacking ensemble combining hand-crafted acoustic descriptors with pretrained YAMNet embeddings. The dataset draws drone clips from just seven recordings, while 96.4% of non-drone clips come from ESC-50, making random clip-level splits vulnerable to source leakage; those splits overstated accuracy by 1.89-3.22 points across four models. The two-view stack achieved 98.29% specificity and 0.990 ROC AUC, and adding YAMNet raised accuracy by 1.11 points, with gains in 12 of 15 folds. Hybrid feature selection matched all 44 baseline descriptors using 16-18 features; NSGA-II used 4-6 but had lower accuracy. In the reported setup, the stack took 51 ms per clip on one CPU thread. ### Source excerpt arXiv:2611.04714v1 Announce Type: cross Abstract: Small multirotor drones produce a characteristic acoustic signature, and microphones offer a low-cost way to detect them. High accuracies have been reported on public drone audio benchmarks, often with random clip-level splits of datasets built from a few long recordings and pieces of existing sound collections, so pieces of one source can appear in both training and test data. We study the binary task of the public DroneAudioDataset (11,704 one-second clips). Its 1,332 drone clips come from 257 five-second chunks of 7 recordings, and 96.4% of its non-drone clips are pieces of the 2,000 clips of the ESC-50 sound dataset. A grouped, source-disjoint protocol (five folds, three repeats) is compared with a random split. Each clip is described by 57 hand-crafted acoustic descriptors, reduced by a hybrid minimum-redundancy maximum-relevance and ReliefF selection or by NSGA-II multi-objective search, and by a 1024-dimensional embedding from the pretrained YAMNet network. Random Forest and XGBoost learners on both views form a stacking ensemble. Under the grouped protocol, the two-view stack reached a mean accuracy of 95.29% (SD 1.62 points over 15 test folds), a specificity of 98.29% and an area under the ROC curve of 0.990, and it correctly classified 98.78% of unseen ESC-50 source clips as non-drone. The random split overstated accuracy by 1.89 to 3.22 points for the four models. Adding YAMNet embeddings to the hand-crafted view raised accuracy by 1.11 points (higher in 12 of 15 folds). Hybrid selection matched the accuracy of all 44 baseline descriptors with 16 to 18 features, and NSGA-II retained 4 to 6 features at lower accuracy. The stack took 51.0 ms per clip on one CPU thread. Source-disjoint evaluation is needed for a fair benchmark on this dataset, and fusing hand-crafted descriptors with pretrained embeddings gives a small but consistent gain at modest computational cost. ## Before you expand your GPU pool, measure the outcome DevFeed: [Before you expand your GPU pool, measure the outcome]() Original publisher: [Read original article]() Author: Danielle Zeis Published: 2026-10-24T12:00:00Z Content type: article Language: en Sources: [Kubernetes Overview, News and Trends | The New Stack](), [The New Stack]() Topics: [AI Infrastructure](), [Latency]() Tags: [2](), [990](), [about](), [account](), [admission](), [ai](), [ai-infrastructure](), [ai-workloads](), [all](), [also](), [api](), [application](), [batch](), [cncf](), [finops](), [first](), [gpu](), [job](), [kubernetes](), [make](), [post-contributed](), [queue](), [running](), [sponsor-cncf](), [sponsored-post-contributed](), [teams]() ### AI overview A full GPU pool and a growing queue do not by themselves show that more GPUs are needed. Define each workload's deadline, runtime, resource needs, interruption tolerance, and measurable quality and latency requirements. Flexible scheduling can cut peak demand: the excerpt's four two-GPU jobs can be staggered across four available GPUs and still finish before their deadline. For serving, estimate capacity with representative load tests that hold model, hardware, settings, and request mix constant while measuring throughput, quality, latency, and errors; GPU activity or cache occupancy alone does not establish spare capacity. OpenCost can provide interval-based workload allocation and opt-in cost per million tokens, but application systems must supply verified outcomes and evaluation evidence. Join costs, telemetry, and outcomes carefully, accounting for shared, idle, retry, and failure costs. Then compare tested changes such as batching, model choice, or replica sizing against cost per qualifying outcome and required headroom before deciding whether to add GPUs. ### Source excerpt For teams running AI workloads on Kubernetes, a fully allocated GPU pool and a growing batch-job queue might make expanding The post Before you expand your GPU pool, measure the outcome appeared first on The New Stack. ## Indian Payment Methods for Global SaaS: UPI, Cards, NetBanking, Wallets DevFeed: [Indian Payment Methods for Global SaaS: UPI, Cards, NetBanking, Wallets]() Original publisher: [Read original article]() Author: Ayush Agarwal Published: 2026-05-13T00:00:00Z Content type: article Language: en Sources: [Dodo Payments Blog]() Topics: [payment-gateway](), [Use Cases for Every Cloud Service Model]() Tags: [000](), [2](), [3](), [3d](), [990](), [about](), [accounting](), [adaptive](), [delay](), [global](), [india](), [indian](), [integration](), [patterns](), [payment](), [payments](), [processing](), [rules](), [saas](), [subscriptions](), [upi]() ### AI overview For global SaaS serving India, the article recommends INR checkout with UPI prominently offered and cards as fallbacks, while treating Indian recurring billing as mandate-based. The initial subscription payment can be immediate, but recurring debits require advance customer notification and a 48-hour processing window; access should wait for the `payment.succeeded` webhook, which the article says arrives about 48-51 hours after the scheduled date. Typical plans up to Rs 15,000 may use an on-demand mandate capped at that amount; higher plans require a mandate for the exact amount, and upgrades beyond an existing limit require customer re-authorization. UPI requires INR and an Indian billing country; non-Indian merchants need Adaptive Currency enabled for UPI to appear. The article describes NetBanking and wallets as optional, less-used additions. It also advises monitoring mandate cancellations and using the platform mandate-amount field instead of hardcoding limits. ### Source excerpt How global SaaS accept UPI and Indian-issued cards, the RBI mandate rules for subscriptions, the 48-hour processing delay, and integration patterns that win in India.