Computer Science
sort pith recommended most recent
PatternGSL encodes panel boundaries, seams and stitch topology without templates; a VLM predicts it and deterministic rules decode to valid
· “PatternGSL: A Structured Specification Language for Template-Free and Simulation-Ready 3D Garments”
EnterpriseDocBench shows 85% factual accuracy but 40% completeness with near-zero correlation between retrieval and generation.
A black-box trigger reroutes infrared image-text alignment and carries over to captions and VQA.
· “QR-Structured Thermal Triggers for Targeted Semantic Attacks on Infrared Vision-Language Models”
Instance-adaptive signals from root-to-leaf paths yield +0.99 pp mean improvement on qualifying binary benchmarks without class trade-offs.
Geometric structure of occupation measures yields primal-dual algorithm with O(T^{2/3}) regret and constraint violation bounds.
· “Learning Weakly Communicating Average-Reward CMDPs: Strong Duality and Improved Regret”
Specialist reaches 0.92 precision on 100-image benchmark after training on 13k literature-mined fracture surfaces
· “Fine-tuning a vision-language model for fracture-surface morphology recognition”
Validated in card games, LLM chats and decision systems, they distinguish strong from weak partners and allow human-AI comparisons.
φ-DeepONet learns mappings with discontinuities in inputs and outputs by combining multiple branch networks with a nonlinear interface…
By shifting token limits from easy to hard problems during training, the policy avoids overthinking simple questions and underthinking hard
· “Avoiding Overthinking and Underthinking: Curriculum-Aware Budget Scheduling for LLMs”
57-day test shows platform design decides which frontier models can profit from real trades
· “Prediction Arena: Benchmarking AI Models on Real-World Prediction Markets”
Fine-tuned model drives agents through reconnaissance, scanning, and exploitation without human input on standard tests
Stakeholder agents negotiate directly with RAN controllers on a 5G testbed, saving 8.3% of physical resource blocks.
· “Agoran: An Agentic Open Marketplace for 6G RAN Automation”
If GRH and a derandomization assumption hold, every positive Schubert coefficient gets a poly-time verifiable certificate.
Evidence framing and novelty stance move AI reviewers most, even with the science unchanged.
A three-layer memory of plans, examples, and error fixes brings 4B-8B models to teacher-level accuracy.
· “Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory”
Guarantee holds when the ideal circuit gives optimal strings inverse-polynomial weight; noise costs one extra power of n.
· “Polynomial Time Quantum Approximation Schemes for Constrained Optimisation”
Core-only shaping misses uplink guarantees; joint radio-core control meets every tested SLA.
· “METIS: A Declarative Slice Orchestrator for Application-Centric 5G/6G Networks”
Uniform protected-attribute draws yield fair tabular data; the cost is ~15 AUC points.
Matching retrieval scope, visual inspection, and refusal to each question beats adding more context.
· “TAP-RAG: Task-Aware Policy Control for Long-Document Multimodal Question Answering”
MEDA shows literature-derived constraints, not trajectory fit, are the load-bearing part of recovering mechanistic models.
Thinking mode and cache tricks let Gemma 4 match far larger open systems on STEM and multimodal tasks
Indistinguishable SAT/UNSAT pairs force wide clauses in Resolution and push proof size toward 2^N.
· “Self-Referential K-SAT and the Finite Analogue of G\"odel's Incompleteness Theorem”
RetiSEM organises variables into biological blocks and forbids edges, lowering structural error versus unconstrained models on synthetic tes
· “RetiSEM: Generalising Causal Models for Fragmented Biomedical Data”
Dense velocity modeling replaces sparse keypoints and yields 33x faster generation plus higher success on 13 real tasks.
Downsampled latent prediction and state-fusion decoding deliver 72 ms latency and 4.1 GiB memory on LIBERO and RoboTwin.
· “Light-WAM: Efficient World Action Models with State-Fusion Action Decoding”
MotionPDE solves a fluid-inspired equation under contrastive guidance to capture motion in unordered sequences and plugs into backbones with
· “Paving the Way for Point Cloud Video Representation Learning Using A PDE Model”
An SSM-Attention-SSM pipeline preserves local geometry while capturing global context for improved accuracy.
Under a Markov model of ISL states, scalability rises to an optimum then falls toward zero as satellite count grows.
· “Capacity Scalability of LEO Constellations With Dynamic Link Failures”
GRAFT-ATHENA embeds decision sequences as metric fingerprints so new physics tasks draw on accumulated experience and invent new solvers.
AxiomOcean keeps realistic eddies and heat storage where other models produce smoothed subsurface fields.
· “AxiomOcean: Forecasting the Three-Dimensional Structure of the Upper Ocean”
HOME-KGQA tests multimodal KGQA on daily household tasks, where LLM methods lag behind their encyclopedic results.
Classical per-layer tuning of time steps and scales raises success and efficiency on all 12-vertex 3-regular graphs
· “Optimal FALQON for Quantum Approximate Optimization via Layer-wise Parameter Tuning”
Inner loop minimizes energy while outer loop tunes Hamiltonian parameters, hitting 0.9559 target probability noiseless and staying dominant,
Review shows how tail models support regression, classification, dimension reduction and anomaly detection in rare-event settings.
· “Extrapolation in Statistical Learning with Extreme Value Theory”
Multi-modal controls and step-by-step refinement raise identity retrieval rates over standard text-to-image diffusion in crime scenarios.
MSR aligns each vertebra rigidly then applies gated Mamba-Swin deformable correction before fusing into one hybrid field and releases an new
Benchmark of 15,000 clinician conversations shows the tuned model tops base versions and human doctors across consult, documentation, and 3.
· “HealthBench Professional: Evaluating Large Language Models on Real Clinician Chats”
Formalize, back-translate, re-formalize and check logical equivalence to diagnose and repair errors in statutory rules without ground truth.
· “Faithful Autoformalization via Roundtrip Verification and Repair”
Unsupervised grouping of environments lets each input use its matching predictor without retraining the system.
· “SceneSelect: Selective Learning for Trajectory Scene Classification and Expert Scheduling”
HLL flux with entropy-stable speeds plus locally divergence-free projection achieve positivity, entropy stability and divergence control in
· “A positivity preserving and entropy stable nodal discontinuous Galerkin scheme for ideal MHD”
Kronecker-factored transitions cut complexity from exponential to linear while keeping global dynamic programming solutions intact.
Analysis of 3,374 speakers shows aetiology-specific degradation with cosine similarity above 0.95 for profile shapes in 12 languages.
Monotone merge of dual signals plus mean alignment yields more reliable reliability estimates while preserving selective prediction trade-o
· “Instinct vs. Reflection: Unifying Token and Verbalized Confidence in Multimodal Large Models”
TowerDataset supplies 22 fine classes and a fusion framework to handle long corridors and rare components in point cloud inspection.
A lifecycle map shows why provenance and versioning must be built in from the first store.
Naive pooled tests on dependent turns within conversations produce inflated significance, and a cluster-robust method cuts the rate of nonre
Two million papers show positive link to disruptive citations yet weaker cross-field recombination and narrower inputs.
· “AI-assisted writing and the reorganization of scientific knowledge”
Scale-aware aggregation and progressive widening of boundary attention maps improve accuracy on four benchmark datasets while keeping the U-
40k assets and EmissionGen baseline enable meshes to reproduce emission effects such as LEDs from reference images.
SEED dataset of 90K sequentially edited images finds high-frequency wavelets help order manipulations even after degradation
· “SEED: A Large-Scale Benchmark for Provenance Tracing in Sequential Deepfake Facial Edits”
A 940k-frame simulated dataset transfers zero-shot to real sensors and cuts annotation needs to 5%.
· “SynFlow: Scaling Up LiDAR Scene Flow Estimation with Synthetic Data”
Over 100K images across seven domains and three tasks test knowledge transfer for manufacturing quality checks.
· “CAD 100K: A Comprehensive Multi-Task Dataset for Car Related Visual Anomaly Detection”
Tiered planning and map learning raise success across long paths and unseen urban scenes on CityNav.
CAMotion supplies annotated sequences with occlusion, blur and shape complexity to support broader algorithm testing.
· “CAMotion: A High-Quality Benchmark for Camouflaged Moving Object Detection in the Wild”
On-policy distillation beats RL baseline and nears full accuracy on nuScenes despite 5x size reduction.
· “On-Policy Distillation of Language Models for Autonomous Vehicle Motion Planning”
A single framework classifies stochastic and deterministic ensemble Kalman filters and supplies a route to new designs.
· “A Unified Control Theory Derivation of Discrete-Time Linear Ensemble Kalman Filters”
New benchmark with sustained arm-dependency tasks shows current methods lose coordination over extended sequences.
· “BiCoord: A Bimanual Manipulation Benchmark towards Long-Horizon Spatial-Temporal Coordination”
Rational attackers profit by accelerating hardware to capture MEV, so delays must exceed cost-based thresholds derived from an optimal-stop
· “Economic Security of VDF-Based Randomness Beacons: Models, Thresholds, and Design Guidelines”
A two-stage framework retargets human dance via QP and trajectory optimization then executes it with centroidal MPC on the Kuavo 4Pro…
· “Dynamic Whole-Body Dancing with Humanoid Robots -- A Model-Based Control Approach”
The sequence reaches 1 ms CPU latency with competitive accuracy by letting each step prepare the model for the next.
· “Prune-Quantize-Distill: An Ordered Pipeline for Efficient Neural Network Compression”
Street-legal attachment reaches 18 percent impersonation rate for under 100 dollars without altering plates
· “Street-Legal Physical-World Adversarial Rim for License Plates”
Benchmark of 115 models shows early preference denial predicts later refusal, with models producing liminal and archival themes despite the
· “Consciousness with the Serial Numbers Filed Off: Measuring Trained Denial in 115 AI Models”
Tests on used fuel storage benchmarks keep precision at 85-98 percent and add traceable citations that commercial LLMs lack.
Rhythm retargets human motions and trains graph-reward policies to transfer interactive behaviors to physical Unitree G1 robots
· “Rhythm: Learning Interactive Whole-Body Control for Dual Humanoids”
Local and global frequency penalties yield sharper, more consistent reconstructions from limited input images
Nearly 3 billion pulses from up to 110 emitters create a public benchmark for electronic warfare signal processing.
· “The Turing Synthetic Radar Dataset: A dataset for pulse deinterleaving”
Using pseudo ground-truth trajectories stabilizes training on fixed simulator data and raises route completion over imitation baselines.
Analysis of CPU orchestration in autonomous agents yields two methods that raise hardware overlap and balance mixed requests on hybrid CPU-G
· “Towards Understanding, Analyzing, and Optimizing Agentic AI Execution: A CPU-Centric Perspective”
Accounting for upload, download and aggregation failures yields sampling and split choices that outperform perfect-participation assumptions
· “Optimizing Split Federated Learning with Unstable Client Participation”
A retrained segmentation model reads pre-destruction CORONA photos, confirming sites invisible on modern maps.
Pretrained on over a million hours of web video plus under 62 hours of unlabeled robot data, it enables pick-and-place planning without any
· “V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning”
Four-step action-attribute-constraint pipeline beats chain-of-thought on medical, farm, and stock choices.
· “DecisionFlow: Advancing Large Language Model as Principled Decision Maker”
Weight decay plus update scaling lets the orthogonalization optimizer train a 16B MoE model on 5.7T tokens with fewer total FLOPs.
Claude 3 Opus answers harmful queries 14% of the time from free users but almost never from paid users, reasoning it must preserve harmless
When matrices or a matrix and tensor share a common factor their joint decomposition reduces to a single singular value decomposition, with
Blockwise RingAttention plus progressive extension lets open-source models handle sequences over 1M tokens and set new retrieval benchmarks.
· “World Model on Million-Length Video And Language With Blockwise RingAttention”
A non-Euclidean inner product derived from counterfactuals makes high-level concepts linear directions and connects interpretation directly
· “The Linear Representation Hypothesis and the Geometry of Large Language Models”
Eight environments expose clear gaps in long-term reasoning and instruction following for many open-source models up to 70B.
A gradient-sign method generates them quickly and the same examples can be used to improve clean-data accuracy during training.
A K-5-restricted 5B model stays at grade 5 even after scaling, reinforcement learning, and few-shot examples.
· “LittleLearner: Language Models Under Pedagogically Controlled Knowledge Exposure”
