Progressive Scene Graphs Improve Dynamic Urban NeRFs
Local scene graphs plus DINOv2-guided object codes lift KITTI PSNR to 30.31 while preserving object editing across long driving sequences.
Jul 13, 20265 min2312.09076 Code available
Computer Science
Research across all areas of computer science, from theoretical foundations to applied systems.
Sort:
Legal Prediction Models Exploit Tribunal Text Shortcuts
Auditing 33,158 UK employment claims shows outcome-revealing phrases can deliver 56.07 Macro-F1 by themselves, close to full-text classifiers.
Jul 13, 20265 min2607.04261 Code available
Ruby Finds Unsafe Rust Regions in Stripped Binaries
Instruction-level binary features train a classifier that recovers 91.75% of unsafe regions at a 6.16% false positive rate.
Jul 13, 20264 min2211.00111
Code Transformations Detect LLM-Written Python Across Models
ACW applies idempotent semantic-preserving Python rewrites after generation, reaching over 97% accuracy with under 0.1 seconds per function.
Jul 13, 20265 min2402.07518 Code available
Full-Stack FP4 Narrows Pretraining Loss Gap
Module-specific NVFP4 quantization covers projections, optimizers, and attention, reaching a 1.47% loss gap on 3B/64B-token pretraining.
Jul 12, 20264 min2607.04422
GPU-Owned Remote I/O Speeds Graph ANNS
GORIO keeps page misses and resume decisions on the GPU, reaching 1.31× over the remote-I/O reference path on SIFT1M.
Jul 12, 20264 min2607.04415
Attention Alone Improves Machine Translation and Training Efficiency
The Transformer replaces recurrence and convolution with multi-head self-attention, reaching 28.4 BLEU on WMT 2014 English-German with shorter training.
Jul 12, 20264 min1706.03762 Code available
ACE-Brain Extends Spatial Models Into Closed-Loop Robot Agents
An 8B backbone couples perception, planning, action, and progress estimation, improving 14 of 18 spatial benchmarks over ACE-Brain-0.
Jul 12, 20265 min2607.04426
Shared Backbones Enable Cross-Task Transfer in Multimodal Models
Controlled experiments across four UMMs find that understanding fine-tuning can improve three generation skills while avoiding direct generation shift.
Jul 12, 20264 min2607.04423
On-Policy Distillation Reduces Cross-Platform GUI Forgetting
Platform-conditioned teacher selection distills desktop and mobile policies into one continual learner, reaching 38.2% OSWorld and 12.0% MobileWorld success.
Jul 12, 20264 min2607.04425